c
CONTENTS
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
Introduction
B.E. Launder and N.D. Sandham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Part A. Physical and Numerical Techniques
1. Linear and Nonlinear Eddy Viscosity Models
T.B. Gatski and C.L. Rumsey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2. Second-Moment Turbulence Closure Modelling
K. Hanjalic and S. Jakirlic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3. Closure Modelling Near the Two-Component Limit
T.J. Craft and B.E. Launder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4. The Elliptic Relaxation Method
P.A. Durbin and B.A. Pettersson-Reif . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5. Numerical Aspects of Applying Second-Moment Closure to Complex Flows
M.A. Leschziner and F.-S. Lien . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6. Modelling Heat Transfer in Near-Wall Flows
Y. Nagano . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7. Introduction to Direct Numerical Simulation
N.D. Sandham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8. Introduction to Large Eddy Simulation of Turbulent Flows
J. Fr ohlich and W. Rodi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
9. Introduction to Two-Point Closures
C. Cambon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
10. Reacting Flows and Probability Density Function Methods
D. Roekaerts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
vi Contents
Part B. Flow Types and Processes and Strategies for Mod-
elling them
Complex Strains and Geometries
11. Modelling of Separating and Impinging Flows
T.J. Craft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
12. Large-Eddy Simulation of the Flow past Blu Bodies
W. Rodi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
13. Large Eddy Simulation of Industrial Flows?
D. Laurence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
Free Surface and Buoyant Eects on Turbulence
14. Application of TCL Modelling to Stratied Flows
T.J. Craft and B.E. Launder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
15. Higher Moment Diusion in Stable Stratication
B.B. Ilyushin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
By-Pass Transition
16. DNS of Bypass Transition
P.A. Durbin, R.G. Jacobs and X. Wu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
17. By-Pass Transition using Conventional Closures
A.M. Savill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
18. New Strategies in Modelling By-Pass Transition
A.M. Savill . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
Compressible Flows
19. Compressible, High Speed Flows
S. Barre, J.-P. Bonnet, T.B. Gatski and N.D. Sandham . . . . . . . . . . . . . . . 522
Combusting Flows
20. The Joint Scalar Probability Density Function Method
W.P. Jones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
21. Joint Velocity-Scalar PDF Methods
H.A. Wouters, T.W.J. Peeters and D. Roekaerts . . . . . . . . . . . . . . . . . . . . . . 626
Contents vii
Part C. Future Directions
22. Simulation of Coherent Eddy Structure in Buoyancy-Driven Flows with
Single-Point Turbulence Closure Models
K. Hanjalic and S. Kenjeres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659
23. Use of Higher Moments to Construct PDFs in Stratied Flows
B.B. Ilyushin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
24. Direct Numerical Simulations of Separation Bubbles
G.N. Coleman and N.D. Sandham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 702
25. Is LES Ready for Complex Flows?
B.J. Geurts and A. Leonard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 720
26. Recent Developments in Two-Point Closures
C. Cambon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 740
CONTRIBUTORS
S. Barre, Universite de Poitiers, 40 Avenue du Recteur Pineau, 86022 Poitiers Cedex,
France
stephane.barre@lea.univ-poitiers.fr
J-P. Bonnet, Universite de Poitiers, 40 Avenue du Recteur Pineau, 86022 Poitiers
Cedex, France
bonnet@univ-poitiers.fr
C. Cambon, Laboratoire de Mecanique des Fluides et dAcoustique, Ecole Centrale
de Lyon, 36 avenue Guy de Collongue, BP 163, 69131 Ecully Cedex, France
cambon@mecaflu.ec-lyon.fr
G.N. Coleman, Aeronautics and Astronautics, School of Engineering Sciences, Uni-
versity of Southampton, Southampton SO17 1BJ, UK
gnc@soton.ac.uk
T.J. Craft, Department of Mechanical, Aerospace and Manufacturing Engineering,
UMIST, PO Box 88, Manchester M60 1QD, UK
tim.craft@umist.ac.uk
P.A. Durbin, Department of Mechanical Engineering, Stanford University, Stanford
CA 94305-3030 USA
pdurbin@stanford.edu
J. Fr ohlich, Institut f ur Hydromechanik, Universit at Karlsruhe, Kaiserstr.12, D-76128
Karlsruhe, Germany
froehlich@ifh.uni-karlsruhe.de
T.B. Gatski, Computational Modeling & Simulation Branch, Mail Stop 128, NASA
Langley Research Center, Hampton VA 23681-2199, USA
t.b.gatski@larc.nasa.gov
B.J. Guerts, Faculty of Mathematical Sciences, University of Twente, PO Box 217,
7500 AE Enschede, The Netherlands
b.j.geurts@math.utwente.nl
K. Hanjalic, Department of Applied Physics, Thermal and Fluids Sciences, University
of Technology Delft, Lorentzweg 1, NL-2628 CJ Delft, Netherlands
hanjalic@ws.tn.tudelft.nl
B. Ilyushin, Institute of Thermophysics SD RAS, Lavrentyev Avenue 1, 630090 Novo-
sibirsk, Russia
ilyushin@itp.nsc.ru
R. Jacobs, TenFold Corporation, Draper, UT 84020, USA
rjacobs@10fold.com
S. Jakarlic, Institute for Fluid Mechanics and Aerodynamics, Darmstadt University
of Technology, Petersenstr. 30, D-64287, Darmstadt, Germany
j.suad@hrzpub.tu-darmstadt.de
Contributors ix
W.P. Jones, Department of Mechanical Engineering, Imperial College of Science Tech-
nology and Medicine, University of London, Exhibition Road, London SW7 2BX,
UK
w.jones@ic.ac.uk
S. Kenjeres, Department of Applied Physics, Thermal and Fluids Sciences, University
of Technology Delft, Lorentzweg 1, NL-2628 CJ Delft, Netherlands
kenjeres@ws.tn.tudelft.nl
B.E. Launder, Department of Mechanical, Aerospace and Manufacturing Engineering,
UMIST, PO Box 88, Manchester M60 1QD, UK
brian.launder@umist.ac.uk
D.R. Laurence, Department of Mechanical Engineering, UMIST, PO Box 88, Manch-
ester M60 1QD, UK; and Electricite de France
dominique.laurence@umist.ac.uk
A. Leonard, Graduate Aeronautical Laboratories, California Institute of Technology,
Pasadena CA 91125, USA
tony@galcit.caltech.edu
M.A. Leschziner, Imperial College of Science Technology and Medicine, Aeronautics,
Department, Prince Consort Rd., London SW7 2BY, UK
Mike.Leschziner@ic.ac.uk
F-S. Lien, Department of Mechanical Engineering, University of Waterloo, Waterloo,
Ontario N2L 3G1, Canada
fslien@sunwise.uwaterloo.ca
Y. Nagano, Department of Environmental Technology, Nagoya Institute of Technol-
ogy, Gokiso-cho, Showa-ku, Nagoya 466-8555, Japan
nagano@heat.mech.nitech.ac.jp
T.W.J. Peeters, Corus Research, Development & Technology P.O. Box 10000, 1970
CA IJmuiden, The Netherlands
tim.peeters@corusgroup.com
B.A. Petterson-Reif, Norwegian Defence Research Establishment, N-2025 Kjeller, Nor-
way
bre@ffi.no
W. Rodi, Institut f ur Hydromechanik, Universit at Karlsruhe, Kaiserstr.12, D-76128
Karlsruhe, Germany
rodi@ifh.uni-karlsruhe.de
D. Roekaerts, Department of Applied Physics, Thermal and Fluids Sciences, Univer-
sity of Technology Delft, Lorentzweg 1, NL-2628 CJ Delft, Netherlands
roekaerts@tnw.tudelft.nl
C.L. Rumsey, Computational Modeling & Simulation Branch, Mail Stop 128, NASA
Langley Research Center, Hampton VA 23681-2199, USA
c.l.rumsey@larc.nasa.gov
N.D. Sandham, Aeronautics and Astronautics, School of Engineering Sciences, Uni-
versity of Southampton, Southampton SO17 1BJ, UK
n.sandham@soton.ac.uk
x Contributors
A.M. Savill, Department of Engineering, University of Cambridge, Trumpington St.,
Cambridge CB2 1PZ, UK
ams3@eng.cam.ac.uk
H.A. Wouters, Corus Research, Development & Technology, P.O. Box 10000, 1970
CA IJmuiden, The Netherlands
huib.wouters@corusgroup.com
X. Wu, Department of Mechanical Engineering, Stanford University, Stanford CA
94305-3030, USA
xiaohua wu@hotmail.com
PREFACE
Material for this volume rst began to be assembled in 1999 when a six-month Pro-
gramme on Turbulence was held at the Isaac Newton Institute for Mathematical
Sciences in Cambridge. The Programme had its own origins in an Initiative on Tur-
bulence by the UK Royal Academy of Engineering, which had identied the prediction
of turbulent ow as a key technology across a range of industrial sectors. Researchers
from dierent disciplines gathered together at the Newton Institute building in Cam-
bridge to work on aspects of the turbulence problem, one of the most important of
which was closure of the averaged (or ltered) turbulent ow equations. An Instruc-
tional Workshop on Closure Strategies for Modelling Turbulent and Transitional Flows
was held from April 6 to 16th. The aim of the workshop was for the experts on closure
to present material leading up to the current state-of-the-art in a form suitable for
research students and others requiring a broad overview of the eld, together with
an appreciation of current issues. A sequence of 38 lectures by 19 dierent lecturers
provided background to techniques, examples of current applications, and reections
on possible future developments. Recognising that a gathering of so many experts
from a variety of backgrounds was somewhat unusual, it was felt that a more polished
version of the lecture material would be of interest to a wider audience, as a reection
upon the current state of prediction methods for turbulent ows. The result has taken
rather longer to appear than originally intended, but the opportunity has been taken
for contributions to be updated wherever possible, to take account of the most recent
developments.
The Editors would like to thank the contributors for their eorts to improve the
papers, with a view to making them more accessible to the intended audience, which
remains that of the original workshop. Especial thanks are due to Dr T. Gatski who
shouldered the not inconsiderable task of shaping three separate contributions into a
single chapter on compressible ow. Thanks are also due to the Newton Institute for
hosting the workshop, and the main sponsors of the Turbulence Programme, including
the then British Aerospace (now BAE Systems), RollsRoyce, The Meteorological Of-
ce, British Gas Technology, British Energy, and the UK Defence Evaluation Research
Agency. Finally we record our appreciation to Mrs C. King who provided invaluable
secretarial support throughout the Workshop and in the editing of the present volume.
ACRONYMS
ABL Atmospheric Boundary Layer
APG Adverse Pressure Gradient
ASM Algebraic Stress Model [1]
BL Binomial Lagrangian [21]
C/D Coalescence/Dispersion [21]
CDF Cumulative Distribution Function [20]
CERT Centre d
, C
displacement thickness
P, p mean pressure scalar invariant (
S
ik
S
ki
)
T turbulent kinetic energy von Karman constant
production term density
R symmetric, traceless tensor
ij
viscous stress tensor
in algebraic stress equation
ij
pressure strain rate
2
ow parameter correlation
( W
2
/S
2
) kinematic viscosity
S
ij
, S mean strain rate tensor
t
,
t
turbulent eddy viscosity
( (U
i
/x
j
+U
j
/x
i
)/2)
ti
,
to
inner and outer eddy viscosity
T characteristic time scale in turbulent time scale (= k/)
wall proximity
ij
, Reynolds stress tensor ( u
i
u
j
)
T
(n)
tensor basis element
ij
arbitrary time-independent
U
e
edge velocity rotation rate of noninertial
U
i
mean velocity component frame
u
friction velocity
r
rotation rate of noninertial
W
ij
, W mean rotation rate tensor frame
in noninertial frame dissipation rate per unit kinetic
( (U
i
/x
j
U
j
/x
i
)/2) energy
W
ij
, W
. (2.1)
The average of a uctuating quantity is zero f
.
The velocity (u
i
) and pressure (p) elds can be decomposed into their mean
(U
i
, P) and uctuating parts (u
i
, p), and the resulting Reynolds-averaged
NavierStokes (RANS) equations can be written as
DU
i
Dt
=
U
i
t
+U
j
U
i
x
j
=
1
P
x
i
+
ij
x
j
ij
x
j
. (2.3)
For an incompressible ow, the mass conservation equation reduces to the
mean continuity equation,
U
j
x
j
= 0. (2.4)
12 Gatski and Rumsey
The viscous stress tensor
ij
for a Newtonian uid and incompressible ow is
given by
ij
= 2S
ij
, (2.5)
where is the kinematic viscosity, and S
ij
is the strain rate tensor
S
ij
=
1
2
_
U
i
x
j
+
U
j
x
i
_
. (2.6)
As equation(2.3) shows, for closure the RANS formulation requires a model
for the second-moment (or Reynolds stress)
ij
(= u
i
u
j
).
3 Linear eddy viscosity models
Using continuity, equation(2.4), and the denition of the viscous stress, equa-
tion(2.5), the RANS equation can be written in the form
DU
i
Dt
+
1
P
x
i
=
ij
x
j
+
x
j
_
U
i
x
j
_
. (3.1)
For linear eddy viscosity models (linear EVMs), the equation is closed by using
a Boussinesq-type approximation between the turbulent Reynolds stress and
the mean strain rate
ij
=
2
3
k
ij
2
t
S
ij
, (3.2)
where k (=
ii
/2) is the turbulent kinetic energy, and
t
is the turbulent eddy
viscosity. In Section 3.4, it will be shown that such a closure model can be
extracted from an analysis of a simple shear ow in local equilibrium. When
equation(3.2) is used as the turbulent closure in linear EVMs, equation(3.1)
can be rewritten as
DU
i
Dt
+
1
p
x
i
=
x
j
_
( +
t
)
U
i
x
j
_
, (3.3)
where the isotropic part of the closure model, 2k/3, is assimilated into the
pressure term so that p = P + 2k/3. In the EVM formulation, the turbulence
eld is coupled to the mean eld only through the turbulent eddy viscosity,
which appears as part of an eective viscosity ( +
t
) in the diusion term of
the Reynolds-averaged NavierStokes equation. Since in general,
t
> , this
formulation of the problem can be rather robust numerically, especially when
compared to the alternative form of retaining the stress gradient
ij
/x
j
explicitly in equation(3.1).
In the remainder of this section, a hierarchy of linear eddy viscosity models
will be presented ranging from the least complex (algebraic) to the most com-
plex (dierential transport) means of specifying the turbulent eddy viscosity
t
.
[1] Linear and nonlinear eddy viscosity models 13
3.1 Zero-equation models
The zero-equation model is so named because the eddy viscosity required in
the turbulent stress-strain relationship is dened from an algebraic relationship
rather than from a dierential one. The earliest example of such a closure is
Prandtls mixing-length theory (Prandtl 1925). By analogy with the kinetic
theory of gases, Prandtl assumed the form for the turbulent eddy viscosity
in a plane shear ow with unidirectional mean ow U
1
(x
2
) = U(y) and shear
stress
12
=
xy
=
t
dU/dy. The eddy viscosity was assumed to have the
form
t
= l
2
dU
dy
, (3.4)
where l is the mixing length that requires specication for each ow under
consideration. In a free shear ow, the mixing length would be a character-
istic measure of the width of the shear layer. In a planar wall-bounded ow,
the mixing length l in the near-wall region would be proportional to the dis-
tance from the wall. These relationships, though simple, give rise to signicant
insights about the structure of turbulent ows. In the case of wall-bounded
ows, the law of the wall and the structure of the outer layer of the boundary-
layer ow can be deduced. Several texts and reviews in the literature provide
an insightful description of the physical and mathematical basis for this type
of modeling. These include Tennekes and Lumley (1972), Reynolds (1987),
Speziale (1991), and Wilcox (1998).
Two of the most popular and versatile algebraic models are the Cebeci
Smith (see Cebeci and Smith 1974) and the BaldwinLomax (see Baldwin and
Lomax 1978) models. Even though the original development of these mod-
els was motivated by application to compressible ows, no explicit account
was taken of compressibility eects. Density eects are simply accounted for
through a variable-mean-density extension of the incompressible formulation
(
t
=
t
). These are two-layer mixing-length models that have an inner layer
eddy viscosity given by
CebeciSmith:
ti
= l
2
_
U
i
x
j
U
i
x
j
_1
2
(3.5)
BaldwinLomax:
ti
= l
2
_
2W
ij
W
ij
, (3.6)
where W
ij
is the rotation tensor
W
ij
=
1
2
_
U
i
x
j
U
j
x
i
_
, (3.7)
and
_
2W
ij
W
ij
represents the magnitude of the vorticity. An outer layer eddy
viscosity is given by
CebeciSmith:
to
= 0.0168U
e
F
K
(y; ) (3.8)
BaldwinLomax:
to
= 0.0269F
wk
F
K
(y; y
m
/0.3). (3.9)
14 Gatski and Rumsey
The mixing length is dened similarly in both models for zero-pressure-grad-
ient ows. That is,
l = y
_
1 e
y
+
/A
+
_
, (3.10)
where = 0.41 is the von Karman constant, A
+
= 26 is the Van Driest damp-
ing coecient, and y
+
is the distance from the wall in wall units (u
y/).
In the expressions for the outer layer eddy viscosity, is the boundary-layer
thickness,
_
6
_
1
, (3.11)
and the wake function F
wk
is given by
F
wk
= min
_
y
m
F
m
; y
m
U
2
dif
/F
m
_
(3.12)
with
F
m
=
1
_
max
y
(l
_
2W
ij
W
ij
)
_
. (3.13)
In the above, y
m
is the distance from the body surface where F
m
occurs, is
the boundary-layer thickness in the CebeciSmith model, and is y
m
/0.3
in the BaldwinLomax model. The quantity U
dif
is the dierence between
the maximum and minimum total velocity in the prole. Unlike the Cebeci
Smith model, the BaldwinLomax model does not need to know the location
of the boundary-layer edge. As equation(3.13) suggests, the BaldwinLomax
model bases the outer layer length scale on the vorticity in the layer rather
than on the displacement thickness, as in the CebeciSmith model. Extensions
and generalizations to more complex ows can be found in Cebeci and Smith
(1974), Degani and Schi (1986), and Wilcox (1998).
A disadvantage of the CebeciSmith and BaldwinLomax turbulence mod-
els is that they possess an inherent dependency on the grid structure: quan-
tities are evaluated and searched for along grid lines normal to walls. This
dependency can be problematic for unstructured grids or for multiple-zone
structured grids. Also, it has been shown that these models, in their original
form, generally do not predict separated ows well. For example, when strong
shock-induced separation is present, these models tend to predict the shock
position too far aft. However, the BaldwinLomax model with the Degani-
Schi modication is often still used in industry for three-dimensional vortical
[1] Linear and nonlinear eddy viscosity models 15
ow applications because other models (including some of the one- and two-
equation eld equation models) can diuse vortices excessively.
3.2 Half-equation models
The motivation for the development of the JohnsonKing model (Johnson and
King 1985) was primarily the need to solve a particular class of ows turbu-
lent boundary layer ows in strong adverse pressure gradients rather than
the development of a universal model. The model was developed to account
for strong history eects that were observed to be characteristic of turbulent
boundary layers subjected to rapid changes in the streamwise pressure gra-
dient. Johnson and King felt that the simple algebraic models (as outlined
in Section 3.1) could be modied suciently, without recourse to the more
elaborate dierential transport formulations (such as the two-equation formu-
lation to be discussed in Section 3.4), to better predict ows with massive
separation. Thus, advection eects were deemed essential, whereas turbulent
transport and diusion eects were assumed to have much less importance.
This level of closure derives its name somewhat subjectively because an
ordinary dierential equation is solved instead of a partial dierential equa-
tion. Nevertheless, this level of closure does generalize the algebraic models
by specifying a smooth functional behavior for the eddy viscosity across the
boundary layer and by accounting in a limited way for history (relaxation)
eects by solving a transport equation for the maximum shear stress. Since
the inception of the JohnsonKing model, it has undergone some modication
(Johnson 1987, Johnson and Coakley 1990) to improve its predictive capabili-
ties for a wider class of ows, and in particular, for compressible ows. For the
present purpose, only the simpler incompressible formulation will be outlined.
The JohnsonKing model is also a two-layer model; however, in this model,
the eddy viscosity changes in a prescribed functional manner from the inner
layer form to the outer layer form. This functional form is given by (Johnson
and King 1985)
t
=
to
[1 exp (
ti
/
to
)] . (3.14)
In the later form of the model (Johnson and Coakley 1990), which was also
used in the solution of transonic ow problems, this functional dependency
was based on a hyperbolic tangent function. The inner layer eddy viscosity is
given by
ti
= l
2
_
xy
[
m
y
, (3.15)
where l is the mixing length dened in equation(3.10) with A
+
= 15, and the
subscript m denotes maximum value along a grid coordinate line normal to
a solid wall surface. In zero-pressure-gradient, two-dimensional ows in which
the law of the wall holds, this expression for
ti
corresponds to the Cebeci
Smith inner layer eddy viscosity given in equation(3.5).
16 Gatski and Rumsey
The outer layer eddy viscosity is given by
to
= 0.0168U
e
F
K
(y; )(x), (3.16)
which is the CebeciSmith form, equation (3.8), with the addition of the factor
(x) that accounts for streamwise evolution of the ow. At each streamwise
station, (x) is adjusted so that the relation
t
[
m
=
xy
[
m
U/y[
m
(3.17)
is satised. The remaining quantity that is needed is
xy
[
m
m
, and this
is determined from a transport equation for the shear stress
xy
. Unlike con-
ventional Reynolds-stress closures in which the transport equation for the
turbulent shear stress contains modeled pressure-strain correlations and tur-
bulent transport terms, this turbulent shear stress equation is extracted from
the turbulent kinetic energy equation (cf. equation(3.30)) by assuming that
the shear stress anisotropy b
12
=
xy
/2k = 0.125 is constant at the point of
maximum shear. The log-layer of an equilibrium turbulent boundary layer ow
is a constant stress layer; therefore, the assumption used here is not without
merit in an equilibrium ow. It is interesting to see that such an assumption
does not adversely impact the model performance for the class of separated
ows for which it was developed. If the viscous diusion eects are neglected,
the evolution equation for
m
is
U
m
d
m
dx
= b
12
_
meq
m
_
m
L
m
C
dif
3/2
m
(0.7 y
m
)
_
1
1/2
(x)
_
, (3.18)
where
meq
is the equilibrium value ((x) = 1) for the shear stress, C
dif
= 0.5
for (x) 1 and zero otherwise, and L
m
is the dissipation length scale given
by
L
m
= y, y
m
/ 0.09/ (3.19)
L
m
= 0.09, y
m
/ > 0.09/. (3.20)
Because (x) is not known a priori at each streamwise station, it is necessary
to iterate on the equation set at each station to determine its value.
While the discussion here has focused on two-dimensional ows, extensions
have been proposed for three-dimensional ows (e.g., Savill et al. 1992) which
have also yielded good ow eld predictions.
The JohnsonKing model suers from the same disadvantage as the Cebeci
Smith and BaldwinLomax models: it relies on the grid structure because
quantities are evaluated and searched for along lines normal to walls. For
this reason, the model has received less attention in the last decade with the
increased use of unstructured and multiple-zone structured grids, for which
eld-equation turbulence models are more ideally suited.
[1] Linear and nonlinear eddy viscosity models 17
3.3 One-equation models
Up to this point, both the zero- and half-equation models have focused on
the specication of an eddy viscosity (which is the underlying basis of the
development of single-point closure schemes) rather than on a specication of
either a turbulent velocity or length scale individually. At the one-equation
level of closure, a transport equation is introduced, which in the earliest mod-
els that date back to Prandtl was for the turbulent velocity scale (turbulent
kinetic energy), with an algebraic prescription for the turbulent length scale.
Modern-day approaches have evolved beyond this formulation to the solution
of transport equations for the turbulent Reynolds number or the turbulent
eddy viscosity (velocity scale length scale). Some of these formulations will
be discussed here, and the interested reader can also refer to the text by Wilcox
(1998) for additional information.
Spalart and Allmaras (1994) devised a one-equation model based primarily
on empiricism and on dimensional analysis arguments. Unlike the zero- and
half-equation models discussed previously, this one-equation model is local;
that is, the equation at one point does not depend on the solution at other
points. Therefore, it is easily usable with any type of grid: structured or un-
structured, single block, or multiple blocks. The eddy viscosity relation is given
by
t
= f
v1
, (3.21)
where f
v1
=
3
/(
3
+ c
3
v1
), and = /. The variable is determined by
using the transport equation
D
Dt
= c
b1
(1 f
t2
)
S +
1
x
k
_
( + )
x
k
_
+
c
b2
_
x
k
_
2
(c
w1
f
w
c
b1
2
f
t2
__
d
_
2
(3.22)
with auxiliary relations
S =
_
2W
ij
W
ij
+
2
d
2
f
v2
(3.23)
f
v2
= 1
1 +f
v1
(3.24)
f
w
= g
_
1 +c
6
w3
g
6
+c
6
w3
_
1/6
=
_
g
6
+c
6
w3
1 +c
6
w3
_
1/6
(3.25)
g = r +c
w2
(r
6
r) (3.26)
r =
S
2
d
2
(3.27)
f
t2
= c
t3
exp(c
t4
2
), (3.28)
where d is the minimum distance to the nearest wall. The closure coecients
are given by: c
b1
= 0.1355, c
b2
= 0.622, = 2/3, = 0.41, c
w1
= c
b1
/
2
+(1 +
18 Gatski and Rumsey
c
b2
)/, c
w2
= 0.3, c
w3
= 2, c
t3
= 1.2, c
t4
= 0.5, and c
v1
= 7.1. Although not
discussed here, Spalart and Allmaras (1994) also developed an additional term
that is used to trip the solution from laminar to turbulent at a desired location.
This feature may be important as the subsequent downstream predictions can
critically depend on the appropriate choice for the onset of turbulence.
Over the years since its introduction, the SpalartAllmaras model has be-
come popular among industrial users due to its ease of implementation and
relatively low cost. Even though this one-equation level of closure is based on
empiricism and dimensional analysis, with characterizing ow features usually
accounted for on a term-by-term basis using phenomenological based models,
it has tended to perform well for a wide variety of ows. As the study by Shur
et al. (1995) has shown, the model can even outperform some two-equation
models in separating and reattaching ows. Recently, Spalart and Shur (1997)
have developed a rotation function, which multiplies the production term and
sensitizes the SpalartAllmaras model to the eects of rotation and curvature.
This function is based on the rate of change of the principal axes of the strain
rate tensor.
Other contemporary one-equation models using the eddy viscosity have been
proposed. One is the Gulyaev et al. (1993) model, which is an improved version
of the model developed by Sekundov (1971). It has been shown in the Russian
literature to solve a variety of incompressible and compressible ow problems
(see Gulyaev et al. 1993 for selected references). Another is the model by Bald-
win and Barth (1991), that has its origins in the k- two-equation formulation
and was a precursor to the SpalartAllmaras model.
In their original forms, both the SpalartAllmaras and BaldwinBarth mod-
els are known to cause excessive diusion in regions of three-dimensional vor-
tical ow. Dacles-Mariani et al. (1995) proposed the use of a modied form
of the production term; rather than basing it on the magnitude of vorticity
(
_
2W
ij
W
ij
) alone, the following functional form is assumed:
_
2W
ij
W
ij
+ 2 min(0,
_
2S
ij
S
ij
_
2W
ij
W
ij
). (3.29)
This method was shown to help for a particular application using the Baldwin
Barth model, but it is not a universally accepted x. The problem of excessive
diusion in some vortical ow applications by these models, in general, still
persists.
3.4 Two-equation models
While the previous closure models discussed have focused on the specication
of a turbulent eddy viscosity to be used directly in the RANS equation (3.3),
the two-equation level of closure attempts to develop transport equations for
both the turbulent velocity and length scales of the ow. Many variations on
[1] Linear and nonlinear eddy viscosity models 19
this approach exist, but the most common approaches use the transport equa-
tion for the turbulent kinetic energy for the turbulent velocity scale equation.
On the other hand, the length scale equation has generally been the most con-
troversial element of the two-equation formulation. For the present purposes,
attention will be focused in this subsection on the k- and k- formulations,
where is the turbulent energy dissipation rate and is the dissipation per
unit turbulent kinetic energy.
The turbulent kinetic energy equation k is easily derived from the uctu-
ating momentum equation for u
i
by forming the transport equation for the
scalar product u
i
u
i
/2. The resulting equation can be written as
Dk
Dt
= T +T, (3.30)
where the right-hand side represents the transport of k by the turbulent pro-
duction T =
ik
U
i
/x
k
, the isotropic turbulent dissipation rate, , and the
combined eects of turbulent transport and viscous diusion T. When equa-
tion(3.2) is used, the turbulent production term can also be written in terms
of the eddy viscosity as
T = 2
t
(S
ik
S
ki
) = 2
t
2
, (3.31)
where the velocity gradient tensor is decomposed into the sum of the symmetric
strain rate tensor S
ij
and the antisymmetric rotation rate tensor W
ij
,
2
=
S
ik
S
ki
(or
2
= S
2
in matrix notation), and the trace S
ik
W
ki
= WS =
0. In such a formulation, the behavior of the individual stress components
is governed by the Boussinesq relation given in equation(3.2), which is an
isotropic eddy viscosity relationship. In general, the evolution of the individual
stress components is not isotropically partitioned among the components. For
this eect to be accounted for, higher-order closures are required such as the
nonlinear eddy viscosity models to be discussed later in this chapter or the
Reynolds stress formulation to be discussed in Chapter [2].
Nevertheless, the Boussinesq relation is not without physical foundation.
For example, in (thin) simple shear ow where an equilibrium layer exists, it
is assumed that
xy
=
_
C
k (3.32)
with C
k
_
2
(3.33)
where equation(3.32) has been used. This yields the familiar closure model for
the turbulent shear stress,
xy
= C
k
2
U
y
. (3.34)
20 Gatski and Rumsey
Dimensional analysis considerations dictate that the eddy viscosity
t
be given
by the product of a turbulent velocity scale and a turbulent length scale.
With the velocity scale given by k
1/2
, the remaining task is the development
of the scale variable. In this chapter, two such alternatives are considered.
The rst is the turbulent energy dissipation rate which implies that k
3/2
/
is proportional to the length scale, and the second is the specic dissipation
rate
1
, , which implies that k
1/2
/ is proportional to length scale. Thus, the
eddy viscosity
t
is given by the relation
t
= C
k
2
= C
k, =
k
(3.35)
for the k- two-equation model, and
t
=
k
(3.36)
for the original k- two-equation model. The modeling coecient C
usually
assumes a value of 0.09. (Note that this value is slightly larger than the value
assumed in the derivation of the half-equation model in Section 3.2.) For the
k- formulation, the kinetic energy equation (3.30) is suitably modied by
using the substitution = C
k
_
k
x
j
_
, (3.37)
where the rst term on the right is the viscous contribution, and the second is
the model for the turbulent transport. The coecient
k
is an eective Prandtl
number for diusion, which is taken as a constant in incompressible ows. The
value of
k
is dependent on the particular scale variable used. The resulting
simple form of the modeled turbulent kinetic energy equation is an obvious
appeal of the formulation.
There are several variations to the modeled form of the transport equation
for the isotropic dissipation rate . A rather general expression (Jones and
Launder 1972) from which many of the forms can be derived and which can
be integrated to the wall is given by
D
Dt
=
1
(C
1
T C
2
) +
x
k
__
+
t
_
x
k
_
, (3.38)
1
i.e., dissipation rate of kinetic energy (k) per unit k.
[1] Linear and nonlinear eddy viscosity models 21
where C
1
1.45 is usually xed from calibrations with homogeneous shear
ows, and C
2
is usually determined from the decay rate of homogeneous,
isotropic turbulence ( 1.90). The closure coecient
=
2
_
C
(C
2
C
1
)
. (3.39)
During the late 1990s, the two-equation shear stress transport (SST) model
of Menter (1994), has gained increasing favor among industrial users, due
primarily to its robust formulation and improved performance for separated
ows over traditional two-equation models. One of the primary features of
Menters model is that it is a blend of Wilcoxs original k- formulation near
walls and a k- formulation in the outer region and in free shear ows. Thus,
the model does not have to contend with the problems often encountered by
k- models near walls (see Section 3.5), while it still retains the k- predictive
capabilities in free shear ows.
Since the transport equation for the turbulent kinetic energy k has been
given previously in equation(3.30), only the transport equation for the specic
dissipation rate of turbulence kinetic energy for the SST model is given here:
D
Dt
=
t
T
2
+
x
k
__
+
t
_
x
k
_
+ 2
1 F
1
k
x
k
x
k
. (3.40)
The function F
1
is the blending function that is used to switch between the
k- (F
1
= 1) and the k- (F
1
= 0) formulations,
F
1
= tanh(
4
), (3.41)
where
= min
_
max
_
k
C
d
;
500
d
2
_
;
4
2
k
CD
k
d
2
_
, (3.42)
and CD
k
represents the cross-diusion term (the last term in the equa-
tion (3.40)), limited to be positive and greater than some very small arbitrary
number. In the derivation of the modied equation, Menter neglects a set of
diusion terms that are demonstrated to be small (Menter 1994) and also ne-
glects the molecular viscosity in the cross-diusion term. The model constants
k
,
, , )
T
= F
1
(
k1
,
1
,
1
,
1
)
T
+ (1 F
1
)(
k2
,
2
,
2
,
2
)
T
. (3.43)
The other important feature of the SST model (which represents a departure
from the component k- and k- models) is a modication to the denition
of the eddy viscosity to account for the eect of the transport of the principal
22 Gatski and Rumsey
turbulent shear stress. The denition of the eddy viscosity
t
in the model is
altered from the forms given previously in equation(3.36):
t
=
2b
12
k
max(2b
12
;
_
2W
ij
W
ij
F
2
)
, (3.44)
where b
12
(= 0.155) is the shear stress anisotropy (see Section 3.2). The blend-
ing function F
2
is given by
F
2
= tanh(
2
2
), (3.45)
where
2
= max
_
2
k
C
d
;
500
d
2
_
. (3.46)
Without the modied form of equation(3.44), most k- and k- linear eddy
viscosity models have been generally found to yield poor results for separated
ows.
The constants for Menters SST model are given by:
k1
= 1.17647,
1
= 2,
1
= 0.075,
k2
= 1,
2
= 1.16822, and
2
= 0.0828. The constant
1
is a
function of
1
and
1
whereas
2
is a function of
2
and
2
as follows:
1,2
=
1,2
/C
2
/(
1,2
_
C
). (3.47)
Notice that the value of
k1
has been recalibrated by Menter from its original
(Wilcox k- model) value of 2 to recover the correct at-plate log-law behavior
when using the modied eddy viscosity equation(3.44). The other coecients
1
,
1
, and
1
are the same as those in Wilcoxs original model. The constants
k2
,
2
,
2
, and
2
have a direct correspondence with the k- coecients:
k2
=
k
2
=
(3.48)
2
= C
(C
2
1)
2
= C
1
1. (3.49)
Menter uses the LaunderSharma (1974) coecients: C
= 0.09, C
1
= 1.44,
C
2
= 1.92, = 0.41,
k
= 1, with
2
/k, is singular at the wall since is
nite, and the turbulent kinetic energy k = 0. There has been an extensive list
of near-wall modications over the last two decades for both the eddy viscos-
ity and the transport equation for the eddy viscosity. The near-wall turbulent
eddy viscosity has taken the form
t
= f
k
2
, (3.52)
where f
_
x
k
_
+E (3.53)
= +D, (3.54)
24 Gatski and Rumsey
where f
1
is a damping function, f
2
is used to ensure that the destruction
term is nite at the wall, and D and E are additional terms added to better
represent the near-wall behavior. A partial list of the various forms for these
functions can be found in Patel et al. (1985), Rodi and Mansour (1993), and
Sarkar and So (1997). While the list is not all inclusive, it does provide the
functional forms which are used today for these near-wall functions.
As seen in Section 3.4, the SST model solves the near-wall problem by
switching on the k- form near the wall. Another alternative to the in-
troduction of damping functions is the elliptic relaxation approach that was
rst proposed by Durbin (1991). This approach has been extended to a full
Reynolds stress closure Durbin (1993a); however, in the context of the current
chapter, the description of the approach will be limited to the two-equation
k- formulation.
In the context of a linear eddy viscosity framework, this is a three-equation
model for the turbulent kinetic energy k, turbulent dissipation rate , and the
normal stress component
22
. (The model has been referred to as the kv
2
or v
2
f model and is based on an elliptic relaxation approach. The notation
used here will be slightly dierent, in keeping with our attempt to have a
unied notation throughout the chapter.) A major assumption of the model
is that the eddy viscosity
t
should be given by
t
= C
22
T, (3.55)
where T is the applicable characteristic timescale of the ow in the proximity
of the wall
T = max
_
, 6
_
_
1/2
_
(3.56)
and the coecient C
1
T C
2
) +
x
k
__
+
t
_
x
k
_
(3.57)
to account for the variability of the characteristic timescale. With the exception
of the production-of-dissipation rate coecient C
1
, the closure coecients are
assigned values close to the standard ones (Durbin 1993b, 1995), but C
1
has
assumed dierent non-constant values (Durbin 1993b, 1995) to optimize the
predictive capability of the model.
The modeled equation for
22
is approximated from the Reynolds stress
transport equation and is given by
D
22
Dt
= kf
22
22
k
+
x
k
__
+
t
k
_
22
x
k
_
, (3.58)
[1] Linear and nonlinear eddy viscosity models 25
where the function f
22
is, in general, obtained from
L
2
2
f
22
f
22
=
22
(3.59)
as discussed in Chapters [2] and [4]. The characteristic length scale L is dened
in a manner analogous to the timescale so that
L = C
L
max
_
_
k
3/2
, C
_
1/4
_
_
, (3.60)
where C
L
0.25 and C
ij
=
2
3
k
ij
+
n
n=1
n
T
(n)
ij
. (4.1)
26 Gatski and Rumsey
Since one of the advantages of the NLEVMs is to be able to capture some
eects of the stress anisotropies that occur at the dierential second-moment
level of closure, it is helpful to recast some of the equations in terms of the
stress anisotropy b
ij
given by
b
ij
=
ij
2k
ij
3
. (4.2)
For example, in terms of b
ij
, equation(4.1) can then be rewritten as
b
ij
=
n
n=1
n
T
(n)
ij
, (4.3)
where T
(n)
ij
are the tensor bases and
n
are the expansion coecients which
need to be determined.
As was shown in Section 3, the linear EVMs, through the Boussinesq approx-
imation, couple to the RANS equations through a simple additive modication
to the diusion term (see equation(3.3)). In the case of nonlinear EVMs, this
coupling can be more complex. The coupling can be either through the direct
use of equation(4.1) in equation(3.1) or through a modied form of equa-
tion(3.3) given by
DU
i
Dt
+
1
p
x
i
=
x
j
_
( +
t
)
U
i
x
j
_
+S.T. (4.4)
where S.T. are nonlinear (source) terms from the tensor representation (4.1).
The degree of complexity associated with the nonlinear source terms is de-
pendent on both the number and form of the terms chosen for the tensor
representation.
The choice of the proper tensor basis is, of course, dependent on the func-
tional dependencies associated with the Reynolds stress
ij
or the correspond-
ing anisotropy tensor b
ij
. As seen from the transport equation for the Reynolds
stresses (e.g. Speziale 1991), the only dependency on the mean ow is through
the mean velocity gradient. Thus, it has been generally assumed in develop-
ing turbulent closure models for the Reynolds stresses, that, in addition to
the functional dependency on the turbulent velocity and length scales, the
dependence on the mean velocity gradient be included as well. The turbu-
lent velocity scale is usually based on the turbulent kinetic energy k and the
turbulent length scale on the variable used in the corresponding scale equa-
tion (which for the purposes here will be the isotropic turbulent dissipation
rate ). The continuum mechanics community has dealt with such questions
on tensor representations for several decades (e.g., Spencer and Rivlin 1959).
Within this context, the stress anisotropy tensor is considered here with the
functional dependencies
b
ij
= b
ij
(k, , S
kl
, W
kl
), (4.5)
[1] Linear and nonlinear eddy viscosity models 27
where the dependence on the mean velocity gradient has been replaced by the
equivalent dependence on the strain rate tensor (see equation(2.6)) and the ro-
tation rate tensor (see equation(3.7)). (In dealing with tensor representations,
it is sometimes better, for notational convenience, to use matrix notation to
eliminate the cumbersome task of accounting for several tensor indices. For
this reason, both tensor and matrix notation will be used in describing the
models in this and subsequent sections.) Equation (4.5) can be rewritten in
matrix notation as
b = b(k, , S, W). (4.6)
In the case of fully three-dimensional mean ow, a symmetric, traceless tensor
function b of a symmetric tensor (S) and an antisymmetric tensor (W) can
be represented as an isotropic tensor function of the following ten (integrity)
tensor bases T
(n)
(= T
(n)
ij
):
T
(1)
= S T
(6)
= W
2
S +SW
2
2
3
SW
2
I
T
(2)
= SWWS T
(7)
= WSW
2
W
2
SW
T
(3)
= S
2
1
3
S
2
I T
(8)
= SWS
2
S
2
WS
T
(4)
= W
2
1
3
W
2
I T
(9)
= W
2
S
2
+S
2
W
2
2
3
S
2
W
2
I
T
(5)
= WS
2
S
2
W T
(10)
= WS
2
W
2
W
2
S
2
W.
(4.7)
The expansion coecients
n
associated with this representation can, in
general, be functions of the invariants of the ow
n
=
n
(S
2
, W
2
, k, , Re
t
), (4.8)
where S
2
= S
ij
S
ji
, W
2
= W
ij
W
ji
are the strain rate and rotation rate
invariants, respectively, and Re
t
= k
2
/ is the turbulent Reynolds number. Of
course, a smaller number of terms could be used for the representation, such as
in three-dimensions where the minimum number is ve to have an independent
basis; however, the expansion coecients will then be more complex (Jongen
and Gatski 1998) and could possibly be singular. An advantage of using the
full integrity basis is that the expansion coecients will not be singular.
The linear term T
(1)
is the strain rate S, as in the linear EVM case, and
its coecient
1
is now the turbulent eddy viscosity
t
, which is used in equa-
tion(4.4). The nonlinear source terms are the remaining terms T
(n)
(n 2)
in the polynomial expansion.
In the remainder of this section, the development of the models will be
categorized based on the methodology used to determine the expansion coef-
cients
n
. First, what is usually termed nonlinear eddy viscosity models are
discussed. These are models in which a polynomial expansion is assumed that
is a subset of equation(4.7), and the expansion coecients are determined
based on calibrations with experimental or numerical data and physical con-
straints. Second, what is usually termed algebraic stress models or algebraic
28 Gatski and Rumsey
Reynolds stress models are discussed. These are models in which a polyno-
mial expansion is, once again, assumed from equation(4.7), but the expansion
coecients are derived in a mathematically consistent fashion from the full
dierential Reynolds stress equation. In both cases, an explicit tensor repre-
sentation for b is obtained in terms of S and W.
4.1 Quadratic and cubic tensor representations
In this subsection, a few examples of nonlinear eddy viscosity models repre-
sented by the tensor expansions in equations (4.1) or (4.3) are discussed. The
models examined include expansions where both terms quadratic (n = 2) and
cubic (n = 3) in the mean strain rate and rotation rate tensors are retained.
Each of these examples (while not all inclusive) provide insight into the variety
of assumptions required in identifying the expansion coecients
n
needed in
the algebraic representation of the Reynolds stresses.
As a rst example of a nonlinear eddy viscosity model using an explicit
representation for the Reynolds stress anisotropy, consider the quadratic model
proposed by Speziale (1987). Speziales approach, while also motivated by the
need to include Reynolds stress anisotropy eects into a linear eddy viscosity
type of formulation, diered in its development of the tensor representation for
the Reynolds stress anisotropy. Speziale assumed that the anisotropy tensor
b
ij
was of the form
b
ij
= b
ij
_
k, , S
kl
,
S
kl
t
_
, (4.9)
where
S
kl
t
=
DS
kl
Dt
U
k
x
m
S
ml
U
l
x
m
S
mk
(4.10)
is the Oldroyd convective derivative (e.g., Aris 1989, p. 185). This dependency
on the convective derivative was used to ensure that the nonlinear polyno-
mial approximation would be frame-indierent, in keeping with the frame-
indierent properties of the anisotropy tensor itself. Calibrations were based
on fully developed channel ow predictions using both k-l and k- two-equation
models. The resulting tensor representation (using the strain rate and rotation
rate tensor notation) was
b =
1
S +
2
(SWWS) +
3
_
S
2
1
3
S
2
I
_
+
D
DS
Dt
, (4.11)
where
D
(k, ) was a closure coecient determined from the calibration. Writ-
ten in this form, it can be seen that the introduction of the frame-indifferent
convective derivative simply modies two of the tensor bases given in equa-
tion(4.7). Preliminary validation studies were done for a rectangular duct ow
and for a backstep ow to highlight the improved predictive capabilities of the
nonlinear model over the corresponding linear eddy viscosity k-l and k- forms.
[1] Linear and nonlinear eddy viscosity models 29
Next consider the quadratic model proposed by Shih et al. (1995),
b =
1
S +
2
(SWWS) . (4.12)
The
i
coecients were determined by applying the rapid distortion theory
constraint to rapidly rotating isotropic turbulence, and the realizability con-
straints
0, no sum (4.13)
and
1
3
S
2
I
_
+
3
(SWWS) +
4
_
W
2
1
3
W
2
I
_
(4.15)
+
5
_
W
2
S +SW
2
2
3
SW
2
I
_
+
6
_
WS
2
S
2
W
_
.
(The form given here diers slightly from the form presented in Craft et al.,
although the two representations can be shown to be equivalent.) Calibration
of the closure coecients was based on an optimization over a wide range of
ows. These included plane channel ow, circular pipe ow, axially rotating
pipe ow, fully developed curved channel ow, and impinging jet ows. This
algebraic representation for the Reynolds stresses was then coupled with low-
Reynolds number forms for the kinetic energy and dissipation rate equations
(see equations (3.30) and (3.52)(3.54)). Attempts at extending the model to
ows far from equilibrium have been undertaken and are discussed in Craft
et al. (1997). Other cubic models have been proposed, for example, by Apsley
and Leschziner (1998) and Wallin and Johansson (2000), and the interested
reader is referred to these papers for further details on their development and
application.
30 Gatski and Rumsey
4.2 Algebraic stress models
The identifying feature of algebraic stress models (ASMs) is the technique
used to obtain the expansion coecients
n
. As noted previously, these co-
ecients have a direct relation to the Reynolds stress model used, or more
specically, to the pressure-strain rate correlation model. The algebraic stress
model used here is based on the model originally developed by Pope (1975)
for two-dimensional ows, and later extended by Gatski and Speziale (1993),
to three-dimensional ows. The implementation has since been rened, and
the formulation to be presented is based on recent work by Jongen and Gatski
(1998b).
4.2.1 Implicit algebraic stress model
The starting point for the development of ASMs is the modeled transport
equation for the Reynolds stress anisotropy tensor b
ij
(see Gatski and Speziale
1993) given by
Db
ij
Dt
=
1
2k
_
D
ij
Dt
ij
k
Dk
Dt
_
= b
ij
_
T
k
_
2
3
S
ij
_
b
ik
S
kj
+S
ik
b
kj
2
3
b
mn
S
mn
ij
_
(4.16)
+(b
ik
W
kj
W
ik
b
kj
) +
ij
2k
+
1
2k
_
T
ij
ij
k
T
_
.
where
ij
is the pressure-strain rate correlation, and T
ij
is the combined
eect of turbulent transport and viscous diusion (T = T
ii
/2). While it is
outside the scope of this chapter to discuss the modeling of the pressure-strain
correlation
ij
, it is necessary for the development of the algebraic stress
model to specify a form for the pressure-strain rate model. For the purposes
here, the SSG model (Speziale, Sarkar, and Gatski 1991) will be used, and can
be written in the form
ij
=
_
C
0
1
+C
1
1
T
_
b
ij
+C
2
kS
ij
(4.17)
+ C
3
k
_
b
ik
S
jk
+b
jk
S
ik
2
3
b
mn
S
mn
ij
_
C
4
k (b
ik
W
kj
W
ik
b
kj
) ,
where the closure coecients can, in general, be functions of the invariants
of the stress anisotropy. It should be noted that the functional form given
in equation (4.17) is representative of any linear pressure-strain rate model
which could be used as well. Substituting equation(4.17) into equation(4.16)
[1] Linear and nonlinear eddy viscosity models 31
and rewriting yields
Db
ij
Dt
1
2k
_
T
ij
ij
k
T
_
=
_
b
ij
a
4
+a
3
_
b
ik
S
kj
+S
ik
b
kj
2
3
b
mn
S
mn
ij
_
(4.18)
a
2
(b
ik
W
kj
W
ik
b
kj
) +a
1
S
ij
_
.
The coecients a
i
are directly related to the pressure-strain correlation model
by
a
1
=
1
2
_
4
3
C
2
_
, a
2
=
1
2
(2 C
4
) ,
(4.19)
a
3
=
1
2
(2 C
3
) , a
4
= g,
and
g =
__
C
1
1
2
+ 1
_
T
+
C
0
1
2
1
_
1
=
_
0
T
+
1
_
1
, (4.20)
where C
0
1
= 3.4, C
1
1
= 1.8, C
2
= 0.36, C
3
= 1.25, and C
4
= 0.4. An implicit al-
gebraic stress relation is obtained from the modeled transport equation for the
Reynolds stress anisotropy equation(4.18) when the following two assumptions
rst proposed by Rodi (1976) are made:
Db
ij
Dt
= 0, or
D
ij
Dt
=
ij
k
Dk
Dt
, (4.21)
and
T
ij
=
ij
k
T. (4.22)
Equation(4.21) is equivalent to requiring that the turbulence has reached an
equilibrium state, Db/Dt = 0, and equation(4.22) invokes the assumption
that any anisotropy of the turbulent transport and viscous diusion is pro-
portional to the anisotropy of the Reynolds stresses. Both these assumptions
impose limitations on the range of applicability of the algebraic stress model.
Later in this section, some alternative assumptions will be proposed that will
improve the range of applicability of the ASM.
With these assumptions, the left side of equation(4.18) vanishes, and the
equation becomes algebraic:
b
ij
a
4
+a
3
_
b
ik
S
kj
+ S
ik
b
kj
2
3
b
mn
S
mn
ij
_
(4.23)
a
2
(b
ik
W
kj
W
ik
b
kj
) +a
1
S
ij
= 0,
32 Gatski and Rumsey
or rewritten using matrix notation
1
a
4
b a
3
_
bS +Sb
2
3
bSI
_
+a
2
(bWWb) = R. (4.24)
For linear pressure-strain rate models and an isotropic dissipation rate, it
follows that R = a
1
S. However, the generalization implied by using R is
intended to indicate that the right-hand side of equation(4.24) can contain any
known symmetric, traceless tensor (Jongen and Gatski 1998). Equation(4.24)
has to be solved for b and is an implicit equation. Such an equation can be
solved numerically in an iterative fashion. Unfortunately, such procedures can
be numerically sti, depending on the complexity of the ow to be solved. It
is desirable to obtain an explicit solution to this equation which still retains
its algebraic character. The rst attempt at this was by Pope (1975) who
obtained an explicit solution of equation(4.24) using a three-term basis (cf.
equations (4.3) and (4.7)) for two-dimensional mean ows
b =
1
S +
2
(SWWS) +
3
_
S
2
1
3
S
2
I
_
, (4.25)
where the
i
are scalar coecient functions of the invariants
_
S
2
_
and
_
W
2
_
.
Gatski and Speziale (1993) derived a corresponding expression for three-di-
mensional mean ows which required all ten terms from the integrity basis
given in equation(4.7). A general methodology will now be presented that
allows for the systematic identication of the coecients
i
from an implicit
algebraic equation such as that given in equation(4.24).
4.2.2 Explicit solution
While it is possible to implement the following methodology by using any
number of terms in the tensor representation T
(n)
, it is dicult to obtain
closed form analytic expressions beyond n = 3. Thus, the discussion here is
limited to n = 3, and the three-term basis T
(1)
, T
(2)
, and T
(3)
(exact for
two-dimensional ows) from equation(4.7) is used for the representation, that
is,
b =
3
n=1
n
T
(n)
, (4.26)
with the same three-term tensor basis T
(n)
shown in equation(4.25).
Equation(4.24) can be solved `a la Galerkin by projecting this algebraic
relation onto the tensor basis T
(m)
itself. For this solution, the scalar product
of equation(4.24) is formed with each of the tensors T
(m)
, (m=1, 2, . . . , n).
This procedure leads to the following system of equations:
n
n=1
n
_
1
a
4
(T
(n)
, T
(m)
) 2a
3
(T
(n)
S, T
(m)
) + 2a
2
(T
(n)
W, T
(m)
)
_
= (R, T
(m)
), (4.27)
[1] Linear and nonlinear eddy viscosity models 33
where the scalar product is dened as (T
(n)
, T
(m)
) = T
(n)
T
(m)
. In a more
compact form,
n
n=1
n
A
nm
= (R, T
(m)
), (4.28)
where the n n matrix A is dened as
A
nm
1
a
4
(T
(n)
, T
(m)
) 2a
3
(T
(n)
S, T
(m)
) + 2a
2
(T
(n)
W, T
(m)
). (4.29)
For a two-dimensional mean ow eld, the matrix A is
A
nm
=
_
1
a
4
2
2a
2
1
3
a
3
4
2a
2
2
a
4
2
0
1
3
a
3
4
0
1
6a
4
4
_
_
, (4.30)
which, when inverted, leads to the following expressions for the representation
coecients
1
=
a
4
2
(RS + 2a
2
a
4
RWS 2a
3
a
4
RS
2
_
, (4.31)
2
= a
4
_
a
2
1
+
RWS
2
_
, (4.32)
3
= a
4
_
2a
3
1
+
6RS
2
4
_
, (4.33)
where
0
=
_
1
2
3
a
2
3
a
2
4
2
+ 2a
2
2
a
2
4
2
_
, and
2
(=
_
W
2
_
/
_
S
2
_
). The ow
parameter
2
is a dimensionless variable that is useful for characterizing the
ow (Astarita 1979, Jongen and Gatski 1998a); for example, for a pure shear
ow
2
= 1, whereas for a plane strain ow
2
= 0. This set of equations is
the general solution valid for two-dimensional mean ow and for any arbitrary
(symmetric traceless) tensor R.
As noted previously, when a linear pressure-strain correlation model is as-
sumed, as well as an isotropic dissipation rate, then R = a
1
S. This expression
leads to a right-hand side for equation(4.28) proportional to
(R, T
(m)
) =
_
_
RS
2RWS
RS
2
_ =
_
_
a
1
2
0
0
_
_. (4.34)
Using equation(4.34) in equations (4.31)(4.33) and substituting into equa-
tion(4.26) leads to the representation for the Reynolds stress tensor
=
2
3
kI + 2k
1
_
S +a
2
a
4
(SWWS) 2a
3
a
4
_
S
2
1
3
S
2
I
__
. (4.35)
34 Gatski and Rumsey
In equation(4.31), a
4
is a function of T/ (and therefore
1
). Gatski and
Speziale (1993) simplied this expression by assuming the coecient g (see
equation(4.20)) and therefore T/ to be constant in the analysis. However,
we follow the approach proposed by Ying and Canuto (1996) and Girimaji
(1996), in which the value of g is not xed; the variation of the production-
to-dissipation-rate ratio in the ow is accounted for in the formulation. It is
easily shown that the production-to-dissipation rate ratio is given by
T
= 2 bS , (4.36)
and that the invariant bS is directly related (Jongen and Gatski 1998a) to
the coecient
1
appearing in the tensor representation through
bS =
1
2
. (4.37)
From equations (4.19) and (4.20), the coecient a
4
can then be written as
a
4
=
_
1
2
0
_
1
. (4.38)
The dependency of a
4
on the production-to-dissipation rate ratio through
1
makes both sides of equation (4.31) functions of
1
. This dependency results
in a cubic equation for
1
given by
2
0
3
1
2
1
1
4
4
2
_
2
1
2
2
0
RS 2
2
2
_
a
2
3
3
2
a
2
2
__
1
+
1
4
6
1
RS + 2
_
a
2
RWS a
3
RS
2
_
_
= 0. (4.39)
The expansion coecients of the nonlinear terms,
2
and
3
, retain the same
functional dependency on
1
as before. When expressed in terms of the pro-
duction-to-dissipation rate ratio with R = a
1
S, equation(4.39) can be shown
(Jongen and Gatski 1998a) to be equivalent to earlier results (Ying and Canuto
1996, Girimaji 1996). Previously (Ying and Canuto 1996, Girimaji 1996), the
selection of the proper root for the solution of equation(4.39) was done on the
basis of continuity arguments. Here, the proper choice for the solution root is
based on the asymptotic analysis of Jongen and Gatski (1999). It was found
that the root with the lowest real part leads to the correct choice for
1
.
The explicit tensor representation given in equation (4.35) with
1
,
2
, and
3
determined by using equations (4.39), (4.32), and (4.33), respectively, is
coupled with a k- two-equation model:
Dk
Dt
= T +
x
k
__
+
t
k
_
k
x
k
_
, (4.40)
D
Dt
= C
1
k
T f
C
2
2
k
+
x
k
__
+
t
_
x
k
_
, (4.41)
[1] Linear and nonlinear eddy viscosity models 35
where is the kinematic viscosity and
t
= C
k is an equilibrium eddy
viscosity. Also,
f
=
_
1 exp
_
Re
k
10.8
__
, Re
k
=
k
1/2
d
, (4.42)
k
= 1.0,
=
2
_
C
(C
2
C
1
)
, = 0.41,
C
1
= 1.44, C
2
= 1.83, C
= 0.096, (4.43)
and d is the distance to the nearest wall. Note in equation(4.35) that
1
=
C
, where C
t
= C
= C
(near 0.09). On the other hand, this explicit algebraic stress solution
has the eect of yielding a variable C
1
3
S
2
I.
4.2.3 Curvature eects and the equilibrium assumption
In a recent study of two-dimensional ow in a U-bend (Rumsey et al. 1999),
the formulation of the explicit algebraic stress model just discussed was unable
to correctly predict the turbulence second-moments in the strongly curved re-
gions of the ow. However, it was shown that a second-moment closure model
could correctly predict the behavior of the turbulence. Since the ASM is closely
coupled with the corresponding second-moment model, the poor performance
of the ASM could be traced to the two underlying assumptions used in the de-
velopment, that is, equations (4.21) and (4.22). The poor performance was
attributed to the equilibrium assumption shown in equation(4.21). It was
found from the second-moment closure computation of the ow that some
components of Db
ij
/Dt were not zero and that their magnitudes were as large
as the components of a
1
S (see equation(4.24)). Since the assumption aecting
the turbulent transport and viscous diusion, equation (4.22), was eectively
satised throughout the ow, any deciency in algebraic stress model predic-
tion can be attributed directly to the imposition of the equilibrium assump-
tion equation(4.21). Similar deciencies have been recognized previously. For
example, Fu et al. (1988) showed that in free shear ows with swirl the ap-
proximated advection terms were independent of swirl whereas the exact form
in the second-moment equation was not. Thus, the equilibrium assumption
precluded an accurate representation of a key dynamic feature.
If the equilibrium condition Db
ij
/Dt = 0 is not correct, then can another
condition on Db
ij
/Dt be chosen such that the results of the second-moment
closure formulation are replicated? For the answer, it is only necessary to
recall that the turbulence second-moment equations are not frame-indierent
36 Gatski and Rumsey
and that Db
ij
/Dt = 0 in one frame does not imply the same in another frame
(D/Dt is only frame-indierent under an extended Galilean transformation,
Speziale 1979, 1998). The answer, then, is to nd a coordinate frame in which
to apply the equilibrium condition Db
ij
/Dt = 0 (Gatski and Jongen 2000).
Under a Euclidean transformation (Speziale 1979), the transformation of the
turbulence anisotropy tensor b
pq
from a Cartesian base system is simply given
by
b
pq
= X
i
p
b
ij
X
j
q
, (4.44)
where X(t) = x/x is a proper orthogonal tensor. It then follows that the
material derivative of equation(4.44) yields
D
b
pq
Dt
= X
i
p
Db
ij
Dt
X
j
q
+
D
Dt
_
X
i
p
_
b
ij
X
j
q
+ X
i
p
b
ij
D
Dt
_
X
j
q
_
. (4.45)
(Note that in a general Euclidean transformation, a constant shift of the time
variable is allowed. This shift is neglected here because it plays no relevant
role in the analysis.) Equation(4.45) shows that the material derivative of
the Reynolds stress anisotropy is not frame-indierent under arbitrary time-
dependent rotations. Since the equilibrium condition Db
ij
/Dt = 0 is a xed
point of equation(4.18) (neglecting the contribution from the turbulent trans-
port and viscous diusion), this result shows that the xed point is also not
frame-indierent. It is assumed at this point that equation(4.45) has trans-
formed to a coordinate frame in which the equilibrium condition D
b
pq
/Dt = 0
does hold so that equation(4.45) can be rewritten as
Db
ij
Dt
= b
ik
kj
ik
b
kj
, (4.46)
where
ij
= X
k
i
D
Dt
_
X
j
k
_
(4.47)
and
ij
is a tensor related to the rate of rotation between the barred and
unbarred (Cartesian) systems. The equilibrium assumption given in equa-
tion(4.21) can now be replaced with equation(4.46), and the resulting implicit
algebraic stress equation is (cf. equation(4.23))
b
ij
a
4
+a
3
_
b
ik
S
kj
+ S
ik
b
kj
2
3
b
mn
S
nm
ij
_
(4.48)
a
2
_
b
ik
W
kj
W
ik
b
kj
_
+a
1
S
ij
= 0,
where
W
ij
= W
ij
+
1
a
2
ij
. (4.49)
The result in equation(4.49) shows that by accounting for curvature eects
through a modication of the condition on Db
ij
/Dt, the resulting implicit
[1] Linear and nonlinear eddy viscosity models 37
algebraic stress equation (in the Cartesian base frame) is only altered through
a change to the mean rotation rate tensor. At this point in the analysis, we
have not yet specically identied the transformation X(t) between the two
systems. The choice for the transformed (barred) system becomes clearer after
the traditional analogy between curvature and rotation is exploited.
Two ows which have been studied by using the algebraic stress formulation
and which t into the general formulation just discussed are rotating homo-
geneous shear and fully developed rotating channel ow. Both of these have
rigid-body rotation about the axis (z-axis) perpendicular to the plane of
shear (x,y plane). Of relevance here is the formulation used in solving these
problems. In both cases, the ow elds are described in the noninertial (barred)
frames where the condition D
b
ij
/Dt = 0 holds exactly. Since equation(4.48)
is an implicit equation for b
ij
in the inertial (base) frame, it needs to be trans-
formed to the noninertial (barred) frame which is given by
b
ij
a
4
+a
3
_
b
ik
S
kj
+ S
ik
b
kj
2
3
b
mn
S
mn
ij
_
(4.50)
a
2
_
b
ik
W
kj
W
ik
b
kj
_
+a
1
S
ij
= 0,
where now
W
ij
= W
ij
ijr
r
+
1
a
2
ij
,
ij
=
_
DX
k
i
Dt
_
X
j
k
(4.51)
and
r
= (0, 0, ). The frame-invariance properties of the strain-rate tensor
and the stress anisotropy equation(4.44) have been used, and the correspond-
ing lack of frame-invariance of the rotation-rate tensor is shown by the ap-
pearance of the term
ijr
r
. Since the noninertial eects are imposed through
a rigid-body rotation perpendicular to the plane of shear, the tensor
ij
is
simply related to the rotation rate
r
by
ij
=
ijr
r
(4.52)
so that
W
ij
= W
ij
_
1 +
1
a
2
_
ijr
r
, (4.53)
where
ijr
is the permutation tensor. Equation(4.53) is the intrinsic rotation
rate tensor and is the form used in previous algebraic stress formulations of
rotating homogeneous shear and rotating channel ow.
With this example, the role and description of
ij
, as a tensor related to
the rate of rotation between the barred and unbarred systems, becomes clear.
The issue of relative rotation rate also arises in the study of non-Newtonian
constitutive equations (e.g., Schunk and Scriven 1990, Souza Mendes et al.
1995). There, the measure is based on the principal axes of the strain rate
38 Gatski and Rumsey
tensor, which are a mutually orthogonal set of axes that rotate at the angular
velocity associated with the rotation rate tensor
ij
. In this coordinate frame,
it will be assumed that the condition D
b
ij
/Dt = 0 will hold. Spalart and Shur
(1997) also used the principal axes frame of reference to account for system
rotation and curvature in sensitizing a one-equation model.
In the principal axes coordinate frame, the transformation matrix X(t) can
be dened through the relation between the unit vectors of the xed and
principal axes system
e
(i)
k
= X
i
j
e
(j)
k
, (4.54)
where superscript (i) (= 1, 2, 3) is the particular unit vector, and subscript
k(= 1, 2, 3) is a component of the unit vector (i). Since in this analysis, the
xed (unbarred) system is a Cartesian system where e
(j)
k
=
j
k
, the unit vectors
in the principal axes frame are simply given by
e
(i)
k
= X
i
k
. (4.55)
When equation(4.55) is substituted into equation(4.47), the eigenvectors in
the principal axes frame can then be expressed in terms of
ij
,
ij
= e
(k)
i
De
j
(k)
Dt
. (4.56)
With the specication of
ij
in equation(4.56), an explicit tensor represen-
tation for the implicit algebraic equation for b
ij
, equation(4.48), can be ob-
tained. In contrast to the previous representation in terms of S and W, the
new representation is now in terms of S and W
.
Figure 1 shows the computed Reynolds shear stress near the inner (con-
vex) wall at 90
12
y
Figure 1: Comparison of the Reynolds shear stress at 90
in the bend of a
U-duct. Db
ij
/Dt ,= 0 condition obtained from equation (4.46).
proposed which modied the turbulent transport hypothesis used in formulat-
ing the algebraic stress model. Later, Fu et al. (1988) concluded that in free
shear ows, where transport eects are signicant, full dierential stress mod-
els should be used rather than algebraic stress models. This conclusion was
reached based on the poor predictive performance in both plane and round
jet ows. In a recent study by Carlson et al. (2001) on the prediction of wake
ows in pressure gradients, poor mean ow predictions in the vicinity of the
wake centerline, where transport eects dominate, were also found. In this sub-
section, the assumption applied to the viscous and turbulent transport term,
equation(4.22), is re-examined and a modication is proposed.
Equation(4.22) assumes that the anisotropy in T
ij
is directly related to
the anisotropy in the Reynolds stresses themselves. A dierent constraint on
T
ij
can be found by rst rewriting equation(4.22) in terms of the anisotropy
tensor b
ij
,
T
ij
ij
k
T = T
ij
2
3
T
ij
2Tb
ij
. (4.57)
The right-hand side of equation(4.57) is now the sum of the deviatoric part
of T
ij
and is a term proportional to the anisotropy tensor b
ij
with scalar
coecient T. If equation(4.57) is now substituted into equation(4.18), the
dierential anisotropy equation can be written as
Db
ij
Dt
1
2k
_
T
ij
2
3
T
ij
_
=
_
b
ij
a
4
+a
3
_
b
ik
S
kj
+S
ik
b
kj
2
3
b
mn
S
mn
ij
_
(4.58)
a
2
(b
ik
W
kj
W
ik
b
kj
) +a
1
S
ij
_
.
40 Gatski and Rumsey
The coecients a
i
have been given previously in equation(4.19), with the
exception of a
4
, which is now given by
1
a
4
=
1
a
4
+
1
k
T (4.59)
or, using equation(3.30) for the turbulent kinetic energy,
a
4
=
__
1
+
1
Dk
Dt
+ 1
_
2 (
0
1)
1
_
1
. (4.60)
To obtain an implicit algebraic stress equation from equation(4.58), it is once
again necessary to assume the equilibrium condition on the anisotropy tensor
Db
ij
/Dt = 0 and to apply the new constraint,
T
ij
2
3
T
ij
= 0, (4.61)
on the viscous and turbulent transport terms. Equation(4.61) simply states
that the deviatoric part of the tensor T
ij
is zero.
In the case of homogeneous shear at equilibrium, an algebraic stress model
should yield the same results as the full second-moment closure from which it
is derived. Thus, any modication to the algebraic stress model is constrained
by this consistency condition. As equation(4.60) shows, the modication in
the formulation is partially through the term
1
Dk/Dt, which for the homo-
geneous shear case at equilibrium, can be related through the condition on the
turbulent time scale,
D
Dt
=
1
Dk
Dt
k
2
D
Dt
= 0, (4.62)
to the production-to-dissipation rate ratio
1
Dk
Dt
=
k
2
D
Dt
= C
1
_
T
C
2
. (4.63)
At equilibrium, the production-to-dissipation rate ratio is given by the relation
_
T
=
C
2
1
C
1
1
(4.64)
which, when used in equation(4.63), gives the new relation for a
4
a
4
=
_
1
2
_
1
, (4.65)
where
0
=
0
1, (4.66)
1
=
1
+ 1 +
_
C
2
C
1
C
1
1
_
. (4.67)
[1] Linear and nonlinear eddy viscosity models 41
The coecients C
1
and C
2
retain the same values used in the EASM formula-
tion with the original assumption on the anisotropy of the turbulent transport
and viscous diusion equation(4.22). This modication meets the requirement
that the homogeneous shear results are unaltered and only slightly aect the
results for the log-layer where the value of C
4
,
0
, and
1
are used instead of the coecients a
4
,
0
, and
1
in the original
formulation.
Figure 2 shows the theoretical values of C
(=
1
/) as a function of for
various levels of T/ for the EASM with the original turbulent transport and
viscous diusion assumption. Note that for regions in an equilibrium log-layer,
where = 1 and T/ = 1, the value of C
near 0.68! As a
result, the turbulent eddy viscosity produced near the center of a wake is very
high, and the velocity proles tend to be somewhat attened. This behavior
is consistent with results obtained by Fu et al. (1988).
-2 -1 0 1 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
P/=0.01
0.10
0.50
1.0
2
3
4
20
1.5
R
C
*
Figure 2: Values of C
is less than 0.19, as opposed to 0.68 for the original model. Consequently,
the resulting wake proles using the modied model are more realistic, as
42 Gatski and Rumsey
shown in Fig. 4 for a wake generated by a splitter plate (Carlson et al. 2001)
developing in zero pressure gradient.
-2 -1 0 1 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
P/=0.01
0.10
0.50
1.0
2
3
4
20
1.5
R
C
*
Figure 3: Values of C
ikm
+u
i
u
m
jkm
)
. .
R
ij
+
p
_
u
i
x
j
+
u
j
x
i
_
. .
ij
2
u
i
x
k
u
j
x
k
. .
ij
+
x
k
_
u
i
u
j
x
k
. .
D
ij
u
i
u
j
u
k
. .
D
t
ij
(u
i
jk
+u
j
ik
)
. .
D
p
ij
_
_
. .
D
ij
. (2.1)
Terms in boxes must be modelled. Note that
k
represents the system rotation
(angular) velocity, which should be distinguished from the rotation-rate vector
of a uid element W
i
=
1
2
ijk
W
kj
associated with the rotation-rate (vorticity)
tensor W
ij
=
1
2
(
U
i
x
j
U
j
x
i
), and the uid vorticity
i
=
ijk
W
kj
.
Each term has been given a short-hand alias so that in further discussion we
may refer just to the symbolic representation of the stress transport equation:
L
ij
+(
ij
= T
ij
+(
ij
+
ij
+
ij
ij
+ (T
ij
+T
t
ij
+T
p
ij
), (2.2)
where L
ij
represents the local change in time; (
ij
the convective transport;
T
ij
the production by mean-ow deformation; (
ij
the production by body
force;
ij
the production/redistribution by rotation force;
ij
the stress redis-
tribution due to uctuating pressure;
ij
the viscous destruction; and T
ij
the
diusive transport.
The modelling of the u
i
u
j
and equations follows the principles for mod-
elling the k- equations, using the characteristic turbulence time scale = k/
and length scales L = k
3/2
/, except that u
i
u
j
itself does not need to be
modelled. The principal task is modelling the pressure-strain term
ij
and the
stress dissipation rate
ij
. The standard modelling practice for the basic model
is outlined below.
2.1.1. Stress dissipation. At high Reynolds numbers the large scale motion
is unaected by viscosity, while the ne-scale structure is locally isotropic, i.e.
unaected by the orientation of large eddies. Consequently, the correlation
ij
= 2
u
i
u
j
x
k
x
k
which is associated with smallest eddies should reduce
to zero if i ,= j, while for i = j all three components should be equal. Hence, a
common way to model the viscous destruction of stresses for high Re-number
ows is:
ij
=
2
3
ij
, (2.3)
where =
u
l
u
l
x
k
x
k
and
ij
is the Kronecker unit tensor. It should be men-
tioned that at high Reynolds numbers can be interpreted as the amount of
50 Hanjalic and Jakirlic
energy exported by large (energy containing) eddies and transferred through
the spectrum towards smaller eddies until ultimately it is dissipated. Hence,
although represents essentially a viscous process, its value (the dissipation
rate) is governed by large, energy-containing eddies. Moreover, the above as-
sumption about the isotropy of
ij
is not very appropriate in non-homogeneous
ow regions such as in the vicinity of a solid wall. Nonetheless, this assump-
tion is widely used in standard DSM models and its deciency is supposedly
compensated by the model of the pressure-strain term
ij
, which accounts for
the turbulence anisotropy.
2.1.2. Turbulent diusion. The form of T
ij
is such that
_
V
T
ij
dV = 0
over a closed domain bounded by impermeable surfaces (as follows from a
Gaussian transformation of the volume integral into a surface integral). Hence
the T
ij
term is of a transport (diusive) nature. The molecular diusion T
ij
can be treated exactly but at high turbulence Re numbers it is negligible.
The remaining two parts, the turbulent diusion by uctuating velocity and
uctuating pressure, T
t
ij
and T
p
ij
, need to be modelled. The most popular
model for T
t
ij
is the generalized gradient diusion (GGD), known also as the
DalyHarlow model: u
k
= C
u
k
u
l
x
l
. The application of GGD to the
turbulent velocity diusion of stress yields:
T
t
ij
=
x
k
(u
i
u
j
u
k
) =
x
k
_
C
s
k
u
k
u
l
u
i
u
j
x
l
_
. (2.4)
A simpler variant is simple gradient diusion (SGD), with an isotropic (scalar)
eddy diusivity (Shir 1973).
T
t
ij
=
x
k
(u
i
u
j
u
k
) =
x
k
_
C
s
k
2
u
i
u
j
x
k
_
. (2.5)
More advanced treatments are discussed in Section 3.
The turbulent transport by pressure uctuations has a dierent nature
(transport by the propagation of disturbances) and none of the gradient trans-
port forms is applicable to modelling T
p
ij
. Yet, it is common to lump T
p
ij
with
T
t
ij
and to adjust the coecient C
s
. In many ows the pressure transport is
smaller than the velocity transport so that this approximation often brings
no adverse consequences. However, in ows driven by thermal buoyancy (e.g.
RayleighBenard convection) this is not the case and such models are not
appropriate.
2.1.3. Pressure-strain interaction. Pressure uctuations act to scram-
ble the turbulence structure and redistribute the turbulent stress among com-
ponents to make turbulence more isotropic. Some insight into the physics and
hints for modelling can be gained from the exact Poisson equation for the
[2] Second-moment turbulence closure modelling 51
pressure uctuations (obtained after dierentiating the equation for u
l
with
respect to x
l
):
2
p
x
2
l
=
2
x
l
x
m
(u
l
u
m
u
l
u
m
) 2
U
l
x
m
u
m
x
l
+
f
l
x
l
, (2.6)
which can be integrated to yield p at x:
p
=
1
4
_
V
_
2
x
l
x
m
(u
l
u
m
u
l
u
m
) + 2
U
l
x
m
u
m
x
l
x
l
_
dV (x
)
[r[
, (2.7)
where r = x
x. Multiplication by
_
u
i
x
j
+
u
j
x
i
_
(at x) and averaging yields
the following exact expression for
ij
:
ij
=
p
_
u
i
x
j
+
u
j
x
i
_
=
1
4
_
V
_
_
_
2
u
l
u
m
x
l
x
m
_
_
u
i
x
j
+
u
j
x
i
_
. .
ij,1
+2
_
U
l
x
m
_ _
u
m
x
l
_
_
u
i
x
j
+
u
j
x
i
_
. .
ij,2
_
f
m
x
m
_
_
u
i
x
j
+
u
j
x
i
_
_
_
. .
ij,3
dV (x
)
[x x
[
+
1
4
_
A
_
_
1
r
_
u
i
x
j
+
u
j
x
i
_
p
_
u
i
x
j
+
u
j
x
i
_
n
_
1
r
_
_
_
dA
. .
w
ij
. (2.8)
Dierent terms in (2.8) can be associated with dierent physical processes,
which can be modelled separately. The most common approach is to split the
term into the following parts:
ij
=
ij,1
+
ij,2
+
ij,3
+
w
ij,1
+
w
ij,2
+
w
ij,3
, (2.9)
where
ij,1
is the return to isotropy of non-isotropic turbulence, (slow term).
In the absence of the mean rate of strain S
ij
and a body force, and away
from any boundary constraint, pressure uctuations will force turbulence
to approach an isotropic state.
ij,2
is the isotropization of the process of stress production due to S
ij
(rapid term). Pressure uctuations will slow down a preferential feeding
of turbulence by S
ij
into a particular component imposed by the active
strain-rate components.
52 Hanjalic and Jakirlic
ij,3
is the isotropization of stress production due to a body force;
w
ij,1
,
w
ij,2
,
w
ij,3
is the wall blockage (eddy splatting) and pressure re-
ection eect associated with
ij,1
,
ij,2
and
ij,3
respectively. The rst
process, which is dominant, impedes the isotropizing action of the pres-
sure uctuations. The pressure reection acts in fact in the opposite way
(the pressure wave reected from a solid surface enhances eddy scram-
bling), but this eect is smaller in comparison with the wall-blockage
eect.
Jones and Musonge (1988) argued that the mean strain rate should appear
also in the exact transport equation for the slow term
ij,1
and that separate
modelling of each part of
ij
may not be fully justied. Their integral model
for the complete
ij
diers, however, only slightly from other models in the
values of the coecients (see next section). However, splitting the terms, even
if not fully justied, enables us to distinguish some physical eects from oth-
ers and gives some basis for their modelling. For that reason we follow the
conventional approach here.
The model of
ij,1
(the slow term). Based on the idea that the pressure
uctuations tend to reduce turbulence anisotropy, Rotta (1951) proposed a
simple linear model by which
ij,1
is proportional to the stress anisotropy
tensor itself (of course with a negative sign). The expression is known as the
linear return-to-isotropy model:
ij,1
= C
1
a
ij
= C
1
_
u
i
u
j
k
2
3
ij
_
, (2.10)
where a
ij
is known as the stress anisotropy tensor; it is just twice the value
of the tensor b
ij
introduced in [1]. To promote a reduction in anisotropy the
coecient C
1
must be greater than 1; the most common value is C
1
= 1.8.
The models of
ij,2
and
ij,3
(the rapid terms). Without going deeper
into the physics, we recall at this point that
ij,2
is associated with the mean
rate of strain, which is usually the major source of turbulence production.
Hence the pressure scrambling action can be expected to modify the very
process of stress production. Following this idea, Naot et al. (1970) proposed
a model of
ij,2
analogous to Rottas model of the slow term, known as the
Isotropization of Production (IP) model:
ij,2
= C
2
_
T
ij
2
3
ij
T
_
. (2.11)
An analogous approach to the pressure eect on stress generation due to body
forces leads to:
ij,3
= C
3
_
(
ij
2
3
ij
(
_
. (2.12)
Note that T 1/2T
kk
and ( 1/2(
kk
[2] Second-moment turbulence closure modelling 53
Interdependence of coecients. The above listed coecients C
1
, C
2
and
C
3
have been obtained mainly from selected experiments where only one of
the processes can be isolated (e.g. C
1
from experiments on free return to
isotropy of initially strained turbulence, C
2
from rapid distortion theory). Of
course, the coecients have also been tuned through subsequent validation
in a series of experimentally well documented ows. Values other than those
quoted above have also been proposed as more suitable for some classes of
ows. However, the validation revealed that a change in one coecient also
requires an adjustment of others in order to reproduce fully the total eect of
the pressure-strain term. A useful correlation between C
1
and C
2
is:
C
1
4.5(1 C
2
). (2.13)
The most frequently used values are: C
1
= 1.8, C
2
= 0.6 and C
3
= 0.55.
Modelling the wall eects on
ij
. Solid walls and free surfaces splat
neighbouring eddies, which leads to a larger turbulence anisotropy. Wall imper-
meability (blocking eect) damps the velocity uctuations in the wall-normal
direction. On the other hand, pressure uctuations reecting from bounding
surfaces will enhance the pressure scrambling eect. Both these eects are
non-viscous in nature and are essentially dependent on the wall distance and
the wall conguration. The blockage eect is stronger resulting in an impedi-
ment of the isotropizing action of pressure uctuations. As a consequence, the
stress anisotropy in the near-wall region is higher than in free ows at similar
strain rates (Launder et al. 1975). Thus, for a simple shear with x
1
the ow
direction and x
2
the direction of shear:
a
11
a
22
a
33
a
12
Homogeneus shear ow (Champagne et al.) 0.30 0.18 0.2 0.33
Near-wall region of a boundary layer 0.55 0.45 0.11 0.24
Wall damping is expected to aect both the slow and the rapid pressure-
induced stress-redistribution processes and we can decompose
w
ij
into two
terms:
w
ij,1
corrects the values of the stress components in such a way as
to slow down the redistribution and diminish the wall-normal component to
the benet of the streamwise and spanwise ones, and impedes the reduction
of the shear stress. Likewise,
w
ij,2
modies the processes of stress production
in the near-wall region. The wall correction should attenuate with distance
from the wall: this is usually arranged through an empirical damping function,
f
w
= L/C
L
y, where L is the turbulence length scale. Close to a wall L y and
the constant C
L
is chosen so that f
w
1. At larger wall-distances L const
so that f
w
0
Based on the above reasoning several models of
w
ij,1
and
w
ij,2
have been
proposed, the most frequently employed schemes being:
54 Hanjalic and Jakirlic
Shir (1973):
w
ij,1
= C
w
1
k
_
u
k
u
m
n
k
n
m
ij
3
2
u
i
u
k
n
k
n
j
3
2
u
k
u
j
n
k
n
i
_
f
w
; (2.14)
Gibson and Launder (1978):
w
ij,2
= C
w
2
_
km,2
n
k
n
m
ij
3
2
ik,2
n
k
n
j
3
2
jk,2
n
k
n
i
_
f
w
; (2.15)
where C
w
1
= 0.5, C
w
2
= 0.3, f
w
=
0.4 k
3/2
x
n
, x
n
is the normal distance to the
wall and n
k
is the unit vector of the coordinate normal to the wall. These
schemes have been and still are being successfully applied for attached ows
on nearly plane, continuous surfaces but they have been superseded for more
complex ows by alternatives discussed later in this chapter and in [3] and [4].
2.2 The model equation for
The exact equation for is not of much help as a basis for modelling except
to give some indication of the meanings and importance of various terms. The
dissipation rate appears as the exact sink term in the k equation and needs to
be provided to close the model (
ij
in u
i
u
j
equation). However, what is needed
for closing other terms is the characteristic time and length scale of the energy-
containing eddies and the equation for energy transfer from these eddies down
the spectrum towards the smaller eddies. This energy transfer rate coincides
with the dissipation rate only under the conditions of spectral equilibrium.
Nevertheless, it is instructive to look at the exact transport equation for :
D
Dt
=
t
..
L
+U
k
x
k
. .
C
= 2
_
u
i
x
l
u
k
x
l
+
u
l
x
i
u
l
x
k
_
U
i
x
k
. .
P
1
+P
2
2u
k
u
i
x
l
2
U
i
x
k
x
l
. .
P
3
+
f
i
x
l
u
i
x
l
. .
G
2
u
i
x
k
u
i
x
l
u
k
x
l
. .
P
4
2
_
2
u
i
x
k
x
l
_
2
. .
Y
+
_
_
_
_
_
_
_
x
k
x
k
. .
D
u
k
. .
D
t
p
x
i
u
k
x
i
. .
D
p
. .
D
_
_
_
_
_
_
_
. (2.16)
[2] Second-moment turbulence closure modelling 55
The physical meaning of the terms can be inferred from comparison with the
transport equation for the mean square of uctuating vorticity
2
i
(enstro-
phy), since for homogeneous turbulence 2
2
i
(Tennekes and Lumley
1972). At high Reynolds numbers the source terms in the third row (T
4
and
Y ) are dominant, while the other production terms can be neglected as smaller
(T
1
+T
2
and (
by Re
1/2
t
and T
3
by Re
t
, where Re
t
= k
2
/ is the turbu-
lence Reynolds number). Of course, for low-Re-number regions of ow, these
terms need to be taken into account. Note that all terms in boxes must be
modelled.
In the DSM closures the same basic form of model equation for is used as
in the k- model, except that now u
i
u
j
is available (and also u
i
if the second-
moment closure level is also used for the thermal and other scalar elds). This
has the following implications:
The T and ( (production of kinetic energy) in the source term of are
treated in exact form;
The generalized gradient-diusion hypothesis is used to model turbulent
diusion
T
t
=
x
k
(u
k
) =
x
k
_
C
u
k
u
l
x
l
_
. (2.17)
Hence, the model equation for has the form (Hanjalic and Launder 1972):
D
Dt
=
x
k
_
C
u
k
u
l
x
l
_
+ (C
1
T +C
3
( +C
4
k
U
k
x
k
C
2
)
k
, (2.18)
where the coecients have the same values as in the k- model except for the
new coecient C
)
The following values of coecients have been recommended:
C
s
C
1
C
2
C
w
1
C
w
2
C
C
1
C
2
C
3
C
4
0.2 1.8 0.6 0.5 0.3 0.18 1.44 1.92 1.44 0.373
2.3 Second-moment closure for scalar elds for high-Peclet-
number ows
The second-moment closure models for scalar elds (thermal, species concen-
tration) follow essentially the same principles as the modelling of the velocity
eld (Launder 1976). This means that the transport equations for the scalar
ux hu
i
, u
i
, cu
i
are modelled starting from their exact parent equations
(here h is the uctuating enthalpy, is the uctuating temperature and c
is the uctuating concentration; for a multi-component mixture each species
concentration is considered separately, i.e. c is replaced by c
(i)
). Because the
principles are the same, except for the source terms, which usually require no
special consideration (nor modelling), we consider here only the equation for
the turbulent heat ux u
i
.
56 Hanjalic and Jakirlic
2.3.1 The Model Equation for Scalar Flux u
i
The exact transport equation for the turbulent heat ux vector for high Peclet
numbers (Pe = Re.Pr) can be derived in a manner analogous to the stress
equation:
Du
i
Dt
= u
i
u
k
x
k
. .
P
i
u
k
U
i
x
k
. .
P
U
i
g
i
2
. .
G
i
+
p
x
i
. .
i
( +)
x
k
u
i
x
k
. .
x
k
_
u
i
u
k
+
p
ik
_
. .
D
t
i
+D
p
i
. (2.19)
The physical meaning of the various terms can be inferred by comparison with
the Reynolds-stress equation:
T
i
is the thermal production (nonuniform temperature eld interacting
with turbulent stresses),
T
U
i
is the mechanical production (mean ow deformation interacting
with the turbulent heat ux),
(
i
is the gravitational production (gravitation interaction with the uc-
tuating temperature eld),
i
is the pressuretemperature-gradient correlation,
i
is the molecular destruction,
T
i
is the diusive transport (where t denotes diusion by turbulent
velocity and p by pressure uctuations).
All three production terms can be treated in exact form, but an additional
transport equation needs to be provided for the temperature variance
2
. Other
terms need to be modelled.
Following the same modelling principles we can express the pressure scram-
bling term as:
i
=
i,1
+
i,2
+
i,3
= C
1
u
i
k
C
2
T
U
i
C
3
(
i
. (2.20)
It is noted that the term corresponding to T
i
is absent because the mean
temperature gradient does not appear in the exact Poisson equation for
i
.
The turbulent diusion by velocity and pressure uctuations is modelled by
GGD. It should be noted that the viscous diusion needs also to be modelled,
except in the case of Prandtl numbers of O(1). For high Pe numbers
i
can
[2] Second-moment turbulence closure modelling 57
be neglected. Hence, the model equation for the scalar ux (with wall eects
omitted
2
) is:
Du
i
Dt
= u
i
u
k
x
k
(1 C
2
)u
k
U
i
x
k
(1 C
3
)g
i
2
C
1
u
i
k
+C
x
k
_
u
k
u
l
k
u
i
x
l
_
. (2.21)
2.3.2 The model equation for scalar variance
2
The model equation for
2
resembles closely the k equation and can be mod-
elled in the same manner. It contains a single production term which can be
treated exactly. The turbulent transport is modelled in the usual gradient-
transport form. The only problem is the sink term (molecular destruction)
= 2
_
x
j
_
2
. A transport equation for
/ = R is constant (where
=
2
/2
and = k/)
leads to the simple approximation
=
R
2
2k
. (2.22)
Hence, the model equation for
2
is
D
2
Dt
= 2u
i
x
i
1
2R
2
2k
+C
x
j
_
u
i
u
j
k
2
x
i
_
. (2.23)
2.3.3 Summary of coecients for scalar ux model
The following values of coecients can be recommended for the scalar ux
model for high Peclet number ows:
C
2 C
C
1
C
2
C
3
R
0.2 0.15 3.5 0.55 0.55 0.5
2.4 The algebraic stress/ux models (ASM/AFM)
A considerable simplication of the dierential equation for u
i
u
j
can be achieved
by eliminating the transport terms in individual stress components in terms
2
Wall-reection eects have been considered in Gibson and Launder (1978) and Launder
and Samaraweera (1979). More elaborate models than equation (20) appear to avoid the
need for explicit wall corrections (see [3], [14], [15].)
58 Hanjalic and Jakirlic
of transport terms of the kinetic energy. The common approach is to assume
the so-called weak non-equilibrium hypothesis (Rodi 1976) by which the time
and space evolution of the stress anisotropy tensor is equal to zero, i.e.:
Da
ij
Dt
= 0. (2.24)
The expansion of a
ij
=
u
i
u
j
k
2
3
ij
leads to:
Du
i
u
j
Dt
T
ij
=
u
i
u
j
k
_
Dk
Dt
T
k
_
=
u
i
u
j
k
(T +( ). (2.25)
Each stress component u
i
u
j
can now be expressed in terms of an algebraic
expression:
u
i
u
j
=
2
3
ij
k +
k
1
_
T
ij
2
3
T
ij
_
+
2
_
(
ij
2
3
(
ij
__
, (2.26)
where
1
and
2
are functions of T/ and (/ (containing also the coecients
from the modelled expressions for the pressure-strain terms). ASMs have some
advantages such as a reduction of computing time in comparison with the
full (dierential) DSM, they usually give better results than the linear k-
model where stress anisotropy is strong and important, e.g. secondary ows in
the straight ducts. However, they are usually derived from a presumed DSM
model and they can at best perform as well as the parent DSM provided
the ow evolution is slow. A major shortcoming is the inability to reproduce
the evolution of the stress anisotropy in nonequilibrium ows, when the ow
history and development cannot be fully accounted for by transport terms in
the k- and -equations.
The same approach can be applied to derive an algebraic ux model (AFM)
for a scalar eld, in which case the weak equilibrium hypothesis is applied to
the scalar-ux correlation coecient, i.e. D(u
i
/
_
(k
2
)/Dt = 0. The
AFM/ASM approach has even greater appeal if convection of scalars is consid-
ered, because even for a single scalar eld the number of dierential transport
equations in a full second-moment closure may be as large as 17.
The above ASM is implicit in u
i
u
j
. Besides, the functions have expres-
sions in the denominator, which may become very small or even zero, leading
to singularities and numerical instability. In order to overcome the numeri-
cal problems, several explicit nonlinear ASMs and AFMs have recently been
proposed in the literature (Speziale and Gatski 1993, Wallin and Johansson
2000). Modelling at this level is more extensively considered in [1].
3 Advanced Dierential Second-Moment Closures
The basic dierential second-moment closure models (DSM) have proved to
perform better than linear Eddy Viscosity Models (EVM) in many ows, but
[2] Second-moment turbulence closure modelling 59
not always, and not by a convincing margin. Over the past two decades there
has been much activity aimed at improving the basic DSM. All improvements
lead, inevitably, to more complex models which may pose additional com-
putational diculties (numerical instabilities, slower convergence). The model
developers often focus on only one or two crucial terms in the u
i
u
j
and equa-
tion and propose more sophisticated expressions which better satisfy physical
and mathematical constraints (realizability, two-component limit, vanishing
and innite Reynolds numbers, etc.). Validation is usually performed in a lim-
ited number of test cases that display particular features which are the focus
of the new development. The resulting complex model is often out of balance
with the usually much simpler models adopted for the rest of the unknown
processes.
We conne our attention here to only a few advancements, which seem to
bring desirable improvement and yet retain the form of the model expressions
at a manageable level of sophistication. The focus is on the models of turbulent
diusion and pressure scrambling in the u
i
u
j
equation and on some proposals
to improve the equation. More complex models of the processes are dealt
with in other chapters in this volume.
3.1 Some improvements to the modelled u
i
u
j
equation
Turbulent diusion of u
i
u
j
. A coordinate-frame invariant model of T
ij
can be derived by tensorial expansion of the GGD hypothesis. Alternatively,
a truncation of the model transport equation for triple velocity correlation
u
i
u
j
u
k
, retaining only the rst order terms, yields (Hanjalic and Launder
1972):
T
t
ij
=
x
k
(u
i
u
j
u
k
) =
x
k
_
C
s
k
_
u
i
u
l
u
j
u
k
x
l
+u
j
u
l
u
k
u
i
x
l
+u
k
u
l
u
i
u
j
x
l
__
.
(3.1)
Application of the moment-generating function leads to still more complex
expressions (e.g. Lumley 1978, Cormac et al. 1978, Magnaudet 1992). In addi-
tion to the invariant expression (3.1), the above-mentioned expressions contain
additional terms, some including the gradients of the turbulent kinetic energy
and its dissipation rate, and even the mean rate of strain. While these more
general expressions lead to some improvements, a comparison with the DNS
results for a plane channel ow showed that none of the above mentioned
models satises all stress components (Hanjalic 1994, Jakirlic 1997). Nagano
and Tagawa (1991) proposed a new way to treat triple velocity and scalar
correlations by solving the transport equations for triple moments for each
of the velocity uctuations, u
3
i
, while evaluating the mixed triple moments
from algebraic correlations. The latter were derived from structural character-
istics of the shear-generated turbulence. This approach led to improvements
in near-wall ows, but it is not coordinate invariant.
60 Hanjalic and Jakirlic
It should be noted, however, that all the more general expressions give rise to
a large number of component terms, particularly in non-Cartesian coordinates.
Because in many ows with a strong stress production the turbulent transport
is relatively unimportant, a simpler model usually suces.
A simpler expression which satises the coordinate-frame invariance and
still retains a relatively simple form is the expression proposed by Mellor and
Herring (1973):
T
t
ij
=
x
k
(u
i
u
j
u
k
) =
x
k
_
C
s
k
2
_
u
j
u
k
x
i
+
u
k
u
i
x
j
+
u
i
u
j
x
k
__
. (3.2)
Pressure-strain interaction: The slow term. The return to isotropy
is in fact a nonlinear process: a tensorial expansion making use of the Cayley
Hamilton theorem leads to a quadratic model of the slow term (Lumley
1978, Reynolds 1984, Fu, Launder and Tselepidakis 1987, Speziale, Sarkar
and Gatski 1991);
ij,1
= [C
1
a
ij
+C
1
(a
ik
a
jk
1
3
ij
A
2
)], (3.3)
where C
1
, C
1
are, in general, functions of turbulence Re-number and stress
anisotropy invariants A
2
= a
ij
a
ij
, A
3
= a
ij
a
jk
a
ki
and the atness parameter
A = 1
9
8
(A
2
A
3
). Speziale, Sarkar and Gatski (1991) (SSG) proposed
C
1
= 1.05 and this value has been generally accepted in the framework of
the complete quadratic pressure-strain model (see below). The UMIST group
(Craft and Launder 1991) proposed a similar expression (validated in free
ows, but subsequently also used in near-wall ows):
ij,1
= C
1
__
1 +
C
1
0
C
1
_
a
ij
+C
1
_
a
ik
a
kj
1
3
A
2
ij
__
, (3.4)
with C
1
= 3.1(A
2
A)
1
2
, C
1
= 1.2 and C
1
0
= 1
Earlier, Lumley (1978) and Shih and Lumley (1993) discussed the quadratic
expression, but due to the lack of evidence, they discarded the second term
and proposed C
1
in the form of a function dependent on Reynolds number
and stress invariants.
The appearance of DNS data for each part of the pressure-strain term makes
it possible not only to verify the proposed expression, but also the values of
the coecients. Equation (3.3) contains two unknowns, C
1
and C
1
. Using any
pair of experimental data for
ij,1
for two components enables candidate values
for C
1
and C
1
to be obtained. In fact, because four components of
ij,1
are
available in a channel ow, the problem is overdened and dierent solutions
can emerge for dierent combinations, if expression (3.3) is not unique. Such a
test in a plane channel ow, using the DNS results of Kim et al. (1987) showed
that both coecients, C
1
and C
1
vary strongly across the ow. However, the
[2] Second-moment turbulence closure modelling 61
Figure 1: Variation of C
1
and C
1
in a plane channel ow. Symbols: evaluation
from pairs of DNS components of
ij,1
, _ 11-22, 11-33, 22-33. Lines:
dierent models, - - - HL, HJ, - SSG, CL, LT, LRR (for
acronyms see Table 1).
results for dierent pairs of
ij,1
collapse on one curve in the region close to a
wall for y
+
< 60 for Re
m
= 5600, though departing substantially away from
the wall, Figure 1. The data also show that a simple expression
C
1
= A
2
C
1
(3.5)
matches the DNS data well in the near-wall region. It should be noted, how-
ever, that C
1
changes sign at y
+
12, exhibiting a peak at y
+
6. Most
models do not reproduce such a behaviour, but impose a monotonic approach
of C
1
to zero at the wall. Figure 1 shows the variation of C
1
and C
1
given by
dierent model proposals, discussed above.
Pressure-strain interaction: The rapid term. The basis for modelling
of the rapid term is the general expression
ij,2
=
U
l
x
m
(b
mi
lj
+b
mj
li
), (3.6)
which represents in a symbolic form the corresponding element of equation
(2.8) after the mean velocity gradient is taken out of the volume integral by
62 Hanjalic and Jakirlic
assuming local mean ow homogeneity. The modelling task is reduced to ex-
pressing the fourth-order tensor b
mi
lj
in terms of available second-order tensors
(turbulent stress tensor u
i
u
j
and Kronecker unit tensor
ij
). A convenient and
more general way is to formulate the complete
ij,2
in the form of a tensorial
expansion series in terms a
ij
, S
ij
and W
ij
. The complete expression (closed
by the CayleyHamilton theorem) contains terms up to fourth order in a
ij
(Johansson and H allback 1994), though cubic models have also been designed
to meet most constraints. An (incomplete) cubic model can be written in the
general form:
ij,2
= C
2
Ta
ij
+C
3
kS
ij
+C
4
k
_
a
ik
S
jk
+a
jk
S
ik
2
3
ij
a
kl
S
kl
_
+C
5
k(a
ik
W
jk
+a
jk
W
ik
)
+C
6
k (a
ik
a
kl
S
jl
+a
jk
a
kl
S
il
2a
kj
a
li
S
kl
3a
ij
a
kl
S
kl
)
+C
7
k (a
ik
a
kl
W
jl
+a
jk
a
kl
W
il
)
+C
8
k[a
2
mn
(a
ik
W
jk
+a
jk
W
ik
) +
3
2
a
mi
a
nj
(a
mk
W
nk
+a
nk
W
mk
)], (3.7)
where S
ij
=
1
2
_
U
i
x
j
+
U
j
x
i
_
and W
ij
=
1
2
_
U
i
x
j
U
j
x
i
_
The line separates the quasi-linear (in a
ij
) from nonlinear expressions (quasi-
linear, because C
2
T a
ij
is, in fact, quadratic since T contains a
ij
, while all
other terms are linear). Although nonlinear models are claimed to satisfy bet-
ter the mathematical constraints and physical requirements, the large number
of terms currently reduce their appeal for industrial applications. In addition
to the rudimentary IP model introduced earlier, equation (2.11), two other
popular models that still retain a simple form are the linear Quasi-Isotropic
model of Launder, Reece and Rodi (1975), denoted as LRR-QI, and the quasi-
linear model of Speziale, Sarkar and Gatski (1991), denoted as SSG.
3
The
models dier in the values of coecients. Unlike the IP model, the LRR-QI
is capable of reproducing some though insucient stress anisotropy in the
near-wall region even without any wall reection term
w
ij
and was shown to
perform slightly better in some homogeneous and free thin shear ows (Laun-
der et al. 1975). However, subsequent broader testing showed that the IP with
the Shir (1973) and Gibson and Launder (1978) wall-echo corrections, (2.14
and 2.15), perform generally more satisfactorily. The SSG model contains the
above mentioned quasi-linear term, which is absent from LLR-QI model, and
one of the coecients is formulated as a function of the second stress invariant
A
2
.
A summary of the coecients for some of the models of
ij
found in the
literature is given in Table 1.
3
The early model of Hanjalic and Launder (1972) also contained the nonlinear term
C
2
Pa
ij
with the same value of the empirical coecient, though with opposite sign (Table 1).
[2] Second-moment turbulence closure modelling 63
Table 1: Summary of coecients in pressure-strain models
ij,1
ij,2
Authors Linear Quadratic Linear Quadratic Cubic
C
1
C
1
C
2
C
3
C
4
C
5
C
6
C
7
C
8
HL 2.8 0 0.9 0.8 0.71 0.582 0 0 0
LRR IP 1.8 0 0 0.8 0.6 0.6 0 0 0
LRR QI 1.8 0 0 0.8 0.873 0.655 0 0 0
SSG 1.7 1.05 0.9 0.8 0.65 0.2 0 0 0
0.625A
1/2
2
LT 6.3AA
1/2
2
0.7C
1
0 0.8 0.6 0.866 0.2 0.2 1.2
Abbreviations: HL: Hanjalic and Launder (1972), LRR: Launder, Reece and Rodi
(1975), SSG Speziale, Sarkar and Gatski (1991), LT: Launder and Tselepidakis (1994).
The model of the rapid pressure-strain term can alternatively be written
in terms of S
ij
, T
ij
and D
ij
=
_
u
i
u
k
U
k
x
j
+u
j
u
k
U
k
x
i
_
(Launder et al. 1975).
The quasi-linear model, corresponding to the rst line of equation (3.7) reads
ij,2
= C
2
Ta
ij
C
3
kS
ij
C
4
(T
ij
2/3T
ij
) C
5
(D
ij
2/3T
ij
). (3.8)
The conversion of the coecients follows from the following relationships
(Hadzic 1999): C
2
= C
2
, C
3
= 4/3C
4
C
3
, C
4
= 1/2(C
4
+ C
5
), and C
5
=
1/2(C
4
C
5
). It should be mentioned that the LRR-IP and LRR-QI models
require the use of wall-echo terms
w
ij
dened earlier, whereas the SSG model
does not. Apparently, the extra quasi-linear term and the function C
3
account
for the stress redistribution modication by a solid wall making the wall-echo
terms redundant. While this statement is not fully true (LRR-IP+wall echo
terms reproduce better the stress anisotropy in a near-wall region, as can be
seen in Hadzic 1999), the SSG has some appeal as a compromise between the
desired accuracy and computational economy: it is more practical than LRR-
QI with wall-correction term
w
ij
, and more accurate than both the IP and
LRR-QI models if used without the
w
ij
.
64 Hanjalic and Jakirlic
The quasi-linear term can be interpreted as an extension of the slow term
ij,1
with the coecient C
1
replaced by a function C
1
(1 + C
2
T/). The ratio
T/ has been used earlier in several models, even at the two-equation level,
to account for departures from local energy equilibrium.
The inclusion of nonlinear terms brings more exibility and potential to
model better various extra strain eects, as well as to satisfy mathematical
and physical constraints. For example, the more complete two-component-
limit model, Craft and Launder (1996), which retains all cubic terms yet with
only two adjustable coecients, was reported to perform very well both in
near-wall and free-ow regions in several complex ows. More details of this
approach can be found in [3] and [14].
Elliptic relaxation concept. Instead of wall damping dened in terms of
local ow and turbulence variables, wall topology and distance, Durbin (1991,
1993) proposed providing the required damping through an elliptic relaxation
(ER) equation, which accounted for the non-viscous wall blocking eect. In
the framework of eddy viscosity models, this approach requires one elliptic
relaxation equation for the wall damping function f, which is used to model
an additional transport equation for another scalar property denoted as v
2
(which essentially reduces to the wall-normal turbulent stress in the near-
wall region of wall-parallel ows). The resulting k--v
2
-f model (with
t
=
C
D
v
2
/ and = max[k/, C
_
(/)], where the latter scale-switch accounts
for viscous eects) allows integration up to the wall. While this model seems
to perform signicantly better than the conventional high- or low-Re-number
eddy viscosity models, it still shows deciencies in reproducing strongly non-
equilibrium ows where the stress anisotropy changes fast and is aected by
the ow history.
Durbin (1993) extended his elliptic relaxation approach to full second-moment
closure by proposing the following elliptic equation for the tensorial damping
function f
ij
corresponding to the stress tensor u
i
u
j
:
L
2
2
f
ij
f
ij
=
h
ij
k
, (3.9)
where
h
ij
is the homogeneous (far from a wall) pressure-strain model, for
which, in principle, any known model (without wall correction) can be used
4
.
The function f
ij
is then used to obtain
D
ij
= k f
ij
. With viscous eects ac-
counted for by imposing Kolmogorov scales as lower bounds on both the time
and length scales of turbulence (just as in the k--v
2
-f model), the model
allows the integration up to a solid wall (for further comments on the imple-
mentation of viscous eects, see Section 5). A fuller account of the physical
basis and the associated analysis for the ER approach, together with examples
of applications is provided in [4].
4
Note that Durbin used
D
ij
= (
ij
1/3
kk
ij
) (
ij
u
i
u
j
/k) instead of the conven-
tional pressure-strain
ij
, where
ij
= D
p
ij
ij
is the velocitypressure-gradient correlation.
[2] Second-moment turbulence closure modelling 65
The ER approach to second-moment closure requires solution up to six el-
liptic relaxation equations, one for each component of the turbulent stress
u
i
u
j
. Despite the demonstrated success in reproducing several types of ows,
the unavoidable additional computational eort, together with some problems
experienced in dening and implementing wall boundary conditions for each
f
ij
, have limited wider testing of this model. While still utilizing the elliptic
relaxation concept within the second-moment closure framework a signicant
simplication can be achieved by solving a single elliptic equation. It is re-
called that the elliptic relaxation equation essentially accounts for geometrical
eects (wall conguration and topology) and provides a continuous modica-
tion of the homogeneous pressure-strain process as the wall is approached to
satisfy the wall conditions. Hence, it should be possible to dene this tran-
sition by a single variable. Manceau and Hanjalic (2000b) have proposed a
blending model which entails the solution of the elliptic relaxation equation
for a blending function ,
L
2
2
=
1
k
, (3.10)
with boundary conditions k[
w
= 0 and k[
ij
= (1 k)
w
ij
+k
h
ij
(3.11)
and between the isotropic and near-wall nonisotropic stress dissipation rate
ij
= (1 Ak)
u
i
u
j
k
+Ak
2
3
ij
. (3.12)
Here
h
ij
can be any known homogeneous model of
ij
, whereas the inhomo-
geneous part (satisfying the wall constraints) is dened as
w
ij
= 5
k
_
u
i
u
k
n
j
n
k
+u
j
u
k
n
i
n
k
1
2
u
k
u
l
n
k
n
l
(n
i
n
j
ij
)
_
. (3.13)
The unit normal vectors are obtained from
n =
[[[[
. (3.14)
The testing of this model in a plane channel and in ow over a backward-
facing step showed very good agreement with the DNS data, though further
validation in more complex ows still remains to be undertaken.
3.2 Some modications to the equation
The rudimentary form of the equation is used in practically all industrial
CFD codes to provide the sink term in the kinetic energy equation, and to
66 Hanjalic and Jakirlic
supply the turbulence scale by which the turbulent diusion is modelled. This
equation has too simple a form for such a task for any turbulent ow sig-
nicantly out of local energy equilibrium. The inability of EVMs to generate
accurately the normal stresses is the root of the failure of such schemes to
model the production of due to normal straining, if these are signicant
(ow acceleration and deceleration, ow recovery after reattachment, etc.)
The use of second-moment closure with the same equation at least obviates
this problem, because the normal-stress components are then (hopefully) well
reproduced. However, any extra strain rates or departure from equilibrium
may require additional modications to the equation.
There have been several proposals to modify and upgrade the simple
equation, though most of them were aimed at curing a specic deciency
related to a particular ow class. Specic tuning usually resulted in achieving
the aim, but in most cases subsequent validation in some other ows produced
unwanted eects. A case in point is the modication which emerged from the
application of the Renormalization Group Theory (RNG). The outcome of
this approach is the insertion of an extra term in the equation:
S
RNG
=
(1 /
0
)
1 +
3
T
k
, (3.15)
where Sk/ is the relative strain parameter (in fact the ratio of the tur-
bulence time scale and the scale of mean rate of strain) and S
_
S
ij
S
ji
is the strain-rate modulus. This term aimed to distinguish large from small
strain rates by increasing or decreasing the source of dissipation depend-
ing on whether the strain parameter is larger or smaller, respectively, than
what is believed to be a typical value for homogeneous equilibrium shear ow,
0
= 4.8. While the term indeed improved the predictions in the recirculation
zone, as well as in the stagnation region, it proved to be harmful in many other
ow cases, particularly when the normal straining is dominant. The reason is
that the term does not distinguish the sign of the strain S
ij
, and produces the
same eect for the same strain intensity irrespective of its sign, i.e., whether
the ow is subjected to acceleration or deceleration, compression or expansion
(e.g. in reciprocating engines). Figure 2 shows the eect of including the RNG
term (3.15) with the LRR stress model to ow in an axisymmetric contraction
and expansion at roughly the same strain = 62.2 and 86.6 respectively, com-
pared with DNS results of Lee (1985) (Hanjalic 1996). The standard model
gives poor results. The RNG modication indeed improves the ow predic-
tions in the expansion, but actually leads to worse agreement in the contrac-
tion. Hence, the use of such a remedy, particularly by inexperienced users, can
yield adverse, instead of benecial, eects. The strain-rate modulus S and the
analogue mean-vorticity modulus W =
_
W
ij
W
ij
have been used by some au-
thors to dene variable coecients in the equation (primarily C
1
), aimed at
accounting for nonequilibrium eects and dissipation anisotropy (e.g. Speziale
and Gatski 1997). Such modications suer from the same deciency as RNG,
[2] Second-moment turbulence closure modelling 67
Figure 2: Predictions of kinetic energy evolution in an axisymmetric contrac-
tion (left) and expansion (right) with the basic LRR model and RNG modi-
cations in the equation (Hanjalic 1996).
because of the inability to distinguish the sign of mean strain rate and vorticity.
A sounder approach is to use turbulent stress invariants or other turbulence
parameters to modify the coecients, or to dene extra source terms. Craft
and Launder (1991) proposed to modify C
2
as
C
2
=
C
0
2
(1 + 0.65AA
1/2
2
)
, (3.16)
(which also requires modifying C
0
1
). This modication was shown to perform
well in several ows, but within the framework of a nonlinear pressure-strain
model; its use in connection with conventional models may require additional
tuning and validation.
Simpler remedies to improve the prediction of complex industrial ows with-
out having to redene the rest of the model, are also possible: for example,
modications that involve introducing two extra terms in the standard equa-
tion, S
and S
l
, discussed below. Both terms have local eects only in the ow
regions where a remedy is needed, while doing no harm in ows where the con-
68 Hanjalic and Jakirlic
ventional dissipation equation serves well. The term S
, dened as:
S
= C
5
kW
ij
W
ij
, (3.17)
was introduced long ago by Hanjalic and Launder (1980) to enhance the ef-
fects of irrotational straining on the production of (note that the coecient
C
5
= 0.1 while C
1
is now 2.6, instead of the conventional 1.44!). Although
not helpful in dealing with streamline curvature, as originally intended, the
term proved to be helpful in reproducing ows with very strong adverse and
favourable pressure gradients (Hanjalic et al. 1997). It is noted that Shimo-
muras (1993) proposal to account for system rotation has the same form,
except that one of the vorticity vectors is replaced by the intrinsic vorticity,
as discussed later.
The second new term S
l
is dened as
S
l
= max
__
_
1
C
l
l
x
n
_
2
1
_
_
1
C
l
l
x
n
_
2
; 0
_
k
A, (3.18)
where l = k
3/2
/ is the turbulence length scale and C
l
= 2.5. This term
resembles the so-called Yap correction, but has a coordinate-invariant form. It
has been introduced to compensate for excessive growth of the length scale,
and it has proved benecial in the prediction of separating and reattaching
ows. The term has indeed a local character, as demonstrated for example,
in a ow behind a backward-facing step, where the standard IP, QI or SSG
models produce anomalous behaviour of the streamline pattern around ow
reattachment (Hanjalic 1996).
4 Potential of Dierential Second-Moment Closures
for Modelling Complex Flow Phenomena
As mentioned earlier, the major advantage of second-moment closure is that
the Reynolds stress u
i
u
j
need not be modelled directly, but is provided from
the solution of a model dierential transport equation. Other benets can be
deduced from the inspection of the exact transport equation for u
i
u
j
. First,
it contains several more terms. While this poses a new challenge (a need for
modelling) and brings additional uncertainty (unavoidable new empirical co-
ecients), it enables the exact treatment of several important turbulent inter-
actions and, thus, it also enables more subtle features of turbulent ows to be
captured. Other terms can be modelled in a dierent, more appropriate way
than with EVMs because of the availability of u
i
u
j
, rather than simply k.
Some of these features and terms in the equation are considered below fo-
cussing on the potential of DSM, as compared with EVM, to reproduce the
physics of the most common complexities in turbulent ows.
[2] Second-moment turbulence closure modelling 69
Stress production. The rst benet comes from the possibility of treat-
ing the stress generation in its exact form. Turbulent stresses are generated
at the expense of mean ow energy by mean ow deformation, T
ij
, and by
body forces (buoyancy, electromagnetic, etc), (
ij
, and rotation,
ij
. Second
moments are explicitly present in all generation terms: T
ij
and
ij
contain
u
i
u
j
and (
ij
= f
i
u
j
is usually replaced by g
i
u
j
for buoyancy generation, (
is the uctuating temperature, or concentration, and is the corresponding
expansion coecient). The advantage of obtaining second moments from their
own transport equations, instead of from an eddy viscosity/diusivity model
becomes particularly obvious when comparing, e.g., the production of kinetic
energy in both EVM and DSM in a two-dimensional ow:
T
EV M
= 2
t
_
U
1
x
1
_
2
+ 2
t
_
U
2
x
2
_
2
+
t
_
U
1
x
2
+
U
2
x
1
_
2
(4.1)
T
DSM
= u
2
1
U
1
x
1
u
2
2
U
2
x
2
u
1
u
2
_
U
1
x
2
+
U
2
x
1
_
. (4.2)
In thin shear ows (dominated by simple shear U
1
/x
2
) both expressions
give a similar value of T, because the eect of normal straining is negligible. In
more complex ows,
U
1
x
1
,
U
2
x
2
,
U
2
x
1
, . . . can have signicant values and dier-
ent signs. Hence, T
DSM
and T
EVM
may be very dierent. This becomes more
evident (and more important) in ows with a complex strain-rate eld, i.e.
when the extra strain rate originates from streamline curvature, ow skew-
ing, lateral divergence, bulk dilation. Unlike EVMs, DSMs account exactly for
stress production by each component of the strain rate. Even a small extra
strain rate can have a signicant eect on stress production. For example, in
a thin shear ow with a mild curvature, such as in a ow over an airfoil,
U
2
x
1
is much less than
U
1
x
2
, but u
2
1
is much greater than u
2
2
near the airfoil surface,
hence both terms in T
12
=
_
u
2
2
U
1
x
2
+u
2
1
U
2
x
1
_
are of importance (Bradshaw
et al. 1981).
The problem becomes even more serious in ows fully dominated by pure
(normal) strain, because the expression for T
EVM
is always positive and cannot
dierentiate the sign of the strain rate, i.e. it cannot distinguish dilatation from
compression, or uid acceleration from deceleration. The exact contribution
to T
DSM
caused by normal straining is the interaction between specic com-
ponents of turbulent normal stress (positive quantities) with corresponding
components of the normal strain-rate that can be either negative or positive.
A simple example is a ow in a nozzle and diuser of the same shape (of
equal contraction and expansion ratio), where, depending on the inow stress
anisotropy (for the same k), very dierent ow development may be expected
in the two cases (see, for example, the DNS of Lee and Reynolds 1985, Han-
jalic 1996). The standard k- EVM yields the same results for the same initial
level of k and for both the compression and expansion. Other, industrially
more relevant examples are stagnation regions, ow impingement on a solid
70 Hanjalic and Jakirlic
Figure 3: A schematic of stress component interactions in a thin shear ow.
surface, boundary layer recovery after reattachment, recirculating ows, the
central region in the cylinder of a reciprocating engine where the ow is sub-
jected to a cyclic compression and expansion, etc. In all these and other cases
the computations with a linear EVM can never be accurate (e.g. Hadzic et al.
2001).
Stress interaction and anisotropy. Another important turbulence fea-
ture that can be reproduced only by models based on the stress-transport
equation, is the stress interaction. In ows with a preferential orientation of
the velocity eld, the dominant strain rate component feeds energy into se-
lective stress component(s). Pressure uctuations redistribute a part of the
largest stress into other components (and also reduce the shear stress) making
turbulence more isotropic. An illustration of stress interaction and the role of
pressure uctuations is given in Figure 3 for a simple thin shear ow, which
shows a general ow chart of turbulence energy (stress) components.
The exact treatment of the stress generation and the possibility of account-
ing for stress-component interactions gives a better prospect for modelling the
important turbulence parameter, the stress anisotropy, which governs, to a
large extent, the wall heat and mass transfer. This is particularly the case in
regions with either small or no wall shear, such as around impingement, sep-
aration or reattachment, where the transport of mean momentum and heat
transfer do not show any correlation or analogy. In these regions, the heat and
mass transport are governed by the wall-normal turbulent stress component,
and its accurate prediction is a crucial prerequisite for computing accurately
the transport phenomena at a solid surface.
Streamline curvature. Most complex ows involve strong streamline cur-
vature, which may occur locally even if the ow boundaries are not curved (e.g.
the curved shear layer and separation bubble behind a step, or on a plane wall
due to adverse pressure gradient), or the whole ow can be curved by the
[2] Second-moment turbulence closure modelling 71
Figure 4: Illustration of eects of streamline curvature: ow in a curved pipe.
imposed boundaries. In the latter case, the centrifugal force aects the ow
and turbulence in much the same way as the Coriolis force in rotating ows
(Ishigaki 1996).
The streamline curvature generates an extra strain rate, which can exert
a signicant eect on stress production. Streamline curvature attenuates the
turbulence when the mean ow angular momentum increases with curvature
radius (e.g. in ow over a convex surface stabilizing curvature), whereas it
amplies the turbulence in the opposite situation (e.g. over a concave surface
destabilizing curvature). The criterion can be related to the direction of the
uid rotation vector or the vorticity
k
=
ijk
W
ji
(which denes the sign
of the shear stress) with respect to the mean angular momentum vector: if
these two vectors have the same direction, the turbulence will be attenuated,
if opposite directions, the turbulence will be enhanced, Figure 4. Eects of
streamline curvature are directly related to the stress production and stress
interactions. Linear EVMs fail to account for curvature eects in most turbu-
lent ows because of their inability to reproduce normal stresses which appear
explicitly in the production term T
ij
. Of course, one can model the eects
of streamline curvature with simple models by ad hoc corrections. This has
been done in EVM schemes, e.g., by expressing the coecients in the sink
term of the dissipation equation in terms of a curvature Richardson number
Ri
t
= (k/R)
2
U
(RU
ijk
, where
j
is the system angular velocity. System rotation directly inuences the inten-
sity of stress components through stress redistribution. It is also known that
rotation aects the turbulence scales. As mentioned earlier, the stress trans-
port equation contains the exact rotational production term
ij
. Because the
Coriolis force acts perpendicular to the velocity vector U
k
, it has no direct
inuence on the k-budget (
ii
= 0), but redistributes the stress among com-
ponents and thus modies the net production of individual stress components.
Rotation also causes a decrease in dissipation, even in isotropic turbulence (see
the DNS by Bardina et al. 1985).
The eects of rotation are illustrated in a plane channel rotating about
an axis perpendicular to the main ow direction with an angular velocity
[2] Second-moment turbulence closure modelling 73
vector (system vorticity)
j
(here
3
), Figure 5. Note that
j
is aligned
with the vector of the mean shear vorticity
k
=
ijk
W
ji
(here
3
= W
21
W
12
=
U
1
x
2
, since
U
2
x
1
= 0). The uctuating Coriolis force amplies the
turbulence on the pressure side, destabilizing the ow, while near the suction
side the tendency is opposite. Comparison of the directions of the angular
velocity and shear vorticity vectors gives a convenient indication of the eects
on turbulence. If
3
and
3
have the same direction (
3
3
), rotation
attenuates the turbulence (stabilizing eect, suction side). In contrast, if
3
and
3
have opposite directions (
3
3
), rotation will amplify the
turbulence (destabilizing eect, pressure side). At high rotation rates, the
stabilizing eect may completely damp the turbulence on the suction side
causing local laminarization. The eect of rotation is often dened in terms of
the local Rossby number, dened as the ratio of the mean shear vorticity to
the system vorticity Ro =
l
/
l
(here Ro =
U
1
/x
2
2
3
), or its reciprocal S =
1/Ro. The sign of Ro or S indicates whether the eect on turbulence will be
amplifying () or attenuating (+). It is noted that for a general classication,
the bulk rotation number N = 2h/U
b
(the inverse of the bulk Rossby number
not to be confused with the local one dened above), is often used, where h
is the channel half width and U
b
is the bulk velocity. The bulk rotation number
serves only to indicate the intensity of rotation, not a stability criterion, nor
as a parameter for modifying the turbulence model. The local Rossby number
Ro (or S) has served in the past to modify the simple k- model by expressing
one of the coecients in the equation in terms of Ro. However, the stress
transport equation accounts exactly for the eects of rotation through the
exact rotational generation term
ij
(stress generation by uctuating Coriolis
force) in equation (1). The table below lists the components of
ij
for the
plane channel example (note,
3
= ).
ij 11 22 33 12
T
ij
2u
1
u
2
dU
1
dx
2
0 0 u
2
2
dU
1
dx
2
ij
4u
1
u
2
4u
1
u
2
0 2(u
2
1
u
2
2
)
In addition to the exact
ij
term in the DSM, system rotation aects also
the stress redistribution induced by uctuating pressure. Hence, the eect of
rotation should be accounted for in the model of the pressure-strain term.
A simple way to do this is to replace T
ij
in
ij,2
(see below) by the total
stress generation T
ij
+
ij
. However, in order to ensure the material frame
indierence, T
ij
should be replaced by T
ij
+
1
2
ij
(e.g. Launder et al. 1987). So
far, no convincing proof has been provided as to which of the two modications
performs better, but both versions have led to substantial improvement in
predicting rotating ows.
Because
ii
= 0, the basic k- model, without modications, cannot mimic
the rotation eects on turbulence. Modications are introduced usually via ad-
ditional term(s) in the -equation, analogous to the modications for curvature
74 Hanjalic and Jakirlic
eects, in terms of rotation Richardson number Ri
r
= 2(k/)
2
(U
1
/x
2
),
or by making the coecient in the eddy-viscosity a function of a rotation
parameter.
In fact, the -equation should, in principle, be modied to account for the
eects of rotation on turbulence scale even in conjunction with DSM models.
Bardina et al. (1987) suggested an additional term:
C
_
1
2
X
ij
X
ij
_
1/2
, (4.3)
where C
= 0.15 and X
ij
= W
ij
+
kji
k
is the intrinsic mean vorticity
tensor.
5
Shimomura (1993) proposed a slightly dierent formulation, C
k
l
l
,
which improved predictions of ow in a rotating plane channel (Jakirlic et al.
1998). It is interesting to note that this term resembles closely that proposed
by Hanjalic and Launder (1980) for nonrotating ows to enhance the eect of
irrotational straining on (energy transfer through the spectrum, see below),
C
3
k
l
l
. This term can be combined with that of Shimomura to yield a joint
term which would be eective in both nonrotating and rotating ows, i.e.
C
3
k
l
(
l
C
/C
3
l
).
Successful DSM computation of rotating channel ows with no modication
of -equation were reported by Launder et al. (1987) for moderate rotation.
These results indicate that the eect of rotation is stronger on the stresses
than on the turbulence scale, which illustrates further the advantage of DSM,
as compared with EVM. The importance of integration up to the wall (using
a low-Re-number model) in ows where the rotation is suciently strong to
cause laminarization on the suction side, has been demonstrated by Jakirlic et
al. (1998) and Pettersson and Andersson (1997), from whose work examples
are drawn in [4], (see also Dutzler et al. 2000).
Swirl. Swirling ows can be regarded as a special case of uid rotation with
the axis usually aligned with the mean ow direction (longitudinal vortex) so
that the Coriolis force is zero. Swirl enhances the turbulent mixing and often
induces recirculation. These features are exploited in internal combustion (IC)
engines and gas-turbine combustors, and for heat transfer enhancement (vor-
tex generators). A common feature of such ows is that the swirl is strong and
conned to short cylindrical enclosures, whose length is of the same order as
the duct diameter. Another type of conned swirl, usually of weak intensity,
is encountered in long tubes, either imposed at the tube entrance, or self-
generated by secondary motions as a consequence of upstream multiple tube
bends (double- or S-bends in dierent planes). A third kind is the swirling
motion inside rotating cylinders or pipes, either developing or fully developed.
5
or, in terms of rotation (dual) vectors, C
(X
l
X
l
)
1/2
, where X
l
=
1
2
l
+
l
[2] Second-moment turbulence closure modelling 75
a. b. c.
Figure 6: A schematic of tangential ow velocity in (a) rotating pipe ow, (b)
swirling pipe ow and (c) in a free swirl.
A special case is the spin-down ow when a rotating pipe is suddenly brought
to a standstill used sometimes in studying experimentally the eect of swirl
in a piston-cylinder assembly related to IC engines. Unconned free swirling
ows such as swirling jets, represent yet another category relevant to swirling
burners, which possess some specic features and are known to modify sub-
stantially the ow characteristics even at a very low swirl intensity.
Although all these cases deal with essentially the same phenomenon ro-
tating uid in axisymmetric geometries their predictions pose dierent chal-
lenges. To illustrate the eect of swirl on turbulence one may follow the same
arguments as outlined earlier in the context of curved ows and system rota-
tion, Figure 6, i.e. by considering the directions of the bulk ow rotation vector
and of the uid element vorticity. In short, the turbulence will be damped in
regions where the rotation vector of the uid vorticity in the ow cross-section
(
z
in Figure 6) has the same direction as the bulk ow rotation vector
z
,
while turbulence will be enhanced in regions where these two vectors have
opposite signs. For example, in a rotating pipe, the uid rotation will damp
turbulence over the whole cross-section, with greatest eect near the pipe wall,
which may cause the ow to laminarize locally, if this eect is stronger than
the turbulence production by shear due to the radial gradient of axial velocity.
If the pipe wall is stationary, with a swirl imposed at its entrance (or gener-
ated by the sudden stopping of pipe rotation, spin-down), in the outer region
close to the pipe wall the turbulence production will be enhanced because
z
has an opposite sign to
z
. In the core region the orientation of
z
is the
same as that of
z
and turbulence will be damped. In a strong swirl the core
may even be laminarized. The same arguments apply for swirling jets. EVM
computations employ swirl-dependent coecients in the modelled equations,
but generally with limited success. In fact, the original DSM computations
did not greatly improve the prediction of a swirling jet. Note that the ro-
tational term
ij
is absent in an inertial coordinate frame and the eect of
swirl should be accounted for by either modifying the models of some terms,
adding extra terms in model equations, or by expressing coecients in terms
of swirl parameters. The simple Isotropization of Production (IP) model of
76 Hanjalic and Jakirlic
ij,2
(see below) seems to perform better than the LRR QI (quasi-isotropic)
model, (Launder and Morse 1979). By recognizing that a major deciency lies
in
ij,2
, Fu, Leschziner and Launder (1987), proposed the inclusion of convec-
tion (
ij
to ensure material frame indierence, (negligible eects in nonswirling
ows):
ij,2
= C
2
__
T
ij
+
ij
2
3
T
ij
_
_
(
ij
2
3
(
ij
__
. (4.4)
It was found, however, that this modication produces little eect (at least in
weak swirls) since the last term in equation (4.4) is smaller than other terms
in the expression. Still better predictions are expected with improvements of
the scale equation.
Several cases of swirling ows reported in the literature show obvious supe-
riority of the DSM model. This is particularly the case for strong swirl in a
combustion-chamber type of geometry (e.g. Hogg and Leschziner 1989; Jakirlic
1997). Similar improvements are achieved for a swirl in a long pipe (Jakirlic
1997).
Mean pressure gradient. The transport equations for turbulent stresses
and scale properties do not contain mean pressure. However, the mean pres-
sure gradient modies the mean rate of strain and depending on the sign
amplies or attenuates turbulence. Extreme cases are laminarization of an
originally turbulent ow, when dp/dx 0 (severe acceleration) and ow sepa-
ration when dp/dx 0 (strong deceleration). Both extreme cases represent a
challenge to turbulence modelling. Turbulent ows subjected to periodic vari-
ations of pressure gradient or other external conditions (pulsating and oscillat-
ing ows) fall into the same category with an additional feature: a hysteresis
of the turbulence eld lagging in phase behind the mean ow perturbations.
Linear EVMs cannot capture these features; DSMs perform generally better,
though additional modications (mainly in the scale equation, see below) are
needed. Wall functions are inapplicable for specifying boundary conditions
and integration up to the wall, with appropriate modications of the model, is
essential to reproduce these phenomena accurately. Predictions with a low-Re-
number DSM of the turbulence evolution and decay in an oscillating ow in
a pipe at transitional Re numbers, displaying a visible hysteresis of the stress
eld, were reported by Hanjalic et al. (1995). An overview of the performance
of DSMs in ows with dierent pressure gradients involving separation is given
in Hanjalic et al. (1999).
Secondary currents and longitudinal vortices. This term refers usually
to a secondary motion with longitudinal, streamwise vorticity
1
, superim-
posed on the mean ow in the x
1
-direction. Skew-induced (pressure-driven)
1
(Prandtls 1st kind of secondary ows) is essentially an inviscid process,
generated by the bending of existing mean vorticity. Viscous and turbulent
[2] Second-moment turbulence closure modelling 77
Figure 7: Schematic of wing-tip vortices and secondary currents in conduits of
noncircular cross-section.
stresses cause
1
to diuse. Turbulent-stress-induced
1
(Prandtls 2nd kind of
secondary ow), is generated by the turbulent stresses due to anisotropy of the
Reynolds-stress eld. Secondary motions can arise in the form of a cross-ow
such as in 3D thin shear ows (
1
U
3
/x
2
), or in the form of recirculating
cross-currents such as occur in noncircular ducts (
1
= U
3
/x
2
U
2
/x
3
),
see Figure 7. Of course, a secondary current can be imposed on the ow in
order to enhance mixing or heat and mass transfer, such as in the case of
vortex generators. Skew-induced secondary velocity can be high, while the
stress-induced currents are usually weak, though still very important for tur-
bulent transport. The importance of turbulence in the dynamics of mean-ow
vorticity can be illustrated by considering the vorticity transport equation,
which, for the
1
(streamwise) component, reads:
D
1
Dt
=
2
1
x
k
x
k
+
1
U
1
x
1
. .
vortex stretching
+
2
U
1
x
2
+
3
U
1
x
3
. .
skew-induced
1
(vortex bending)
2
u
2
u
3
x
2
2
+
2
u
2
u
3
x
2
3
+
2
x
2
x
3
(u
2
2
u
2
3
)
. .
stress-induced generation of
1
. (4.5)
Skew-induced secondary motion, driven essentially by mean-ow deformation
and by the mean pressure eld does not require a complex turbulence model.
However, stress-induced motion cannot be handled with an EVM and requires
a model which can compute individual turbulent stress components (DSM or
ASM). In fully developed ows in ducts of non-circular cross-section the ap-
plication of ASM is sucient to capture the stress-induced secondary motion.
Illustrations of the prediction of secondary currents in square ducts have been
published inter alia by Demuren and Rodi (1984), in pipe bends by Anwer et
al. (1989), and in U-ducts by Iacovides et al. (1996). See also Chapters [1, 3, 4].
Three-dimensionality. Finally, a few remarks should be added concern-
ing the ow three-dimensionality eects. Even a mild three-dimensionality of
the mean ow produces signicant changes in turbulence structure. In strong
78 Hanjalic and Jakirlic
Figure 8: Schematic of mean velocity proles in ow with a longitudinal vortex
and in a pressure-skewed ow.
cross-ows the eects can be dramatic as, for example, in the case of a unidi-
rectional uid stream with a superimposed longitudinal vortex, such as is pro-
duced by a vortex generator to enhance wall heat transfer. The resulting mean
velocity proles may look very skewed, as shown in Figure 8, which is dicult
to reproduce by simple eddy-viscosity or similar models. Other examples of
relatively simple 3D boundary layers include wing-body (or blade-rotor) junc-
tions, ows encountering laterally moving walls imposing a transverse shear,
such as in the stator-rotor assembly in turbomachinery. In fully 3D separating
ows the problem is even more challenging. Current turbulence models have
been developed on the basis of our knowledge of 2D ows. Plausible extensions
to the third dimension do not always yield satisfactory results. This is partic-
ularly the case for linear eddy-viscosity models. Even in a simple 3D boundary
layer, the eddy viscosity is not isotropic, as discussed earlier, i.e.:
u
1
u
2
U
1
/x
2
,=
u
3
u
2
U
3
/x
2
; (4.6)
This nding illustrates that complex ows require turbulence models of a
higher order than the EVM. Further illustrations of the inadequacy of the eddy
viscosity concept in 3D ows can be found in Hanjalic (1994) and elsewhere.
5 Advanced Models for Near-wall and Low
Re-number Flows
5.1 The wall-function approach and its deciency
All industrial CFD codes use wall functions to serve as the wall boundary
conditions. The viscosity-aected near-wall region is bridged by placing the
[2] Second-moment turbulence closure modelling 79
rst grid point outside the viscous layer in the fully turbulent region: such
high-Reynolds-number turbulence closures can thus be used for ows at high
bulk Reynolds numbers. This simplies to a great degree both the modelling
and computational tasks. By obviating the need to resolve steep gradients very
close to the wall, a relatively coarse numerical grid often suces for reaching
grid-independent solutions. However, the wall functions have been derived on
the basis of wall scaling in attached boundary layers in local turbulent energy
equilibrium. Their use in nonequilibrium ows is therefore considered inappro-
priate, particularly in separating and recirculating ows, around reattachment,
in strong pressure gradients and in rotating ows. Figure 9 shows two examples
of the deviation of the mean velocity prole from the conventional logarithmic
law which serves as a basis for most wall functions for the mean velocity. The
upper three proles correspond to three locations in a ow behind a backward-
facing step in the recovery zone downstream from reattachment. Even greater
deviation is exhibited in the recirculation zone. The lower gure shows axial
and tangential velocity in a swirl generated by cylinder rotation some time af-
ter the rotation was stopped (spin-down). Good agreement with experiments
was achieved only with the application of a low-Re-number DSM integrated
up to the wall.
Various modications have been proposed to improve and extend the valid-
ity of wall functions to non-equilibrium and separating ows, but none of the
proposals showed general improvement. The incorporation of pressure gradi-
ent is most straightforward and can easily be done by extending the near-wall
ow analysis, e.g. Ciofalo and Collins (1989), Kiel and Vieth (1995), Kim and
Choudhury (1995). Such modications generally lead to some improvement
of attached thin shear wall ows with pressure gradient (with convection still
being neglected), but their validity is conned only to such situations. A more
general two-layer approach, based on splitting the wall layer in viscous and
nonviscous parts with assumed variation of shear stress and kinetic energy
in each layer, was earlier proposed by Chieng and Launder (1980) and John-
son and Launder (1982). The assumed proles for uv and k enable the stress
production and dissipation over the rst control volume next to a wall to be
integrated separately, instead of assuming wall-equilibrium values. However,
despite some improvement of wall friction and heat transfer behind a back step
and sudden pipe expansion, the approach still has serious deciencies. More
general wall functions that would be applicable to various complex ows (with
separation, stagnation, laminar-to-turbulent transition, buoyancy and other
eects) are still awaited. Development of such wall functions has recently been
taken up again by the UMIST group, and the initial results are encouraging
(Craft et al. 2001, 2002).
80 Hanjalic and Jakirlic
Figure 9: Mean velocity proles at selected locations in the recovery region
behind a backward-facing step ow (Hanjalic and Jakirlic 1998) (above) and
proles of axial and tangential velocity in a spin-down ow in an engine cylin-
der at the beginning of compression (below), (Jakirlic et al. 2000).
5.2 Models with near-wall and low-Re-number modications
Integration up to the wall is a more generally applicable strategy than adopt-
ing wall functions. This approach requires rst the introduction of substantial
modications to the turbulence models in order to account for complex wall
eects, primarily for viscosity (low-Re-number models), but also for non-
viscous ow blocking and pressure-reection by a solid wall. This in turn
requires a much ner grid resolution in and around the viscosity-aected wall
sublayer, and, consequently, increases demands for computational resources,
with often formidable requirements on the numerical solver to ensure conver-
gent solutions. Low-Re-number models are at present available both at the
EVM and DSM level. While primarily developed for treating the near-wall
[2] Second-moment turbulence closure modelling 81
viscous region, some of the models are reasonably successful also in predict-
ing turbulence transition. In fact, general low-Re-number models, which can
distinguish appropriately between the non-viscous blocking and viscous wall
eects are indispensable for predicting the laminar-to-turbulent and reverse
transition (at least the forms of transition which can be handled within the
framework of the Reynolds averaging approach, such as by-pass transition, or
the revival of inactive background turbulence and its laminarization). By-pass
transition is the subject of [17] and [18].
A number of proposals for modifying DSMs to account for low-Reynolds-
number and wall-proximity eects can be found in the literature. These mod-
ications are based on a reference high-Re-number DSM which serves as the
asymptotic model to which the modications should reduce for suciently
high Reynolds number and at a sucient distance form a solid wall. Han-
jalic and Launder (1976), Launder and Shima (1989), Hanjalic and Jakirlic
(1993), Shima (1993), Hanjalic et al. (1995) base their modications on the
basic DSM with linear pressure-strain models, in which the coecients are
dened as functions of turbulent Reynolds number and invariant turbulence
parameters. Earlier models also use the distance from a solid wall. More recent
models are based on DNS data and term-by-term modelling, which ensures
model realizability, compliance with the near-wall stress two-component limit,
as well as with the limit of vanishing turbulence Reynolds number. Launder
and Tselepidakis (1991), Craft and Launder (1996), Craft (1997) use the cubic
pressure-strain model in which the coecients were determined by imposing,
in addition to basic constraints discussed earlier, the two-component limit.
A larger number of terms with associated coecients reduces the need for
introducing functional dependence and additional turbulence parameters.
As mentioned earlier, Durbin (1991, 1993) uses elliptic relaxation to account
for non-viscous blockage eect of a solid wall, whereas a switch of time and
length scale from the high-Re-number energy-containing scales to Kolmogorov
scales (lower bound), when the latter become dominant, accounts for viscosity
eects
6
.
All models require a very ne numerical mesh within the viscosity aected
wall sub-layer, so that the computation becomes time-consuming and imprac-
tical for more complex ow cases. Durbins elliptic relaxation approach seems
to be somewhat less demanding in this respect, but at present it cannot be
used to predict transitional phenomena.
In principle, all modications involve the following:
Inclusion of viscous diusion in all equations;
6
Jakirlic (1997, p. 44) has noted that the current ER proposal appears to be inconsistent:
the time scale and the length-scale switching (and the introduction of viscous eects in each)
occur at very dierent distances from a wall: e.g. in a plane channel ow at Re
m
= 14000
at y
+
6 and at y
+
150, respectively. Moreover, the length scale switch implies that the
Kolmogorov scale (and, consequently, the viscous eect) prevails over almost the whole ow
width for this Reynolds number.
82 Hanjalic and Jakirlic
Provision of a non-isotropic model for
ij
;
Addition of further terms to the -equation (supposedly to model P
3
)
Replacement of certain constant coecient by functions of turbulent
Re number, Re
t
= k
2
/(), dimensionless wall distance and/or other
turbulence parameters.
5.2.1 A low-Re-number DSM
An example of a low-Re-number second-moment closure (DSM) is the model
proposed by Hanjalic et al. (1995) (see also Hanjalic et al. 1997, Jakirlic 1997,
Hadzic 1998). This model has been used with reasonable success in a number
of 2D and some 3D high- and low-Re-number wall ows including cases of
severe acceleration (laminarizing 2D and 3D turbulent boundary layers), by-
pass and separation-induced laminar-to-turbulent transition, oscillating ows
at transitional and high Re numbers, rotating and separating ows. Some
examples will be shown in the next subsection.
The model is based on the basic DSM (Section 2) in which a low-Re-number
version of the equation is used. The coecients in the u
i
u
j
equation are
expressed as functions of Re
t
and invariants of the stress and dissipation rate
tensors to account for the viscous and inviscid wall eect respectively. The
model satises the two-component and vanishing Reynolds number limits, thus
enabling integration up to the wall.
Viscosity has a scalar character (it dampens all stress components and is
independent of the wall distance and its topology), and it is only indirectly
related to the wall presence via no-slip conditions. Hence, its eect can be
conveniently accounted for through the turbulent Reynolds number Re
t
=
k
2
/(). This should be formulated in a general manner to be applicable both
close to a wall and away from it.
Inviscid eects are basically dependent on the distance from a solid wall
and its orientation, as seen from the Poisson equation for uctuating pres-
sure (Stokes term). This term accounts for the wall blockage and pressure
reection. However, the DNS data for a plane channel show that this term de-
cays fast with the wall distance and becomes insignicant outside the viscous
layer. Yet, a notable dierence in the stress anisotropy between a homogeneous
shear ow and an equilibrium wall boundary layer for comparable shear in-
tensity shows that the eect of the walls presence extends much further from
the wall into the log-layer. This indicates an indirect wall eect through the
strong inhomogeneity of the mean shear rate, a fact that is ignored by almost
all available pressure-strain models
7
.
In view of above discussion, the use of wall distance through the function f
w
and wall orientation represented by the unit normal vector n
i
in the adopted
7
An exception is the model of Craft and Launder (1996), see [3].
[2] Second-moment turbulence closure modelling 83
models for
w
ij
seems reasonable, despite some opposing views in the literature.
However, these modications, introduced for and tuned in high-Re-number
wall attached ows, cannot account for inviscid eects closer to a wall (in the
buer and viscous layer). Wall impermeability imposes a blocking on the uid
velocity and its uctuations in the normal direction, causing the stress eld to
be strongly non-isotropic. This fact has been exploited by Hanjalic et al. (1995)
by introducing, in addition to Re
t
, invariants of both the turbulent-stress and
dissipation-rate anisotropy, a
ij
= u
i
u
j
/k 2/3
ij
and e
ij
=
ij
/ 2/3
ij
,
respectively, A
2
, A
3
, E
2
and E
3
, as parameters in the coecients. This enables
one to account separately for the wall eect on the anisotropy of the stress-
bearing and dissipative scales, shown by the DNS data to be notably dierent
(Hanjalic et al. 1997, 1999), Figure 10a. The sensitivity of stress invariants to
pressure gradient is illustrated in Figure 10b, where Lumleys two-component
(atness) parameter A is plotted for boundary layers in zero, favourable and
adverse pressure gradients. Also, the predictions of A with the here presented
low-Re-number second-moment closure model are shown.
Based on the above arguments, the following modications were introduced:
Stress transport equation:
ij
: Linear models are adopted for the slow, rapid and wall terms, equations
(2.10), (2.11), (2.14) and (2.15), in which the coecients are dened as
follows:
C
1
= C +
AE
2
; C = 2.5AF
1/4
f; F = min0.6; A
2
f = min
_
_
Re
t
150
_
3/2
; 1
_
; f
w
= min
_
k
3/2
2.5x
n
; 1.4
_
C
2
= 0.8A
1/2
; C
w
1
= max(10.7C; 0.3); C
w
2
= min(A; 0.3),
where
A = 1
9
8
(A
2
A
3
); A
2
= a
ij
a
ji
; A
3
= a
ij
a
jk
a
ki
; a
ij
=
u
i
u
j
k
2
3
ij
E = 1
9
8
(E
2
E
3
); E
2
= e
ij
e
ji
; E
3
= e
ij
e
jk
e
ki
; e
ij
=
ij
2
3
ij
.
ij
, the stress dissipation rate model:
ij
= f
s
ij
+ (1 f
s
)
2
3
ij
ij
=
k
[u
i
u
j
+ (u
i
u
k
n
j
n
k
+u
j
u
k
n
i
n
k
+u
k
u
l
n
k
n
l
n
i
n
j
)f
d
]
1 +
3
2
u
p
u
q
k
n
p
n
q
f
d
f
s
= 1
AE
2
; f
d
= (1 + 0.1Re
t
)
1
.
84 Hanjalic and Jakirlic
Figure 10: Lumleys two-component (atness) parameter for turbulent stress
(A) and its dissipation rate (E) in the recirculation region behind a back-
ward facing step (above); Stress atness parameter in boundary layers in zero,
favourable and adverse pressure gradients (below).
equation:
Equation (2.18) for is modied to the form:
D
Dt
=
x
k
_
kl
+C
u
k
u
l
x
l
_
+
_
C
1
T +C
3
( +C
4
k
U
k
x
k
C
2
f
_
k
+C
u
j
u
k
2
U
i
x
j
x
l
2
U
i
x
j
x
l
+S
+S
l
, (5.1)
where
f
= 1
C
2
1.4
C
2
exp
_
_
Re
t
6
_
2
_
, = 2
_
k
1/2
x
n
_
,
and S
and S
l
have been dened earlier, equations (3.17) and (3.18).
[2] Second-moment turbulence closure modelling 85
6 Some Illustration of DSM Performance
Many complex ows contain several types of strain rate and it is not easy to
distinguish their individual eects on turbulence. Moreover, dierent types of
strain or other eects may dominate dierent regions so that improvements in
one region can lead to a deterioration in others. Improvements can often be
achieved with dierent remedies and it is not always clear which modications
have a better physical meaning. Some illustrations of the eects of various
model modications are provided in Figure 11 for the case of a turbulent
jet issuing from a circular tube and impinging normally on a plane wall. In
this ow the eect of mean pressure is dominant so that the inuence of
the turbulence model on the predicted mean velocity eld is not strong, at
least in the stagnation region, Figure 11a. However, the predicted turbulence
depends greatly on the model applied, as shown in Figure 11b, where the
performance of model variants is shown. The standard k- model yields a
far too high kinetic energy, because of poor modelling of the normal stress
production, as discussed earlier. The next three curves illustrate the eect of
the model of the pressure-strain term in the various DSMs. The basic DSM
(BDSM) is that of Launder, Reece and Rodi (1975) (LRR-IP) with Gibson
and Launder (1978) wall-echo correction (hereafter denoted LRRG), performs
better, though still not satisfactorily owing to the inadequacy of the wall-
echo term for impinging ows. Indeed, better results are obtained by simply
omitting the wall-echo term. Further improvement is achieved when using the
pressure-strain model of Speziale, Sarkar and Gatski (1991), denoted as SSG,
which contains no extra wall-echo term. The application of the cubic model
of Craft and Launder (1991) was reported to perform best (not shown here),
but at the expense of greater complexity. While the above discussion clearly
demonstrates the importance of the pressure-strain model, it would be wrong
to conclude that this is the sole cause of unsatisfactory performance. The scale
equation (here ) in its simplest form is clearly inadequate to model complex
ows with extra strain rates. The eects of three possible modications, each
involving an extra term, are illustrated by the last three curves in Figure
11b, showing further possibilities for improving the predictions. Accounting
for the eect of irrotational strain (term S
, /U
ij
. (2.1)
104 Craft and Launder
The exact generation term P
ij
u
i
u
k
U
j
/x
k
u
j
u
k
U
i
/x
k
, whilst the
remaining terms, which require modelling, are the pressure-strain correlation
ij
, diusion d
ij
and dissipation rate
ij
. In the shear ows discussed here,
diusion is a relatively unimportant process, so the simple gradient diusion
model of Daly and Harlow (1970),
d
ij
=
x
k
__
lk
+c
s
k
u
k
u
l
_
u
i
u
j
x
l
_
, (2.2)
is often employed. The modelling of the remaining terms is made so as to
ensure compliance with the two-component limit (TCL). Broadly, two strate-
gies are adopted, which are exemplied in the modelling of the pressure-strain
processes.
2.1 Pressure-Strain Processes
In many ows, it is the pressure-strain correlation
ij
which is the most impor-
tant term requiring modelling, and consequently much eort has been put into
developing improved models for this process. As reported in [2], equation (6)
et seq., integration of the Poisson equation for pressure uctuations for regions
where the surface integral is unimportant leads to
ij
p
_
u
i
x
j
+
u
j
x
i
_
(2.3)
=
1
4
_
V ol
_
3
u
k
u
l
u
i
r
l
r
k
r
j
+
3
u
k
u
l
u
j
r
l
r
k
r
i
_
dV
[r[
. .
ij1
1
2
_
V ol
_
2
u
l
u
i
r
k
r
j
+
2
u
l
u
j
r
k
r
i
_
U
k
r
l
dV
[r[
. .
ij2
,
where non-primed quantities are evaluated at the point where
ij
is being
determined, whilst primed quantities are evaluated at positions within the
integration volume at a displacement of r from this point.
From this, it can be seen that there are two distinct contributions to
ij
:
one involving interactions between uctuating quantities and one dependent
on mean strain rates. In buoyancy-aected ows there is a further contribution,
which will be considered separately in [14].
The volume of integration in equation (2.3), although formally being the
entire uid domain, can in practice be regarded as the region where the time
averaged two-point correlations are non-zero, corresponding to some relatively
small region surrounding the point at which
ij
is being evaluated with a radius
typically of the local integral lengthscale.
[3] Closure modelling near the two-component limit 105
It is noted that equation (2.3) contains no surface integral discussed both
in [2] and, in more detail, in [4]. This represents a fundamental dierence of
strategy between closure schemes which are otherwise of the same type. The
philosophy explored at UMIST is that if one ensures compliance of the model of
ij
with the two-component limit, there should, in many cases, be no require-
ment for any further wall correction - at least if one remains outside the buer
layer where the eects of very rapid spatial variations must be accounted for.
2.1.1 Mean-Strain (or Rapid) Part of
ij
If the mean strain is assumed to vary much more slowly in space than the two-
point correlation gradients in equation (2.3), it can be regarded as uniform over
the volume of integration, so that
ij2
can be modelled as
ij2
=
_
X
li
kj
+X
lj
ki
_
U
k
x
l
, (2.4)
where the tensor X
li
kj
represents the integral of the two-point velocity-derivative
correlations:
X
li
kj
=
1
2
_
V ol
2
u
l
u
i
r
k
r
j
dV
[r[
. (2.5)
This approach has been employed by Naot et al. (1973) and Launder et al.
(1975) to derive models of
ij2
in which X
li
kj
is simply a linear function of the
Reynolds stresses:
X
li
kj
= u
l
u
i
kj
+ (u
i
u
j
lk
+u
l
u
j
ik
+u
l
u
k
ij
+u
i
u
k
lj
)
+u
k
u
j
il
+k (
lj
ik
+
lk
ij
) +k
il
kj
, (2.6)
where the Greek symbols are coecients to be determined. One can note that
the exact integral in equation (2.5) satises
Continuity: X
li
ki
= 0
Normalization: X
li
kk
= 2u
l
u
i
.
By applying these two constraints, all the coecients except one in equa-
tion (2.6) can be determined and the resultant model, known as the Quasi
Isotropic (QI) Model, may be written as
ij2
=
+ 8
11
(P
ij
1
/
3
ij
P
kk
)
30 2
55
k
_
U
i
x
j
+
U
j
x
i
_
8 2
11
(D
ij
1
/
3
ij
D
kk
), (2.7)
where D
ij
u
i
u
k
U
k
/x
j
u
j
u
k
U
k
/x
i
.
106 Craft and Launder
However, it is not possible to choose a value of that will enable this linear
form to satisfy the 2-component limit, which requires
2
= 0 if u
= 0.
An obvious extension of the approach is to recognise that the two-point
correlations appearing in the integral of equation (2.5) will not depend linearly
on the second-moments, and thus to allow X
li
kj
to be a nonlinear function of
the Reynolds stresses. If one includes both quadratic and cubic terms, then
the most general expression satisfying the required symmetry properties can
be written as
X
li
kj
k
=
1
li
kj
+
2
(
lj
ki
+
lk
ij
)
+
3
a
li
kj
+
4
a
kj
li
+
5
(a
lj
ki
+a
lk
ij
+a
ij
lk
+a
ki
lj
)
+
6
a
li
a
kj
+
7
(a
lj
a
ki
+a
lk
a
ij
) +
8
a
lm
a
mi
kj
+
9
a
km
a
mj
li
+
10
(a
lm
a
mj
ki
+a
lm
a
mk
ij
+a
im
a
mj
lk
+a
km
a
mi
lj
)
+
11
a
mn
a
mn
li
kj
+
12
a
mn
a
mn
(
lj
ki
+
lk
ij
)
+
13
a
li
a
km
a
mj
+
14
a
kj
a
lm
a
mi
+
15
(a
lj
a
km
a
mi
+a
lk
a
im
a
mj
+a
ij
a
lm
a
mk
+a
ki
a
ml
a
mj
)
+
16
a
mn
a
np
a
pm
li
kj
+
17
a
mn
a
np
a
pm
(
lj
ki
+
lk
ij
)
+
18
a
mn
a
mn
a
li
kj
+
19
a
mn
a
mn
a
kj
li
+
20
a
mn
a
mn
(a
lj
ki
+a
lk
ij
+a
ij
lk
+a
ki
lj
). (2.8)
The approach outlined below, following the analysis of Fu (1988), is to
assume the coecients
1
, . . . ,
20
to be constants, and to apply the continu-
ity, normalization and 2-component-limit constraints in order to determine as
many of the coecients as possible.
Applying the continuity constraint (X
li
ki
= 0), and making use of the Cayley
Hamilton theorem, leads to six equations:
1
+ 4
2
= 0 (2.9a)
3
+
4
+ 5
5
= 0 (2.9b)
6
+
7
+
8
+
9
+ 5
10
= 0 (2.9c)
10
+
11
+ 4
12
= 0 (2.9d)
13
+
14
+ 4
15
+ 2
18
+ 2
19
+ 10
20
= 0 (2.9e)
16
+ 4
17
+
1
/
3
(
13
+
14
+ 2
15
) = 0 (2.9f)
Similarly, the normalization constraint leads to a further six equations:
3
1
+ 2
2
= 4/3 (2.10a)
3
3
+ 4
5
= 2 (2.10b)
2
7
+ 3
8
+ 4
10
= 0 (2.10c)
9
+ 3
11
+ 2
12
= 0 (2.10d)
13
+ 2
15
+ 3
18
+ 4
20
= 0 (2.10e)
4
15
+ 9
16
+ 6
17
= 0 (2.10f)
[3] Closure modelling near the two-component limit 107
The 2-component-limit constraint is most conveniently handled in principal
axes of the Reynolds stresses, where u
i
u
j
= 0 if i ,= j. If u
2
2
is taken as the
vanishing component, then the other two normal stresses can be written as
u
2
1
= (1 +)k and u
2
3
= (1 )k. The 2-component limit requires that
X
l2
k2
U
k
x
l
= 0, (2.11)
and substituting the above values for the stresses into this equation leads to
a further four relations between the model coecients:
1
+
2
2
3
(
3
+
4
)
7
3
5
+
4
9
6
+
10
9
7
+
4
9
(
8
+
9
) +
11
9
10
+
2
3
(
11
+
12
)
8
27
(
13
+
14
)
34
27
15
2
9
(
16
+
17
)
4
9
(
18
+
19
)
14
9
20
= 0 (2.12a)
3
5
2
7
+ 2
10
+ 2
20
= 0 (2.12b)
3
10
6(
11
+
12
) 2
15
6(
16
+
17
) + 4(
18
+
19
) + 6
20
= 0 (2.12c)
2
20
= 0. (2.12d)
These equations can be solved, leaving four undetermined parameters:
1
=
8
15
8
=
1
5
+t 2p
15
= 3s
9
4
t
2
=
2
15
9
=
15
2
t 2p
16
= 2s +t
3
=
14
15
+
4
3
t
10
= p
17
= s
4
=
1
15
+
11
3
t
11
= 3t +p
18
=
15
4
t r
5
=
1
5
t
12
=
3
4
t
1
2
p
19
= r
6
=
1
10
+ 8t 2p
13
= 6s
27
4
t + 3r
20
= 0.
7
=
3
10
3
2
t +p
14
= 6s +
33
4
t 3r
However, it can be shown that the contributions to
ij2
arising from the
terms with coecients s and p are identically zero. The resulting model can
thus be written as
ij2
= 0.6 (P
ij
1
/
3
ij
P
kk
) + 0.3a
ij
P
kk
0.2
_
u
k
u
j
u
l
u
i
k
_
U
k
x
l
+
U
l
x
k
_
u
l
u
k
k
_
u
i
u
k
U
j
x
l
+u
j
u
k
U
i
x
l
__
c
2
[A
2
(P
ij
D
ij
) + 3a
mi
a
nj
(P
mn
D
mn
)]
+c
2
__
7
15
A
2
4
_
(P
ij
1
/
3
ij
P
kk
)
+0.1 [a
ij
1
/
2
(a
ik
a
kj
1
/
3
ij
A
2
)] P
kk
0.05a
ij
a
lk
P
kl
108 Craft and Launder
+0.1
__
u
i
u
m
k
P
mj
+
u
j
u
m
k
P
mi
_
2
/
3
ij
u
l
u
m
k
P
ml
_
+0.1
_
u
l
u
i
u
k
u
j
k
2
1
/
3
ij
u
l
u
m
u
k
u
m
k
2
_ _
6D
lk
+ 13k
_
U
l
x
k
+
U
k
x
l
__
+ 0.2
u
l
u
i
u
k
u
j
k
2
(D
lk
P
lk
)
_
,
where A
2
= a
ij
a
ij
. There are two free coecients, c
2
and c
2
, which can be set
by tuning the model to simple shear ows. Fu et al. (1987) recommended values
of c
2
= 0.6, c
2
= 0, which considerably simplies the task of implementing the
model in a computer code. Later, however, Fu (1988) concluded that slightly
better agreement for free shear ows could be obtained with c
2
= 0.55 and
c
2
= 0.6, values which greatly improved the performance in near-wall ows
since in many cases if one remains outside the viscosity-aected sublayer, no
wall corrections of the type described in [2] are then needed, Launder and Li
(1994).
2.1.2 Turbulence (or Slow) Part of Pressure-Strain
No-one has so far devised a successful analytical route for modelling
ij1
anal-
ogous to that for
ij2
. Thus the two-component limit is imposed empirically
through stress invariants of which, for a second rank tensor, there are two inde-
pendent parameters. One of these, A
2
has already appeared in the expression
for
ij2
. The natural second parameter might be thought to be
A
3
a
ij
a
jk
a
ki
. (2.13)
However, Lumley (1978) showed that, for modelling purposes, a combined
invariant A, dened as
A 1 9/8(A
2
A
3
), (2.14)
was a particularly powerful choice because, in the limit of two-component
turbulence, the parameter always goes to zero.
1
By including the parameter A
in a model for
ij1
, one may thus arrange that the model of
ij1
is consistent
with the two-component limit.
Thus, for
ij1
, a nonlinear extension of the return to isotropy model of Rotta
(1951) could be written as:
ij1
= c
1
a
ij
c
1
(a
ik
a
kj
1
/
3
A
2
ij
) c
1
A
2
a
ij
(2.15)
1
We can conveniently map the range of attainable states of the turbulent stress eld as
an A
2
-A
3
plot (Lumley 1978), Figure 3a. All realizable states fall within or on the boundary
of this triangle, the upper line corresponding to two-component turbulence while the two
curved lines represent axisymmetric turbulence (that is, where two of the normal stresses
are equal). The origin corresponds to isotropic turbulence. Alternatively, on an A
2
-A plot,
two-component turbulence corresponds with states lying on the A
2
axis, Figure 3b.
[3] Closure modelling near the two-component limit 109
(Although it might appear that the cubic term (a
ik
a
kl
a
lj
1
/
3
A
3
ij
) should
also be included in equation (2.15), the CayleyHamilton theorem means that
the term is proportional to A
2
a
ij
, and its inclusion would not thus add any
more generality to the form shown).
Figure 3: The anisotropy invariant map in A
2
-A
3
and A
2
-A space.
The approach followed is simply to make the coecients functions of the
invariants A and A
2
, to ensure that they vanish in the 2-component limit. The
UMIST group, for example, have employed the form
ij1
= c
1
_
a
ij
+c
1
(a
ij
a
jk
1
/
3
A
2
ij
)
A
a
ij
, (2.16)
where
c
1
= 3.1(A
2
A)
1/2
c
1
= 1.1 f
A
= A
1/2
.
2.2 Dissipation
Since the dissipative processes arise predominantly from the smallest scales
of turbulence,
ij
is normally considered to be essentially isotropic, even if
the stress eld is signicantly anisotropic. From this assumption,
ij
is often
modelled as
2
/
3
ij
.
However, local isotropy is not consistent with the 2-component limit which
requires
22
to vanish at a wall. A simple way of ensuring compliance with this
limit is to devise a model where
ij
(u
i
u
j
/k) close to a wall (or, indeed, in
other circumstances where the stress eld is near the two-component limit).
A transition function, based on the atness parameter A, can be employed
to switch between the two forms:
ij
=
2
/
3
ij
f
+
u
i
u
j
k
(1 f
). (2.17)
110 Craft and Launder
If f
2
k
+
x
k
__
lk
+c
u
k
u
l
_
x
l
_
, (2.18)
with coecients c
1
= 1.44, c
2
= 1.92. At UMIST, workers have retained this
general form, but have included some account of the eects of dierent stress
anisotropy on by allowing c
2
to be a function of A and A
2
, and reducing
the value of c
1
. The recommended form for the coecients is:
c
1
= 1.0 c
2
= 1.92/(1 + 0.7A
1/2
2
A). (2.19)
3 Applications to the computation of dynamic eld
The performance of the model described in Section 2 is now considered, rst
for free ows, then for ows near a single plane wall or free surface and then,
nally, for composite walls. To provide as accurate an impression as possible of
the capabilities of the TCL approach, we limit attention to computations made
with a single form of the model. Consequently, earlier TCL forms adopted by
Tselepidakis (1991) (see Launder and Tselepidakis 1993) and Launder and
Shima (1989) (see also Shima 1993, 1998) have not been included here even
though the latter, in particular, has been successfully applied to a wide range
of two- and three-dimensional boundary layers near walls. Nor do we include
the very recent publications by Leschziner and his group of the TCL model
applied to transonic and supersonic ows (Batten et al. 1999b,a).
3.1 Free Shear Flows
The free coecients in the model were originally assigned to secure satisfactory
agreement in various homogeneous shear ows and plane strains. Thus one
can hardly claim to be predicting these ows since they formed part of the
overall optimization process. A typical example is the homogeneous shear ow
considered in Figure 4. This particular example is a severe test as it starts
from isotropic turbulence. The growth of the anisotropy is predicted broadly
as the DNS of the four non-zero stress components indicates, although the
[3] Closure modelling near the two-component limit 111
Figure 4: Development of stress anisotropies in homogeneous shear ow. Lines:
TCL model predictions, Symbols: DNS of Matsumoto et al. (1991).
Flow Experimental value Basic model TCL model
Plane jet 0.110 0.100 0.110
Round jet 0.093 0.105 0.101
Plane wake 0.086 0.070 0.069
Table 1: Predicted and measured spreading rates of some self-preserving free
shear ows.
initial growth of the anisotropy of u
2
1
and u
2
2
is rather too weak. It is, however,
considerably better than that achieved by the Basic Model, presented in [2].
Turning to inhomogeneous shear ows, Table 1 reports the computed and
measured asymptotic growth rates of three free shear ows: the plane and
round jets and the plane wake. These three ows collectively provide a severe
test for any model. The table indicates that the TCL scheme again comes
close to mimicking the growth rates of all three, whereas the Basic Model
does badly for both the plane wake and the round jet. It is worth noting
that these results were obtained with a full elliptic solution of the transport
equations rather than the usual thin-shear-ow approximation. This practice
led to a reduction in growth rates (compared with a thin-shear-ow treatment)
of about 12% for the round jet and about 4% for the plane jet (El Baz et al.
1993). This dierence reected the rapid axial decay of the round jet. There
was negligible dierence between the two treatments for the wake. There are
similar improvements in the prediction of growth rates for buoyantly-driven
plumes, considered in [14].
112 Craft and Launder
Finally, Figure 5 compares the development of plane wakes created by two
dierent bodies, thus providing dierent initial states of turbulence energy
and dissipation, but with the same momentum decit. The experiments show
that the eects of the dierent initial conditions carry over very far down-
stream well beyond the region of measurement. The Basic Model, however,
quickly forgets about the dierent initial conditions, showing identical growth
for the two cases. In contrast, the TCL scheme clearly displays a very similar
development to that recorded.
Airfoil Expt.
Pred.
Solid Strip Expt.
Pred.
Airfoil Expt.
Pred.
Solid Strip Expt.
Pred.
Airfoil Expt.
Pred.
Solid Strip Expt.
Pred.
Airfoil Expt.
Pred.
Solid Strip Expt.
Pred.
0
600
450
300
150
0
0 500 1000 1500 2000
500 1000 1500 2000
0.0
0.1
0.2
0.3
0.4
0 500 1000 1500 2000
0.0
0.1
0.2
0.3
0.4
x/
x/
x/
u
2 c
/
(
U
2
)
c
u
2 c
/
(
U
2
)
c
(
y
1
/
2
/
)
2
(
y
1
/
2
/
)
2
600
450
300
150
0
0 500 1000 1500 2000
x/
(
y
1
/
2
/
)
2
around a square-
sectioned U-bend. Computations of Iacovides et al. (1996), : Basic model;
- : TCL model; Symbols: experiments of Chang et al. (1983). From Iacovides
et al. (1996).
4 Scalar ux modelling
Similar considerations relating to the two-component-limit behaviour can be
applied to the modelling of the scalar-ux transport equations. The exact
transport equations (see [2], equation (2.19)) can be written symbolically as
Du
i
Dt
= P
i
+
i
+d
i
i
, (4.1)
[3] Closure modelling near the two-component limit 117
where the production P
i
u
i
u
k
/x
k
u
k
U
i
/x
k
which does not re-
quire modelling, whilst
i
represents the pressure-scalar gradient correlation,
d
i
the diusion and
i
the dissipation rate of the scalar ux.
In this case, the assumption of isotropic dissipation leads to
i
= 0, and
consequently the main modelling task is to approximate
i
.
An analytical expression similar to equation (2.3) can be obtained for
i
:
i
=
1
4
_
V ol
3
u
k
u
r
l
r
k
r
i
dV
[r[
1
2
_
V ol
2
u
r
k
r
i
U
k
r
l
dV
[r[
, (4.2)
from which
i
is traditionally modelled as
i
=
i1
+
i2
, (4.3)
where
i1
represents the turbulence interactions and
i2
depends on the
mean strains. In buoyancy-aected ows there is a further contribution which
will be discussed in [14].
4.1 Mean-Strain (or Rapid) Part of Pressure-Scalar Gradient
Correlation:
i2
Again, if the mean strain is assumed to be essentially constant over the volume
of integration in equation (4.2), the
i2
process can be modelled as
i2
= 2b
l
ki
U
k
x
l
, (4.4)
where the tensor b
l
ki
represents the integral:
b
l
ki
=
1
4
_
V ol
2
u
r
k
r
i
dV
[r[
. (4.5)
Adopting the same approach to modelling this tensor as was done for
ij2
,
b
l
ki
can be modelled in terms of the Reynolds stresses and scalar uxes. How-
ever, the linearity principle (noting that the integral in equation (4.5) is linear
in the scalar ) requires that an expansion for b
l
ki
, whilst possibly being nonlin-
ear in the Reynolds stresses, should only depend linearly on the scalar uxes.
Including all possible terms which satisfy the required symmetry in i and k,
such an expansion up to cubic order can be written as
b
l
ki
=
1
u
l
ik
+
2
_
u
k
li
+u
i
lk
_
+
3
u
l
a
ik
+
4
_
u
k
a
li
+u
i
a
lk
_
+
5
u
m
a
ml
ik
+
6
u
m
(a
mk
li
+a
mi
lk
)
+
7
u
m
a
ml
a
ik
+
8
u
m
(a
mk
a
il
+a
mi
a
kl
)
+
9
u
l
a
mi
a
mk
+
10
a
ml
_
u
k
a
im
+u
i
a
km
_
+a
mn
a
mn
_
11
u
l
ik
+
12
_
u
i
lk
+u
k
li
__
+u
n
a
mn
_
13
a
ml
ik
+
14
(a
mk
li
+a
mi
lk
)
_
. (4.6)
118 Craft and Launder
Constraints similar to those applied in the modelling of
ij2
can now be ap-
plied to determine as many of the model coecients as possible. The equivalent
continuity and normalization conditions give:
Continuity: b
k
ki
= 0
Normalization: b
l
kk
= u
l
1
+ 4
2
= 0 (4.7a)
3
+
4
+
5
+ 4
6
= 0 (4.7b)
7
+
8
+
9
+
10
+
13
+ 4
14
= 0 (4.7c)
10
+
11
+ 4
12
= 0 (4.7d)
3
1
+ 2
2
= 1 (4.7e)
2
4
+ 3
5
+ 2
6
= 0 (4.7f)
2
8
+ 2
10
+ 3
13
+ 2
14
= 0 (4.7g)
9
+ 3
11
+ 2
12
= 0. (4.7h)
Before considering what 2-component limit constraint should be applied,
note that if only the linear terms are retained (so only
1
and
2
are non-
zero) the above equations yield
1
= 0.4,
2
= 0.1, leading to the linear QI
model of Launder (1973) (see also, Launder 1975; Lumley 1975):
i2
= 0.8u
k
U
i
x
k
0.2u
k
U
k
x
i
. (4.8)
Returning to the question of satisfying the 2-component limit, Shih and
Lumley (1985) applied a constraint that ensured the Schwarz inequality,
_
u
_
2
u
2
2
, (4.9)
could not be violated. They did this by imposing the condition that the rate
of change of the dierence
_
u
_
2
u
2
2
should be zero when equality held
or, mathematically,
2u
Du
Dt
= u
2
D
2
Dt
+
2
Du
2
Dt
, (4.10)
when
_
u
_
2
= u
2
2
.
However, this relation links the models for
ij2
and
i2
, and the outcome
was that not only did Shih and Lumley (1985) determine all the coecients
in b
l
ki
, but the above constraint also led to both free coecients in the TCL
model of
ij2
being determined as zero.
[3] Closure modelling near the two-component limit 119
Unfortunately, it was the c
2
and c
2
terms that enabled good agreement with
simple shear ow experiments. Without them, Shih and Lumley were forced to
add some additional, arbitrary, higher-order correction terms to their model in
order to get the correct stress levels in shear ow. There is, moreover, a further
objection to the Shih and Lumley formulation, in that, if a genuine passive
scalar is being considered, the thermal eld modelling should not inuence the
modelling of the underlying dynamic eld.
For these reasons, workers at UMIST have adopted an alternative approach,
by ensuring that the nett contribution to the u
i
transport equation,
2
+P
2
,
vanishes when u
2
2
is zero (Craft 1991). For
i2
, this condition translates to
U
k
x
l
b
l
k2
=
1
2
u
l
U
2
x
l
(4.11)
when u
2
= 0.
By again considering the situation in principal axes of the stresses, this
condition leads to the six equations
2
3
3
+
1
3
2
9
7
+
4
9
9
+
2
3
11
+
1
9
13
=
1
2
(4.12a)
2
3
7
+
2
3
13
= 0 (4.12b)
2
11
+
13
= 0 (4.12c)
2
3
4
+
1
3
2
9
8
+
4
9
10
+
2
3
12
+
1
9
14
= 0 (4.12d)
2
3
8
+
2
3
14
= 0 (4.12e)
2
12
+
14
= 0. (4.12f)
Solving these, together with the earlier continuity and normalization equa-
tions, leads to the result
1
= 0.4
8
= 1/8 1/2
7
2
= 0.1
9
= 1/8 +
7
3
= 1/6
10
= 1/2
7
4
= 1/6
11
= 1/20 1/2
7
5
= 1/15
12
= 1/80 + 1/4
7
6
= 1/15
13
= 1/10 +
7
14
= 1/40 1/2
7
.
with, apparently, one free coecient. However, the term multiplied by
7
can
be shown to be identically zero, and hence the resulting model for
i2
can be
120 Craft and Launder
written:
i2
= 0.8u
k
U
i
x
k
0.2u
k
U
k
x
i
+
1
/
3
k
u
i
0.4u
k
a
il
_
U
k
x
l
+
U
l
x
k
_
+0.1u
k
a
ik
a
ml
_
U
m
x
l
+
U
l
x
m
_
0.1u
k
(a
im
P
mk
+ 2a
mk
P
im
) /k
+0.15a
ml
_
U
k
x
l
+
U
l
x
k
_
_
a
mk
u
i
a
mi
u
k
_
0.05a
ml
_
7a
mk
_
u
i
U
k
x
l
+u
k
U
i
x
l
_
u
k
_
a
ml
U
i
x
k
+a
mk
U
i
x
l
__
, (4.13)
where there are no free coecients.
4.2 Turbulence (or Slow) Part of Pressure-Scalar Gradient
Correlation:
i1
To model
i1
in a similar manner to
ij1
, the obvious extension to the linear
model of Monin (1965), would be to employ an expression of the form
i1
= c
1
k
u
i
c
k
a
ij
u
j
c
k
a
ik
a
kj
u
j
c
k
A
2
u
i
. (4.14)
Considering this expression in principal axes of the stresses, it is clear that such
a model satises the condition that
21
should vanish when u
2
2
= 0 regardless
of the values of the model coecients. The UMIST group has therefore em-
ployed a form similar to this, allowing the coecients to be functions of the
stress invariants, and tuning them to a range of free shear ows.
The exact expression for
i1
does not depend explicitly on mean scalar
gradients, and these have thus not traditionally appeared in the modelled pro-
cess. However, Jones and Musonge (1983) argued that the uctuating quanti-
ties do, nevertheless, depend on mean gradients, and thus included a term in
their model for
i
which did explicitly contain the mean scalar gradient. Craft
(1991) also found it benecial to include some explicit mean scalar gradient
dependence in order to capture simple homogeneous shear ows at dierent
strain rates. The form employed for
i1
in this latter work was
i1
= c
1
r
1/2
k
_
u
i
(1 +c
1
A
2
) +c
1
a
ik
u
k
+c
1
a
ik
a
kj
u
j
1
rka
ij
x
j
,
(4.15)
[3] Closure modelling near the two-component limit 121
where
c
1
= 1.7
_
1 + 1.2(A
2
A)
1/2
1
= 0.8, c
1
= 1.1,
c
1
= 0.6, c
1
= 0.2A
1/2
,
and the timescale ratio r is dened
2
as r = (2
/
2
)(k/). In this case the factor
A
1/2
in the coecient c
1
ensures that this part of the model also satises the
2-component limit.
The parameter r represents the ratio of mechanical to thermal timescales,
where 2
. A number of such
equations have been proposed (see, for example Newman et al. 1981; Jones
and Musonge 1983; Shih et al. 1985; Craft and Launder 1989; Nagano et al.
1991) although it must be conceded that few of these have been applied over
a very wide range of ows, and there is thus relatively little agreement on the
exact form that such an equation should take. In the examples below, in order
to focus attention on the modelling of the scalar uxes, the timescale r has
either been prescribed (from available data) of obtained from the correlation
of equation (4.16). A comprehensive account of recent approaches to modelling
the
2
p = 2
U
j
x
i
u
i
x
j
u
j
x
i
u
i
x
j
+
u
j
x
i
u
i
x
j
(1.1)
(we are considering constant density ow 1); the boundary condition
is usually taken to be p/x
n
= 0, ignoring a small viscous contribution.
The boundary condition inuences the pressure of the interior uid through
the solution to (1.1). Mathematically this is quite simple: the solution to the
linear equation (1.1) consists of a particular part, forced by the right-hand
side, and a homogeneous part, forced by the boundary condition. The fact
that the boundary condition adds to the solution interior to the uid can be
described as a non-local, kinematic eect.
Figure 1 schematizes non-locality in the Poisson equation as a reected
pressure wave, but for incompressible turbulent uctuations the wall eect is
instantaneous, though non-local. Pressure reection enhances pressure uc-
tuations; indeed, Manceau et al. (2001) show that pressure reection can in-
crease redistribution of Reynolds stress anisotropy. Redistribution is due to the
pressure-strain correlation: the notion that it is increased by the wall eect
is contrary to most second moment closure (SMC) models, which represent
pressure echo as a reduction of the redistribution term.
The idea of associating inviscid wall eects with pressure reection is natu-
ral, because the pressure enters the Reynolds stress transport equation through
the velocity-pressure gradient correlation. Suppression of the normal compo-
nent of pressure gradient by the wall should have an eect on the rate of
redistribution of variance between those components of the Reynolds stress
tensor that contain the normal velocity component i.e., u
n
u
i
, where n de-
notes the wall-normal direction. This eect enters the evolution equation for
the Reynolds stress (equation (1.3) below).
127
128 Durbin and Pettersson-Reif
IMAGE VORTICITY
PRESSURE REFLECTION
Figure 1: Schematic representations of non-local wall inuences.
However, there is another notion about how anisotropy of the Reynolds
stress tensor is altered non-locally by the presence of a wall. The inviscid
boundary condition on the normal component of velocity is the no-ux con-
dition u n = 0. This constraint on the normal velocity produces another
non-local, elliptic inuence of the boundary. If the vorticity,
k
, is known the
velocity solves the kinematic equation
2
u
i
=
ijk
k
x
j
. (1.2)
The boundary condition u n = 0 alters the ow interior to the uid. At
a plane boundary this is termed the image vorticity eect (gure 1); for
instance, (1.2) can be solved by extending the tangential vorticity components
anti-symmetrically inside the boundary, and extending the normal component
symmetrically into the boundary. Again, the wall alters the ow interior to
the uid through non-local kinematics. This perspective on ellipticity is often
referred to as kinematic blocking (Phillips 1955, Hunt and Graham 1978).
Kinematic blocking is not an alteration of the redistribution tensor. It can
be perceived as a continuity eect: instead of (1.2), suppose that a homoge-
neous eld of turbulence u
, where is a
velocity potential. Incompressibility, u = 0 implies that
2
= 0. The
boundary condition is n = u
ij
=
h
ij
+
w
ij
(1.4)
where
h
ij
is a homogeneous pressure-strain model, such as the simple IP model,
h
ij
= c
1
a
ij
c
2
_
P
ij
1
3
P
kk
ij
_
. (1.5)
Here a
ij
u
i
u
j
/k
2
3
ij
is the anisotropy tensor. The term
w
ij
in (1.4) is
the additive wall correction. In addition to wall distance, d, it is a function of
the wall normal n. The latter must be used interior to the uid, where it is ill
dened. (For example, if it is dened as the normal vector at the nearest wall
location then it is discontinuous on a surface emanating from corners. While
one might dene a wall normal function to be the gradient of a smooth wall
distance function, that has not been used in the literature.)
For concreteness, formulas that have been used in conjunction with the IP
model (Gibson and Launder 1978), can be cited. Their basic type is
w
ij
= c
w
1
k
_
u
m
u
l
n
m
n
l
ij
3
2
u
i
u
m
n
m
n
j
3
2
u
j
u
m
n
m
n
i
_
L
d
+ (1.6)
130 Durbin and Pettersson-Reif
Here indicates that further terms are contained in the formulas that have
been used in the literature. Expression (1.6) illustrates that wall corrections are
tensorial operators that act on the Reynolds stress tensor. The n
i
dependence
of these operators has to be adjusted to properly damp each component of
u
i
u
j
(Craft et al. 1993). The formula for the correction function,
w
ij
, has to
be readjusted in this manner for each homogeneous redistribution model to
which it is applied (Lai and So 1990).
Elliptic relaxation (Durbin 1993) is a rather dierent approach to wall ef-
fects. Instead of adding a wall echo term, the homogeneous redistribution
model is used as the source in a non-homogeneous, modied Helmholtz equa-
tion
1
equation:
L
2
2
f
ij
f
ij
=
h
ij
k
. (1.7)
Here f
ij
is an intermediate variable, related to the redistribution tensor by
ij
= kf
ij
. The turbulent kinetic energy, k, is used as a factor in order to
enforce the correct behavior,
ij
0, at a no-slip boundary. The anisotropic
inuence of the wall on Reynolds stresses interior to the uid arises by imposing
suitable boundary conditions on the components of the u
i
u
j
f
ij
system. For
instance, at a no-slip surface the normal intensity u
n
u
n
= O(d
4
) and tangential
intensity u
t
u
t
= O(d
2
) can be given their correct asymptotic limits. The wall
normal now enters only into the wall boundary condition.
The precise form of elliptic relaxation models can be found in Durbin (1993),
Wizman et al. (1996) and Pettersson and Andersson (1997). In these references
the general formulation is applied to several particular
h
ij
models: IP, LRR
(Launder et al. 1975), SSG (Speziale et al. 1991), FLT (Fu et al. 1987) and
RLA (Ristorcelli et al. 1995). While the wall echo approach (1.6) requires a
completely new formulation for each model, the elliptic relaxation equation
(1.7) is unchanged, except for the source on the right side. Implementation of
a new model into a computer code can be done entirely in a subroutine that
denes
h
ij
.
2 A justication
The elliptic relaxation formulation can be justied by a modication to the
usual rationale for pressure-strain modeling. The analysis in this section is in
the vein of formulating a template for elliptic relaxation, rather than being a
derivation per se.
For simplicity write the Poisson equation for the pressure uctuation (1.1)
as
2
p = S(x). (2.1)
1
The modied Helmholtz equation is
2
k
2
= 0; the unmodied equation has a +
sign.
[4] The elliptic relaxation method 131
The zero normal derivative condition on p can be imposed on a plane bound-
ary by extending the source, S(x), symmetrically into the surface. Let that
extension be understood. Then a formal solution to the Poisson equation is
obtained by inverting (2.1) with its free-space Green function:
p(x) =
1
4
___
S(x
)
[x x
[
d
3
x
. (2.2)
The redistribution term in the exact, unclosed second moment closure equa-
tions includes the velocity-pressure gradient correlation. Dierentiating (2.2)
with respect to x
j
, then integrating by parts, give
u
i
p
x
j
(x) =
1
4
___
u
i
(x)
S(x
)
x
j
[x x
[
d
3
x
. (2.3)
after multiplying by u
i
and averaging. The integrand of (2.3) contains a two-
point correlation. This embodies one origin of the lack of closure of the single-
point moment equations: single-point moment equations depend on two-point
statistics.
If the turbulence is homogeneous, the non-local closure problem is masked
by translational invariance of two-point statistics. Analyses for homogeneous
turbulence (Chou 1945) note at this point in the derivation that u
i
(x)
S(x
)
x
j
is a function of x x
)
[x x
[
d
3
x
=
h
ij
(2.4)
where
h
ij
is not a function of x.
The original form of the source in (1.1) motivates the standard practice of
splitting
h
ij
into slow and rapid parts,
h
ij
= N
ij
+M
ijkl
U
l
x
k
. (2.5)
The second (rapid) term is motivated by the rst term in (1.1); the rst
(slow) term is motivated by the second, nonlinear term in (1.1). The split
(2.5) is invoked in the homogeneous redistribution models that are used on
the right side of (1.7); for instance the rst term of (1.5) is the slow part, with
N
ij
= c
1
a
ij
, and the second is the rapid part, with M
ijkl
= c
2
(u
i
u
k
jl
+
u
j
u
k
il
2
3
u
k
u
l
ij
).
If the turbulence is not homogeneous (2.4) is not applicable and the role of
non-homogeneity in (2.3) must be examined (Manceau et al. 2001). This re-
quires a representation of the spatial correlation function in the integrand. To
132 Durbin and Pettersson-Reif
this end, an exponential function will be used as a device to introduce the cor-
relation length of the turbulence into the formulation. Letting u
i
(x)
S(x
)
x
j
=
R
ij
(x
)e
|xx
|/L
in (2.3) gives
u
i
p
x
j
(x) =
___
R
ij
(x
)
e
|xx
|/L
4[x x
[
d
3
x
. (2.6)
The representation of non-locality in this formula can be described as both
geometrical and statistical.
_
() e
|xx
|/L
. .
statistical
decorrelation
_
4[x x
[
. .
geometrical
spreading
d
3
x
.
The kernel in the above equation is the Greens function for the modied
Helmholtz equation (1.7); the large dot stands for the source, R
ij
in the present
case. In other words, if L is constant, then (2.6) is the solution to
2
u
i
p
x
j
1
L
2
u
i
p
x
j
= R
ij
. (2.7)
Of course, the specic equation (1.7) could be obtained by dividing (2.3) by
k(x) and letting f
ij
= u
i
p/x
j
/k and R
ij
=
h
ij
/kL
2
. However, the present
rationalization of (2.7) is meant only to suggest a template for (1.7). Ultimately
the motivation for the elliptic relaxation method is to enable boundary con-
ditions and anisotropic wall eects to be introduced into the second-moment
closure model in a exible and geometry-independent manner.
The elliptic relaxation procedure accepts a homogeneous redistribution
model on its right side and operates on it with a Helmholtz type of Greens
function, imposing suitable wall conditions. The net result can be a substan-
tial alteration of the near-wall behavior from that of the original redistribution
model. Figure 2 illustrates how elliptic relaxation modies the SSG (Speziale
et al. 1991) homogeneous redistribution model. The dashed lines in this gure
are the homogeneous SSG model for
h
ij
, evaluated in plane channel ow with
friction velocity Reynolds number of R
12
predicted by elliptic relaxation agrees quite well with the data. Equation
(1.7) is linear, so its general solution can be written as a particular part, forced
by the source, plus a homogeneous part that satises the boundary condition.
The particular part would tend to have the same sign as the source and could
[4] The elliptic relaxation method 133
y
+
u
v
c
o
m
p
o
n
e
n
t
0 100 200
-0.3
-0.2
-0.1
0
0.1
12
via elliptic relaxation
h
12
SSG formula
DNS data
12
y
+
v
2
c
o
m
p
o
n
e
n
t
0 100 200
0
0.05
0.10
0.15
0.20
y
+
u
2
c
o
m
p
o
n
e
n
t
0 100 200
-0.3
-0.2
-0.1
0
Figure 2: The eect of elliptic relaxation on the SSG formula for
h
ij
. Compu-
tation of plane channel ow with R
= 395.
not cause the sign reversal shown by the solution for uv. So it is the homoge-
neous part of the general solution that causes the sign of
12
to be opposite
to
h
12
near the wall, and brings
12
into agreement with the DNS data.
A similar reduction in magnitude, and drastic improvement in agreement
with the data, is seen also for
11
and
22
. Near the wall, the homogeneous
model shows too large transfer out of u
2
and into v
2
(and w
2
). The elliptic
relaxation procedure greatly improves agreement with the data. This includes
creation of a slightly negative lobe near the wall in the
22
prole, correspond-
ing with the data. The negative lobe is required if v
2
is to be non-negative
(see (3.5)).
It is rather intriguing that the elliptic relaxation equation is able to auto-
matically produce such improvement to the redistribution model. How this
occurs is not well understood mathematically.
3 Its use with Reynolds Stress Transport equations
In practice one solves (1.3) and (1.7) with a model for the transport term T
ij
to complete the closure. The Daly and Harlow (1970) formula
T
ij
u
k
u
i
u
j
x
k
=
x
k
_
T
kl
x
l
u
i
u
j
_
(3.1)
134 Durbin and Pettersson-Reif
is commonly used. Here
T
ij
is a turbulent eddy viscosity tensor that is given
by
T
ij
= c
u
i
u
j
T (3.2)
and T is the turbulence time-scale
T = max
_
k/, 6
_
/
_
. (3.3)
This denition of the time-scale simply uses the Kolmogoro scale,
_
/, as
a lower bound, applicable near the wall, and the integral scale k/ outside the
viscous wall layer. It is not meant to be valid in truly low Reynolds number
turbulence (such as the nal period of decay), to which Kolmogoro scaling
does not apply.
The complete set of closed Reynolds stress transport and redistribution
equations that have been used in most elliptic relaxation models to date are
Du
i
u
j
Dt
+
l
(
ikl
u
k
u
j
+
jkl
u
k
u
i
) +
u
i
u
j
k
= P
ij
+
ij
+
x
k
_
T
kl
x
l
u
i
u
j
_
+
2
u
i
u
j
L
2
2
f
ij
f
ij
=
h
ij
k
a
ij
T
(3.4)
where
P
ij
= u
i
u
k
U
j
x
k
u
j
u
k
U
i
x
k
l
(
ikl
u
k
u
j
+
jkl
u
k
u
i
)
is the production tensor and
ij
= kf
ij
. The coordinate system rotation vector
l
has been included; the equations are written in a rotating frame of reference
for use in 4. For non-rotating frames
l
= 0.
It can be seen that one modication has been made to the formulation
described in 1:
ij
has been eliminated from (1.3) by adding u
i
u
j
/k to the
left side of the transport equation and a
ij
/T to the right side of the elliptic
relaxation equation. If the turbulence is homogeneous, then the solution to
the second of equations (3.4) is
ij
=
h
ij
+
_
u
i
u
j
k
2
3
ij
_
.
When this is inserted into the rst of (3.4), u
i
u
j
/k cancels from both sides. So,
the aggregate eect of this last modication is to replace
ij
by the isotropic
formula
2
3
ij
in the quasi-homogeneous limit. This is the behavior far from
walls. Near to surfaces, anisotropic dissipation is lumped with the redistribu-
tion model:
ij
ij
+
ij
u
i
u
j
/k. Equation (3.4) can be understood to
model this term in aggregate.
The mathematical motive for inserting u
i
u
j
/k on the left side of (3.4) is to
ensure that all components of u
i
u
j
go to zero at least as fast as k as the wall
[4] The elliptic relaxation method 135
is approached (Durbin 1993). In particular, the tangential components of u
i
u
j
are O(d
2
) as d 0. The normal component becomes O(d
4
) if the boundary
condition
f
nn
= 5 lim
d0
_
u
n
u
n
k
2
_
(3.5)
is imposed on d = 0. This condition is derived by noting that as d 0 the
dominant balance for the normal stress in the rst of (3.4) becomes
u
n
u
n
k
=
nn
+
2
u
n
u
n
x
2
k
.
Assuming u
n
u
n
(d
4
) gives
u
n
u
n
k
12
u
n
u
n
d
2
=
nn
= kf
nn
.
But k d
2
/2, so the left side is 5u
n
u
n
/k, giving (3.5). [The asymptote
k d
2
/2 follows from the limiting behavior of the k-equation
=
2
k/x
2
d
upon integrating twice with the no-slip condition k = k/x
d
= 0 on d = 0.]
Boundary conditions on slip surfaces can be derived in a similar manner,
although there has been little work on that subject.
A similar analysis shows that the boundary condition on the o-diagonal
f
nt
must also be (3.5). It is not possible to impose the condition u
n
u
t
=
O(d
3
). However, this asymptotic behavior is only valid as d
+
0, where
molecular transport dominates over turbulent mixing. So, the precise power is
less important than the condition u
n
u
t
(U
t
/x
n
) as d
+
0, that is met
by the model.
The tangential components f
t
1
t
1
and f
t
2
t
2
are only required to be O(1) as
d 0. This is because the dominant balance
u
t
u
t
k
=
tt
+
2
u
t
u
t
x
2
d
+O(d
2
)
causes the viscous and dissipation terms to cancel: with u
t
u
t
(d
2
) and k/ =
d
2
/2 this equation gives
tt
= O(d
2
). Guaranteeing the d
2
behavior was
the motivation for writing
ij
= kf
ij
. As long as f
tt
= O(1) as d 0 the
correct tangential balance will be achieved. Demuren and Wilson (1995) use
the condition f
t
1
t
1
= f
t
2
t
2
= 1/2f
nn
to ensure that
ij
is trace-free. In two-
dimensional ows Durbin (1993) used f
tt
= 0 to obtain the Reynolds stresses
in the xy plane, and computed the third normal stress from w
2
= 2ku
2
v
2
.
The Demuren and Wilson (1995) condition is probably more satisfactory in
general (see also Durbin (1991), Appendix B).
136 Durbin and Pettersson-Reif
The length scale in (3.4) is prescribed by analogy to (3.3) as
L = max
_
_
_
c
L
k
3/2
, c
_
1/4
_
_
_
. (3.6)
Although most implementations of elliptic relaxation to date have used these
simple formulas for L and T, they are not crucial to the approach. The only
important feature is that L and T do not vanish at no-slip surfaces. If they
vanished then the equations would become singular. But it would also be
unphysical for the turbulence scales to vanish at a wall: they represent the
correlation length and time of turbulence, not the intensity, which are not zero.
In fully turbulent ow it has been found from direct numerical simulations that
the Kolmogoro scaling collapses near-wall data quite eectively.
3.1 Variants
The elliptic relaxation approach can be invoked in a variety of manners, as is
already implied by (3.4). In principle, there is freedom to choose the source
term
h
ij
in the elliptic equation; the only constraint is that the model relax
to
ij
ij
=
h
ij
2
3
ij
in the quasi-homogeneous limit. However, to date
the primary variations of this ilk have been to substitute dierent existing
closuressuch as IP or SSGfor the source term.
Variants of the elliptic operator have been explored by Wizman et al. (1996)
and by Dreeben and Pope (1997). Wizman et al. (1996) note that in the
constant stress, or logarithmic, layer the source term in (1.7) is proportional
to 1/y and hence f
ij
is as well. This means the the Laplacian of f
ij
does
not vanish. If f
ij
is redened via
ij
= kf
ij
/L then the source term becomes
L
h
ij
/k, making it and f
ij
constant in the log-layer. This makes the Laplacian
vanish. Wizman et al. (1996) entitle this the neutral formulation. They also
consider replacing the Laplacian by (L
2
f) = L
2
2
f +2L(L)f. These
modications were explored in the interest of improving predictions in the
central region of channel ow. Research into such variants of the formulation
is currently in progress (Manceau et al. 2001).
Dreeben and Pope (1997) invoked the representation
ij
= u
i
u
k
f
kj
+u
j
u
k
f
ki
2
3
ij
u
l
u
k
f
kl
in order to make elliptic relaxation compatible with Langevin stochastic mod-
els. This formulation automatically satises the redistribution property
ii
=
0. However, f
ij
is then no longer symmetric. Although attractive in concept,
this variant adds greatly to the computational complexity of the model.
Elliptic relaxation has been simplied into a scalar eddy viscosity formula-
tion by the v
2
-f model (Durbin 1995, Parneix et al. 1998a). To this end the
[4] The elliptic relaxation method 137
set of equations for the Reynolds stress tensor is replaced by a pair of equa-
tions for a velocity scalar v
2
and a function f that is somewhat analogous to
redistribution. The governing equations are
Dv
2
Dt
+
v
2
k
= kf +
x
k
_
T
v
2
x
k
_
+
2
v
2
L
2
2
f f = c
2
P
k
+c
1
_
v
2
k
2
3
_
/T
(3.7)
where P =
T
_
U
j
x
i
+
U
i
x
j
_
U
j
x
i
. These are solved in conjunction with the k-
equations, which are needed to obtain the length and time scales. The bound-
ary condition on f is that of equation (3.5), for the normal component of
Reynolds stress:
f = 5 lim
d0
_
v
2
k
2
_
= 20 lim
d0
_
2
v
2
d
4
_
(3.8)
for a wall located at d = 0. The eddy viscosity is predicted by the formula
T
= c
v
2
T. The motivation behind (3.7) and (3.8) is to represent the tendency
of the wall to suppress transport in the normal direction. The variable v
2
is a
scalar, not the normal component of a tensor, but the boundary condition on
f makes it behave like u
n
u
n
near to solid walls.
The v
2
-f variant is motivated by the need for a practical prediction method.
In this model the mean ow is computed with an eddy viscosity. The transport
equations for k, , and v
2
are solved to obtain the spatial distribution of eddy
viscosity.
Elliptic relaxation is not a panacea, but it has intriguing properties. Other
avenues to geometry-independent near-wall modeling are treated elsewhere
in this book. In particular, tensorally nonlinear representations for the two-
component limit (Launder and Li 1994) are discussed in [3] and [14].
4 Applications
4.1 Insensitivity to the homogeneous model
Figure 2 illustrated that near to walls the elliptic relaxation closure overwhelms
the homogeneous redistribution model. There is an extent to which this makes
the prediction of surface properties insensitive to the detailed homogeneous
redistribution model. Figure 3 shows the friction coecient in ow over a
backward facing step using both the IP and SSG models. The two models
give very similar solutions for C
f
. The same is true of the entire mean ow
eld. Although this is not always the case, it is clear that in this particular
calculation elliptic relaxation largely determines the solution. The specics
138 Durbin and Pettersson-Reif
x/H
1
0
3
C
f
-5 0 5 10 15 20 25
-2
0
2
4
6
IP
SSG
DNS data
Figure 3: Friction coecient in ow over a backward facing step. The step is
at x = 0. Model solutions compared to DNS data from Le et al. (1997).
of the homogeneous redistribution model are less important, although they
are not irrelevant. In other ows they play a role and the complexity of the
homogeneous model also has an impact on numerical tractability
An under-prediction of the minimum C
f
is seen in gure 3 near x/H = 4.
This is a common feature of predictions of most second-moment transport
models in ow over a backward step, irrespective of the near wall treatment
(Parneix et al. 1998b).
As a second example, consider the ow over a convex bump. This geometry
is characterized by substantial favorable and adverse pressure gradients. While
that might seem at rst to make it a stringent test, in fact the experimental
data of Webster et al. (1996) are predicted well by many models. Velocity pro-
les computed with the IP model via elliptic relaxation are shown on the left
side of gure 4. Predictions by the SSG model are virtually indistinguishable
from those portrayed.
The closure model predictions are in excellent accord with the data. Al-
though the data do not extend to the upper wall, the predicted boundary layer
thickness must be correct; otherwise the central velocity would be wrong, due
to the need to maintain a constant mass ux.
The friction coecient, C
f
(x), and pressure coecient, C
p
(x), predictions by
either SSG or IP are also in good agreement with data; again, only predictions
by the IP are shown in gure 4. In this case, as well as for the backstep, elliptic
relaxation makes the results relatively insensitive to the homogeneous SMC
model.
Transport of mean momentum towards a solid boundary is controlled largely
by a thin layer immediately next to the wall. Success in computing wall-
bounded ows therefore depends strongly on the model used to account for
the non-local wall eects that occur in the proximity of solid boundaries. As
[4] The elliptic relaxation method 139
-0.5 0 0.5 1 1.5 2 2.5
U profiles
0
0.1
0.2
0.3
0.4
0.5
Y
x
-0.5 0 0.5 1.0 1.5 2.0
-0.75
-0.50
-0.25
0
0.25
0.50
0.75
C
p
100C
f
Figure 4: Velocity proles, skin friction and surface pressure coecients in ow
over a bump. Solutions to the IP model with elliptic relaxation are compared
to experimental data from Webster et al. (1996).
elucidated in 3, the boundary conditions for the elliptic relaxation equation
(3.4) assures a proper suppression of the normal component of turbulent trans-
port in the vicinity of the surface. This has a large inuence on skin friction
predictions in the two examples considered so far. In strongly curved or ro-
tating ows the near-wall modeling inuences relaminarization of the viscous
wall layer.
The majority of turbulent ows of engineering interest are characterized
by nonequilibrium near-wall turbulence as well as by eects of inertial forces
arising from streamline curvature or system rotation. The most natural level
of closure modeling to adopt in these cases is full Reynolds stress transport
models, which in a natural and systematic way account for rotational eects.
The next sections contain examples to illustrate ows in which rotation and
curvature are important.
4.2 Rotating cylinder
Pettersson et al. (1996) considered the turbulent boundary layer around an
innitely long, axially rotating cylinder in a quiescent uid. The cylinder ro-
tates with a constant angular velocity as dened in gure 5. The mean ow
equation in this case is
u
r
u
=
R
2
w
r
2
+r
U
r
with U
> 0 if
w
> 0. The second moment closure model is required to predict
u
r
u
.
Figure 5 compares predictions by the IP model in conjunction with both
elliptic relaxation and with the LaunderShima low-Reynolds number, sec-
ond moment closure (Launder and Shima 1989). The LaunderShima model
140 Durbin and Pettersson-Reif
R
(r)
r
Schematic of ow conguration
0.0 0.2 0.4
0.0
1.0
U
/
U
w
IP
Launder & Shima
Exp
(r-R)/R
Mean velocity distribution
10
-3
10
-2
10
-1
0.0
0.1
0.2
0.3
0.4
IP
Launder & Shima
Exp
(r-R)/R
uv/2k
Structural parameter
Figure 5: Rotating cylinder in a quiescent uid Re R
2
/ = 20,000. Exper-
imental data of Andersson et al. (1991).
utilizes IP with the additive wall-echo methodology described prior to equa-
tion (1.6). Dierences between the model predictions therefore can be at-
tributed mainly to the method of near-wall modeling.
In the limit as the solid boundary is approached, the structural parame-
ter uv/k should tend to 0 due to kinematic blocking. However, the additive
wall-correction in LaunderShima does not provide sucient suppression of
turbulent normal and shear stresses; the structural parameter predictions in
gure 5 behave incorrectly near the surface. The mean ow prediction in the
upper part of gure 5 also is incorrect. When the same IP closure is used in
conjunction with elliptic relaxation the predictions improve dramatically. Both
the mean ow and the structure parameter are brought into good agreement
[4] The elliptic relaxation method 141
y
x
z
(b) (a)
wall U
Figure 6: Schematic of plane channel ow in orthogonal mode rotation. (a)
Poiseuille ow (b) Plane Couette ow.
with the experimental data. The importance of modeling suppression of the
wall-normal intensity is made clear by this example.
4.3 Rotating Channel Flows
Non-inertial frames of reference are encountered in a wide variety of engi-
neering ows. When the momentum equations are transformed to a rotating
frame, a Coriolis acceleration, 2 u, is added. One half of the Coriolis ac-
celeration comes from transforming the time-derivative, the other half comes
from rotation of the velocity components relative to an absolute frame. When
the Reynolds stress transport equations (3.4) are similarly transformed, the
frame rotation adds the same term,
l
(
ikl
u
k
u
j
+
jkl
u
k
u
i
), to both the time
derivative and to the production tensors, by this same reasoning. It is im-
portant to distinguish these two contributions because the production tensor,
P
ij
, often appears in closure models; the IP formula (1.5) is a case in point.
Only the contribution of rotation to the production tensor should be added
to the closure formula. If this is not done correctly the equations will not be
coordinate frame independent.
The Coriolis acceleration profoundly aects turbulent ow. Depending on
magnitude and orientation of the rotation vector, , relative the mean ow
vorticity, = U, turbulence can be augmented or reduced: the turbulence
is suppressed if the imposed rotation is cyclonic (that is, if the mean ow
vorticity is parallel to the rotation vector); the turbulence is enhanced if the
imposed rotation is anticyclonic.
An imposed system rotation may also contribute to the formation of orga-
nized large scale structures. These roll-cells are found in turbulent ows sub-
jected to anticyclonic rotation (Andersson 1997). The presence of a rotation-
induced secondary mean ow inevitably alters the turbulence eld as well.
142 Durbin and Pettersson-Reif
DNS
RLA
y/h
0.0
0.0
0.0
0.0
1.0
-1.0 0.0 1.0
Ro = 0.05
Ro = 0.10
Ro = 0.20
Ro = 0.50
U/U
b
Mean velocity distribution
DNS
RLA
y/h
0.0
0.0
0.0
0.0
1.0
-1.0 0.0 1.0
Ro = 0.05
Ro = 0.10
Ro = 0.20
Ro = 0.50
-uv/u
*
2
0
Turbulent shear stress distribution
DNS
RLA
y/h
0.0
0.0
0.0
0.0
0.5
-1.0 0.0 1.0
Ro = 0.05
Ro = 0.10
Ro = 0.20
Ro = 0.50
-uv/k
Structural parameter
Figure 7: Spanwise rotating Poiseuille ow at Re
h/ = 194; Ro
2h/U
b
. DNS: Kristoersen and Andersson (1993).
The mean ow eld can be directly altered by the imposed rotation, or it can
respond indirectly in consequence of alterations to the turbulence. Test cases
where the latter dominates are attractive because the response of the closure
model to system rotation is then all important. Fully developed turbulent ow
between two innite parallel planes in orthogonal mode rotation constitutes
one such example. Experimental (Johnston et al. 1972) and DNS data bases
(Kristoersen and Andersson 1993; Lamballais et al. 1996) are available to
assist the model development. These data have contributed to making this
particular ow a standard benchmark test case.
The most frequently adopted conguration is pressure-driven (Poiseuille)
ow subjected to spanwise rotation (case a of gure 6). In the absence of
[4] The elliptic relaxation method 143
rotationand in laminar ow, even with rotationthe mean velocity prole
is symmetric about the middle of the channel. Imposed rotation breaks the
mean ow symmetry. The mechanism is indirect; rotation alters the turbulence
eld, producing asymmetry in the Reynolds shear stress. Because the mean
ow vorticity changes sign across the channel, the ow eld is simultane-
ously subjected to both cyclonic and anticyclonic rotation. The side on which
the turbulence is suppressed (enhanced) is usually referred to as the stable
(unstable) side of the channel. In gure 6 the vorticity in the upper part of the
channel is counter-rotating with the frame; this is destabilizing. The increase
in turbulent mixing on that side steepens the velocity prole. Asymmetry de-
velops as shown in the gure; it can be seen more clearly in gure 7. Note that
y is increasing downward in gure 6. The proles in the upper left of gure 7
are consistent with gure 6 if the former are rotated clockwise by 90
.
At high cyclonic rotation rates, the ow eld tends to relaminarize. Near-
wall modeling then plays a crucial role. The commonly used wall-function
approach assumes fully developed turbulence and must be abandoned in this
case.
Figure 7 displays model predictions by Pettersson and Andersson (1997) of
unidirectional, fully developed rotating Poiseuille ow, U = [U(y), 0, 0] and
= [0, 0, (y)]. The transport of mean momentum is governed by
0 =
P
x
+
y
_
U
y
uv
_
P
y
= 2U
v
2
y
(4.1)
where P
= P
1
2
2
(x
2
+y
2
) is the mean reduced pressure. The Coriolis ac-
celeration does not directly aect the mean ow: it is balanced by the pressure
and turbulent normal stress in the y-direction, as stated in the y-momentum
equation. In the x-momentum equation P
z
Spanwise averaged mean streamwise velocity
z
y
U(2h)= U
w
U(0)= 0
Figure 8: Spanwise rotating plane Couette ow at Re U
w
h/ = 2600 and
Ro 2h/U
w
= 0.1. Vectors: secondary ow eld in a plane perpendicular to
the streamwise direction; Contours: streamwise mean velocity. From Anders-
son et al. (1998).
1993). However, the observed roll-cell patterns were not steady. In this case
the turbulence model is assumed to represent the entire spectrum of unsteady
motion. But, in rotating Couette ow, which is case b of gure 6, steady roll
cells have been observed. A computation of such ow should include all three
velocity components to allow such cells to form. They were obtained in the
following computations of plane turbulent Couette ow subjected to spanwise
rotation.
In contrast to the pressure-driven channel ow, when the ow eld is driven
solely by a moving wall the mean vorticity is single signed and the mean
velocity prole is antisymmetric (gure 6b). Hence the ow is exposed entirely
[4] The elliptic relaxation method 145
either to cyclonic or to anticyclonic rotation. The antisymmetry of the mean
ow eld is therefore preserved in a noninertial frame of reference.
Pettersson and Andersson (1997) computed plane Couette ow subjected to
anticyclonic rotation, assuming a unidirectional mean ow eld and ignoring
the roll cells. That assumption met with limited success. It was conjectured
that the signicant discrepancies that existed between model predictions and
DNS data could be attributed to organized large scale structures. The latter
were observed in the DNS of Bech and Andersson (1997). In that study it was
found that the rotation induced streamwise vortices within an intermediate
range of rotation numbers. The vortices spanned the entire channel, from top
to bottom. In contrast to rotating Poiseuille ow, these roll-cells were observed
to be steady. They should truly be considered part of the mean ow. Correct
usage of Reynolds averaged closure then requires that the vortices be computed
explicitly; the turbulence model is only responsible for the incoherent portion
of the uid motion.
Andersson et al. (1998) therefore adopted a more correct and physically
appealing approach: they treated the secondary ow as an integral part of a
two-dimensional, three-component mean velocity eld. The ow eld was of
the form
U = [U(y, z), V (y, z), W(y, z)].
The turbulence model was then left to represent only the real turbulence. The
mean ow roll-cells appeared automatically in the computation.
In this case the governing mean momentum equations for the two-dimen-
sional, three-component ow eld are
0 = 2V +
y
_
U
y
uv
_
+
z
_
U
z
uw
_
P
y
= 2U +
y
_
V
y
v
2
_
+
z
_
V
z
vw
_
P
z
=
y
_
W
y
vw
_
+
z
_
W
z
w
2
_
0 =
V
y
+
W
z
(4.2)
All six components of the Reynolds stress tensor are required in this ow.
Even though u
2
does not appear in the mean ow equations (4.2), it enters
the computation because the model depends on k. The full set of Reynolds
stress transport and elliptic relaxation equations (3.4) were solved.
Figure 8 shows the predicted secondary ow in a cross-plane of the channel.
The maximum value of the secondary ow at Ro = 0.1 is approximately 13% of
the mean streamwise bulk velocity. The width of each of the roll-cells is about
146 Durbin and Pettersson-Reif
Figure 9: Axially rotating pipe ow at Re 2RU
b
/ = 20000. N U
wall
/U
b
.
Experiment by Imao et al. (1996).
equal to the height of the channel. The dimensions of the cells are relatively
insensitive to the width of the computational domain, as long as it is large
enough to capture the pair of vortices; the particular vortices illustrated here
are from a computation with a domain wide enough to capture two complete
pairs. The predicted spanwise averaged, mean streamwise velocity is shown in
the upper part of gure 8; it is in excellent agreement with the DNS data.
A word of warning: it should be emphasized that the validity of this ap-
proach requires a truly steady roll-cell pattern. When such steady secondary
ow is present it should be computed as part of the mean ow. Similarly if
coherent, periodic vortex shedding is present, it too will be part of the mean
ow computation. Turbulence models are not designed to represent determin-
[4] The elliptic relaxation method 147
istic structures. But if large-scale, incoherent unsteadiness is present it must
be regarded as part of the turbulence and to lie within the province of the
model.
2
4.4 Axially Rotating Pipe
Fully developed turbulent ow inside an axially rotating pipe constitutes an-
other interesting test case. It has relevance to internal cooling in turbomachin-
ery. The ow eld can be assumed to be one-dimensional with two velocity
components: U = [0, U
(r), U
z
(r)] see gure 9. The imposed circumferen-
tial wall velocity, U
,wall
, is stabilizing. It suppresses the turbulence across the
entire pipe. The direction of the mean velocity changes rapidly through the
wall-layer, from being circumferential at the wall, to becoming nearly axial
away from the vicinity of the wall. This rapid turning of the mean ow direc-
tion close to a solid boundary is analogous to what is seen in three-dimensional
turbulent boundary layers. It produces skewing between axes of the Reynolds
stress and of the mean rate of strain. Therefore it is desirable to integrate the
governing set of equations all the way to the wall. Pettersson and Andersson
(1997) employed several pressure-strain models in conjunction with elliptic
relaxation to compute this ow. The mean momentum equations in this case
are
0 =
P
z
+
_
2
U
z
r
+
1
r
U
z
r
_
1
r
(ru
r
u
z
)
r
0 =
_
2
U
r
2
+
1
r
U
r
_
1
r
(ru
r
u
)
r
.
(4.3)
As in the previous case, all six components of the Reynolds stress tensor are
non-zero and the full set of transport equations has to be solved.
Figure 9 displays the mean axial and mean circumferential velocity com-
ponents across the pipe using the SSG and IP models. The SSG model cap-
tures the departure of the circumferential velocity from solid body rotation
(U
r) much better than the IP model. The departure from solid body rota-
tion is a feature that is peculiar to turbulent rotating pipe ow. It originates in
Reynolds stress production terms created by the rotation: in particular, u
r
u
is the solution to
0 = P
r
C
r
+
r
+T
r
. (4.4)
The crucial terms in this equation are production and convection:
P
r
= u
2
r
U
r
+u
2
r
and C
r
= (u
2
r
u
2
)
U
r
.
Under solid body rotation rU
/r = U
.
2
Editors note. Readers will note that a dierent view is taken in [22].
148 Durbin and Pettersson-Reif
x
y
z
2h
2h
Schematic of square duct
0.0 1.0 2.0
0.0
1.0
2.0
z/h
y/h
Figure 10: Fully developed ow in a straight square duct at Re
2hu
/ =
600. Vectors: Secondary mean ow eld; Contours: Mean streamwise velocity.
SSG model. (Petterssson-Reif and Andersson 1999.)
The second of (4.3) implies that in the inviscid limit u
r
u
= 0 since it must
be non-singular at the center of the pipe. The U
2hu
/ =
600. DNS data of Huser and Biringen (1993). (From Petterssson-Reif and
Andersson 1999.)
4.5 Square Duct
The majority of geophysical and engineering ows exhibit a non-zero mean
uid motion in the plane perpendicular to the primary ow direction. The
generation of this secondary motion can be attributed to two fundamentally
dierent mechanisms: (i) quasi-inviscid deection of the mean ow eld due
to body forces (such as the Coriolis force discussed above); (ii) an imbalance
between gradients of the Reynolds-stress components. The latter mechanism is
termed Prandtls secondary ow of the second kind. The standard illustration
of secondary ow of the second kind is turbulent ow inside noncircular ducts.
That is the subject of this nal example.
From a modeling perspective, the square duct constitutes a demanding test
case for any turbulence model. Regions approximating one, two and three-
component mean ow exist and the model must perform well in all of them
simultaneously.
Figure 10 shows a computation in this geometry. The pattern of velocity
vectors has 8-fold symmetry with respect to rotations and reections. The
secondary ow depicted in gure 10 can be attributed to streamwise vorticity.
The transport equation for streamwise vorticity is
V
x
y
+W
x
z
=
_
x
y
2
+
2
x
z
2
_
+
2
(v
2
w
2
)
yz
2
vw
y
2
+
2
vw
z
2
(4.5)
150 Durbin and Pettersson-Reif
where
x
=
W
y
V
z
. The two last terms are referred to as turbulent source
terms: the rst involves normal stress anisotropy; the second involves the sec-
ondary Reynolds shear stress. The rst has traditionally been recognized as
driving the streamwise vorticity. However, recent DNS of Huser et al. (1994)
indicate, to the contrary, that the net source term is dominated by the shear
stress contribution.
Pettersson-Reif and Andersson (1999) computed the fully developed turbu-
lent ow in a straight square duct. A full SMC was for the rst time integrated
all the way to the wall in this ow. The secondary ow had the correct eight-
fold symmetry about the two wall mid-planes and the corner bisectors of the
eld shown in gure 10. The discrepancy most noted by Pettersson-Reif and
Andersson (1999) is that the predicted secondary ow was even weaker than
that of the data. However, the secondary ow is quite weak, being only a few
percent of the centerline velocity, and its prediction is of questionable practical
importance. Figure 11 compares model predictions with DNS data from Huser
and Biringen (1993) at Re
2hu
ij
, (1)
156 Leschziner and Lien
where
d
ij
=
x
k
_
C
s
k
u
k
u
l
u
i
u
j
x
l
_
,
P
ij
=
_
u
i
u
k
U
j
x
k
+u
j
u
k
U
i
x
k
_
,
ij,1
= C
1
k
_
u
i
u
j
2
3
ij
k
_
,
ij,2
= C
2
_
P
ij
1
3
P
kk
ij
_
,
w
ij
=
_
kl
n
k
n
l
ij
3
2
ik
n
j
n
k
3
2
jk
n
i
n
k
_
f,
ij
= C
1w
k
u
i
u
j
+C
2w
ij,2
,
(2)
and f is a wall-distance function to be dened later.
For further considerations, it is advantageous to treat the stress components
and the related equations explicitly, i.e. as separate entities, and to adopt a
general non-orthogonal coordinate framework; this is, in fact, the environment
into which the closure has been implemented. Without loss of generality, in
terms of numerical-implementation issues, and to enhance clarity, attention is
conned to plane 2D ow, for which the set of equations (1) may be written,
after some manipulation and insertion of (2) into (1), as follows:
1
J
_
_
U
c
1
J
(q
11
+q
12
)
_
+
_
V
c
1
J
(q
12
+q
22
)
__
=
1
P
11
+
2
P
22
+
3
P
12
+
4
P
k
+
k
_
5
u
2
+
6
v
2
+
7
uv
_
+
8
,
= u
2
, v
2
, uv,
(3)
where J is the Jacobian linking the (x, y) and (, ) systems, the subscripts
and denote dierentiation with respect to these coordinates,
q
11
=
C
s
k
_
u
2
y
2
2uvx
+v
2
x
2
_
, (4)
q
22
=
C
s
k
_
u
2
y
2
2uvx
+v
2
x
2
_
, (5)
q
12
=
C
s
k
_
u
2
y
uv(x
+x
) +v
2
x
_
, (6)
the contravariant velocities U
c
and V
c
are given by:
U
c
= Uy
V x
, V
c
= V x
Uy
, (7)
and the turbulence production P
ij
arise as:
P
11
=
2
J
_
u
2
(U
) +uv(U
)
_
(8)
[5] Applying Second-Moment Closure to Complex Flows 157
P
22
=
2
J
_
v
2
(V
) +uv(V
)
_
(9)
P
12
=
J
_
u
2
(V
) +v
2
(U
)
+uv(U
+V
)
_
, (10)
and P
k
= 0.5(P
11
+ P
22
). The coecients
i
in the above expressions for the
individual stress components are summarized in Table 1. In this table, the
wall-damping functions f
x
, f
y
and f
xy
for any wall are:
f
x
= n
2
1
f, f
y
= n
2
2
f, f
xy
= n
1
n
2
f, (11)
where, as shown in Figure 1, (n
1
, n
2
) are directional cosines relating the
global system unit-vectors (e
1
, e
2
) to the local wall-normal vector e
2
, and
f = (k
3/2
/)/(C
l
l
n
), with l
n
being the wall-normal distance.
Figure 1: Global and local wall-aligned systems of unit vectors for any curved
wall.
Other closures, e.g. the low-Re versions of Jakirlic and Hanjalic (1995) and
Craft and Launder (1996), contain additional terms associated with the uid
viscosity and/or with higher-order approximations for the pressure-strain in-
teraction terms. The implementation issues discussed below apply to all vari-
ants, except for aspects pertaining to the imposition of wall boundary condi-
tions an issue requiring more careful attention in high-Re models coupled to
wall laws.
3 Numerical Issues
3.1 General Considerations
Experience shows that the incorporation of second-moment closure into general-
ow solvers is a non-trivial task, mainly because the equations exhibit a nu-
merically sti behaviour. Specically, the equations contain large source-like
terms, are highly nonlinear and are strongly coupled. In addition, momentum
158 Leschziner and Lien
Table 1: Coecients associated with the sources of the Reynolds-stress equa-
tions.
= u
2
1
1 C
2
+ 2C
2
C
2w
f
x
5
(C
1
+ 2C
1w
f
x
)
2
C
2
C
2w
f
y
6
C
1w
f
y
3
C
2
C
2w
f
xy
7
C
1w
f
xy
4
2
3
(C
2
2C
2
C
2w
f
x
+C
2
C
2w
f
y
)
8
2
3
(C
1
1)
= v
2
1
C
2
C
2w
f
y
5
C
1w
f
x
2
1 C
2
+ 2C
2
C
2w
f
x
6
(C
1
+ 2C
1w
f
y
)
3
C
2
C
2w
f
xy
7
C
1w
f
xy
4
2
3
(C
2
2C
2
C
2w
f
x
+C
2
C
2w
f
y
)
8
2
3
(C
1
1)
= uv
1
1.5C
2
C
2w
f
xy
5
1.5C
1w
f
xy
2
1.5C
2
C
2w
f
xy
6
1.5C
1w
f
xy
3
1 C
2
+ 1.5C
2
C
2w
(f
x
+f
y
)
7
[C
1
+ 1.5C
1w
(f
x
+f
y
)]
4
2C
2
C
2w
f
xy
8
0
diusion does not arise in the form of second-order derivatives of the subject
variable (momentum), and this can have a dramatically adverse eect on the
stability of the solution.
The impact on stability depends critically on the manner in which the trans-
ported variables are stored spatially, both relative to one another and rela-
tive to the cells over which conservation is satised (cell-vertex/cell-centred,
staggered/collocated storage), the type and disposition of the numerical grid
(structured/unstructured, skewness, aspect ratio), the degree of implicitness
and inter-variate coupling maintained by the solution algorithm, and, to some
extent, also on the complexity of the ow and its geometry (e.g. boundary-layer
vs. massively separated ow).
There are some major dierences between semi-implicit, pressure-based
schemes (e.g. SIMPLE), which are used extensively for incompressible ows,
and density-based schemes used for compressible ows. In the latter cate-
gory, two main groups of schemes are conventional explicit, decoupled, time-
marching formulations (e.g. RungeKutta), which solve the conservation laws
[5] Applying Second-Moment Closure to Complex Flows 159
in their basic form in a segregated manner, and the more modern Riemann-
solver-based schemes which account for the propagation of waves and which
are usually combined with implicit time-integration. The degree of implicit-
ness and coupling in the latter framework can vary greatly, however: it can
encompass the entire set of transport equations or only subsets; it can include
only coupling via the uxes or via the source terms; it can involve coupling
between variables at each point or include spatial coupling.
All the above factors inuence stability, in general, and the degree of di-
culty encountered in the incorporation of second-moment closure, in particu-
lar. The diversity of approaches and the resulting dierences in their numerical
properties means that there are no universally applicable rules or algorithmic
measures for a stable and ecient implementation of second-moment closure.
Rather, stability-promoting practices have evolved in response to the need to
achieve stability or increase the convergence rate of dierent algorithms.
In what follows, some advantageous stability-promoting practices, which are
used for incompressible and compressible ows, are introduced. None of the
practices is essential in all circumstances, and no practice can be claimed to
secure stability unconditionally.
3.2 Collocated-storage, Pressure-based Algorithms
Apart from the problems associated with non-orthogonality, especially at bound-
aries, the main diculty in combining a stress-transport model with a collo-
cated nite-volume scheme for complex geometries arises from the fact that
storage of all variables at the same spatial location tends to lead to che-
querboard oscillations, caused by an inappropriate decoupling of velocity and
Reynolds stresses when linear interpolation is used to approximate cell-face
stresses in terms of nodal values. The practices outlined below re-establish
coupling by use of interpolation methods analogous to those proposed by Rhie
and Chow (1983) for momentum interpolation in the solution of the mean-
ow equations within a pressure-based strategy. While the considerations to
follow assume the ow to be incompressible, the practices introduced also ap-
ply, subject to some minor variations, to compressible ows computed with
pressure-based schemes. Computational studies on transonic ows combining
pressure-based schemes with Reynolds-stress-transport models include those
of Lien and Leschziner (1993b) and Leschziner and Ince (1995).
To highlight the underlying rationale, it is instructive to focus rst on the
Boussinesq relationship,
u
i
u
j
=
T
_
U
i
x
j
+
U
j
x
i
_
2
3
ij
k. (12)
As is evident, u
2
is driven by U/x, while uv is driven by U/y and
V/x. To retain this physical coupling in the numerical representation, it
is necessary to store u
2
between U-velocity locations approximating U/x.
160 Leschziner and Lien
Figure 2: Staggered Reynolds-stress storage locations.
Similarly, uv needs to be stored between U- and V -velocity nodes approx-
imating U/y and V/x, respectively. This rationale is reected by the
arrangement shown in Figure 2. In a 3D environment, a total of seven sepa-
rate control volumes are thus required if the Reynolds-stress model is used.
Examples of 3D ows computed with this arrangement may be found in Lin
and Leschziner (1993) and Iacovides and Launder (1985), the latter using an
algebraic Reynolds-stress model.
The above methodology is, of course, untenable in a collocated arrange-
ment. To retain coupling, a nonlinear interpolation practice has been devised
which prevents stress-related odd-even oscillations. In order to facilitate trans-
parency, the method is explained rst by reference to a 2D Cartesian arrange-
ment having a uniform mesh x = y, with generalization pursued later.
With the Reynolds-stress model of Gibson and Launder (1978) chosen to rep-
resent turbulence transport, it may be shown that the discretized form of the
transport equation for the normal stress u
2
at the location P may be written:
u
2
P
=
m=E,W,N,S
A
m
u
2
m
+S
C
. .
H
P
/A
P
+
P
11
(U
w
U
e
)
P
x
, (13)
where
P
11
=
__
2
4
3
C
2
+
8
3
C
2
C
2w
f
x
+
2
3
C
2
C
2w
f
y
_
u
2
_
P
xy
A
P
, (14)
and S
C
includes a cross-diusion term arising from Daly and Harlows stress-
diusion model (1970) and some fragments of the production, pressure-strain
and dissipation processes. Analogous expressions for u
2
E
and u
2
e
are:
u
2
E
=
H
E
A
E
+
E
11
(U
w
U
e
)
E
x
, u
2
e
=
H
e
A
e
+
e
11
(U
P
U
E
)
x
, (15)
[5] Applying Second-Moment Closure to Complex Flows 161
with
e
11
= (
P
11
+
E
11
)/2. If the H/A terms are also linearly interpolated, the
resulting form of u
2
e
is:
u
2
e
=
1
2
_
u
2
P
+u
2
E
_
. .
linear interpolation
(16)
+
1
2
__
P
11
+
E
11
_
(U
P
U
E
)
P
11
(U
w
U
e
)
P
E
11
(U
w
U
e
)
E
_
. .
Uvelocity smoothing
/x,
which is identical to the form proposed by Obi et al. (1989).
The above practice of extracting apparent viscosities (e.g.
11
associated
with the mean strain U/x) had already been suggested by Huang and
Leschziner (1985) who employed the staggered Reynolds-stress arrangement.
Their objective was to enhance the iterative stability by increasing the magni-
tude of the diagonal coecient A
P
. In contrast, it may be seen from (16) that
the same apparent-viscosity approach applied to the collocated arrangement
for all Reynolds stresses introduces fourth-order smoothing, here depending
on the U-velocity component rather than, as is the case with Rhie and Chows
interpolation, on pressure. Equation (16) has been derived for steady-state
conditions without the use of under-relaxation in the solution sequence. The
generalization to unsteady conditions and with under-relaxation included re-
quires care to ensure that the solution will not depend on the under-relaxation
factor, and that the cell-face interpolants do not depend on the time step. De-
tails on this extension may be found in Lien and Leschziner (1994a).
In deriving the above expression for the apparent viscosity, an element of
uncertainty is that the level of this viscosity is a strong function of the manner
in which the source in the u
2
-equation is transformed into the equivalent quasi-
linear form via:
JS
u
2
S
C
+S
P
u
2
P
. (17)
To appreciate the nature of the problem, it must rst be pointed out that
11
depends strongly on A
P
via (14). But A
P
contains S
P
, and this value depends
on how the terms encountered in S
u
2
are treated. It is a general objective
here to maximize the magnitude of the (negative!) fragment S
P
. If any term
in S
u
2
is found to be negative, but does not contain u
2
as a factor, it may
nevertheless be allocated to S
P
by dividing that term by u
2
(the previous
iterate) and then adding the result to S
P
. Because S
P
varies greatly from
node to node, the averaging process (16) may yield a value for A
e
which
gives an inappropriate level of
e
11
(which equals 0.5[
P
11
+
E
11
]). This leads to
the conclusion that it would be desirable to construct a scheme which allows
11
(and other apparent viscosities) to be determined without reference to the
discretization process and its details. To this end, attention is directed towards
162 Leschziner and Lien
the dierential equation governing u
2
. This may be written in the following
compressed form:
C
11
D
11
= AB +
_
P
11
+
11
2
3
+A+B
_
, (18)
with
A =
k
(C
1
+ 2C
1w
f
x
) u
2
(19)
B =
_
2
4
3
C
2
+
8
3
C
2
C
2w
f
x
+
2
3
C
2
C
2w
f
y
_
u
2
U
x
, (20)
and C
11
, D
11
, P
11
and
11
being, respectively, convection, diusion, production
and redistribution. As regards terms A and B, suce it to say here that both
arise naturally from fragments of P
11
and
11
. Combination of (18) to (20)
yields:
u
2
=
11
U
x
+
_
k
_
C
11
D
11
(P
11
+
11
2
3
+A+B)
C
1
+ 2C
1w
f
x
, (21)
with
11
=
2
4
3
C
2
+
2
3
C
2
C
2w
(4f
x
+f
y
)
C
1
+ 2C
1w
f
x
ku
2
. (22)
Note that
11
in (21) and (22) is associated with the dierential gradient; in
contrast,
P
11
in (13) is applicable to the gradient approximation at grid-node
P. The derivation of
22
follows a path analogous to (18) to (22). Hence, only
the end result is given, namely:
22
=
2
4
3
C
2
+
2
3
C
2
C
2w
(4f
y
+f
x
)
C
1
+ 2C
1w
f
y
kv
2
. (23)
Attention is turned next to the interpolation formula for the shear stress uv.
A treatment consistent with the one applied in relation to u
2
would involve
extracting a viscosity
12
by reference to
U
y
and
V
x
. This is not possible,
however, because the fragments multiplying
U
y
and
V
x
in the stress model
are not identical. Hence, here, uv is only sensitized to one of the two strains;
which one is chosen is dictated by the direction of the derivative of the shear
stress. Since in the U-momentum equation the shear-stress gradient is
uv
y
, it
is natural to relate uv to
U
y
, leading to the stability-promoting second-order
derivative
(
12
U/y)
y
. The same rule can be applied to the V -momentum
[5] Applying Second-Moment Closure to Complex Flows 163
equation, resulting in the diusion term
(
21
V/x)
x
. In order to derive the
two apparent viscosities
12
and
21
from the uv-equation, this equation is
rst written as follows:
C
12
D
12
=
k
_
C
1
+
3
2
C
1w
(f
x
+f
y
)
_
uv
_
1 C
2
+
3
2
C
2
C
2w
(f
x
+f
y
)
_ _
v
2
U
y
+u
2
V
x
_
. (24)
Then (24) can be reformulated in either of the two forms:
uv =
12
U
y
+
_
k
_
C
12
D
12
+
_
1 C
2
+
3
2
(f
x
+f
y
)
_
u
2
V
x
C
1
+
3
2
C
1w
(f
x
+f
y
)
, (25)
or
uv =
21
V
x
+
_
k
_ C
12
D
12
+
_
1 C
2
+
3
2
(f
x
+f
y
)
_
v
2
U
y
C
1
+
3
2
C
1w
(f
x
+f
y
)
, (26)
where
12
=
1 C
2
+
3
2
C
2
C
2w
(f
x
+f
y
)
C
1
+
3
2
C
1w
(f
x
+f
y
)
_
kv
2
_
(27)
21
=
1 C
2
+
3
2
C
2
C
2w
(f
x
+f
y
)
C
1
+
3
2
C
1w
(f
x
+f
y
)
_
ku
2
_
. (28)
To extend the above concepts to the general curvilinear environment, attention
is next focused on the U- and V -momentum equations, written in terms of the
general coordinates (, ):
(U
c
U)
+
(V
c
U)
=
(P +u
2
)y
+
(P +u
2
)y
+
(uv)x
(uv)x
(29)
(U
c
V )
+
(V
c
V )
=
(P +v
2
)x
(P +v
2
)x
(uv)y
+
(uv)y
.
(30)
164 Leschziner and Lien
It is clear from (29) and (30) that no physical (second-order) diusion terms
arise naturally. In order to extract apparent viscosities from the Reynolds-
stress equations in terms of the (, ) coordinate system, a tedious, but other-
wise rather straightforward manipulation of the transformed equations, anal-
ogous to that in the Cartesian framework, may be carried out. Interestingly,
the nal expressions are identical to equations (22), (23), (27) and (28), except
that the wall-related damping function in the pressure-strain model assumes
dierent forms. Thus, for a single x-directed wall in the Cartesian framework,
f
y
is given by f
y
= (k
3/2
/)/(C
l
y), while, in contrast, f
y
along a curved sur-
face becomes f
y
= n
2
2
(k
3/2
/)/(C
l
l
n
) (see equation (11)). Introduction of the
apparent viscosity into (29) and (30) leads to:
_
U
c
_
q
1
J
_
_
+
_
V
c
_
q
2
J
_
_
= JS
(31)
where for = U:
q
u
1
=
11
y
2
+
12
x
2
, q
u
2
=
11
y
2
+
12
x
2
, (32)
JS
u
=
_
P
_
y
+
_
P
_
y
_
(u
2
)
u
_
y
+
_
(u
2
)
u
_
y
+
_
(uv)
u
_
x
_
(uv)
u
_
x
,
(33)
with
(u
2
)
u
= u
2
+
_
11
y
J
_
U
, (u
2
)
u
= u
2
11
y
J
_
U
,
(uv)
u
= uv
_
12
x
J
_
U
, (uv)
u
= uv +
_
12
x
J
_
U
,
(34)
while for = V :
q
v
1
=
21
y
2
+
22
x
2
, q
v
2
=
21
y
2
+
22
x
2
, (35)
JS
v
=
_
P
_
x
_
P
_
x
+
_
(v
2
)
v
_
x
_
(v
2
)
v
_
x
_
(uv)
v
_
y
+
_
(uv)
v
_
y
,
(36)
with
(v
2
)
v
= v
2
22
x
J
_
V
, (v
2
)
v
= v
2
+
_
22
x
J
_
V
,
(uv)
v
= uv +
_
21
x
J
_
V
, (uv)
v
= uv
_
21
y
J
_
V
.
(37)
[5] Applying Second-Moment Closure to Complex Flows 165
3.3 Analysis of Fourth-Order Smoothing
While equation (31), incorporating (32)(37), is mathematically identical to
the original RANS equation written in a curvilinear coordinate system, im-
portant dierences between the two forms arise from particular choices of
discretization practices. One specic practice is to approximate both the ap-
parent diusion term on the LHS of equation (31) and the source terms (JS
)
on the RHS by appropriate second-order central dierences. The result is the
introduction of an articial fourth-order smoothing term which prevents the
stress-velocity decoupling that would arise if equation (31) were written in
its original form, without the apparent diusion terms. The mechanism by
which this is achieved is best conveyed by considering equation (31) in its
Cartesian-coordinate form:
(U)
x
+
(V )
y
_
x
_
q
x
_
+
x
_
q
y
__
. .
apparent diusion
= S
(38)
where, for = U,
q
u
1
=
11
, q
u
2
=
12
S
u
=
P
x
(u
2
)
ux
x
(uv)
uy
y
,
(39)
with
(u
2
)
ux
= u
2
+q
u
1
U
x
, (uv)
uy
= uv +q
u
2
U
y
. (40)
By reference to the collocated arrangement shown in Figure 3, the rst ap-
parent diusion term on the LHS of equation (38) may be approximated by:
x
_
q
x
_
(q
u
1
)
e
U
E
[(q
u
1
)
e
+ (q
u
1
)
w
] U
P
+ (q
u
1
)
w
U
W
(x)
2
. (41)
The normal-stress gradient on the RHS of equation (39) is also approximated
by a central dierence, yielding:
x
_
(u
2
)
ux
_
_
(u
2
)
ux
_
E
_
(u
2
)
ux
_
W
2x
=
_
q
u
1
U
x
_
E
_
q
u
1
U
x
_
W
2x
+
_
u
2
_
x
.
(42)
With
_
q
u
1
U
x
_
E
=
(q
u
1
)
E
(U
EE
U
P
)
2x
,
_
q
u
1
U
x
_
W
=
(q
u
1
)
W
(U
P
U
WW
)
2x
, (43)
166 Leschziner and Lien
Figure 3: Collocated cell storage arrangement.
equation (42) becomes:
x
_
(u
2
)
ux
_
(q
u
1
)
E
U
EE
[(q
u
1
)
E
+ (q
u
1
)
W
] U
P
+ (q
u
1
)
W
U
WW
4(x)
2
+
_
u
2
_
x
.
(44)
Note that approximation (41) and the penultimate term of equation (42) sup-
posedly represent one and the same second-order derivative,
x
_
q
x
_
. If
one assumes that q
u
1
is a constant, it is evident that the dierence between the
two approximations, D
U
, is proportional to:
D
U
U
EE
4U
E
+ 6U
P
4U
W
+U
WW
, (45)
which is a fourth-order velocity-based smoothing. This is analogous to the
smoothing term in equation (16), except that the apparent viscosity
11
is
here derived from the dierential form, instead of the discretized form of the
second-moment closure.
3.4 Source-Term Linearization
Integration of equation (3) over the nite volume shown in Figure 3, applica-
tion of the Gauss Divergence Theorem and approximation of the uxes with
appropriate second-order schemes yields the following algebraic expression for
the cell-centroidal value of any scalar ow property :
A
P
p
=
m=E,W,N,S
A
m
m
+JS
. (46)
The form of the coecients A
P
and A
m
(m = E, W, N, S) depends upon the
precise nature of the ux-approximation schemes (see Lien and Leschziner
[5] Applying Second-Moment Closure to Complex Flows 167
1994a, 1994b). To ensure (or promote) positivity of , in the case of the
Reynolds normal stresses, turbulence kinetic energy and turbulence dissipation
rate, the source term JS
= S
C
+S
P
P
, (47)
where S
C
0 and S
P
0, so that S
P
can be absorbed into A
P
, i.e. A
P
(A
P
S
P
), in order to increase diagonal dominance of the coecient matrix.
As a result, numerical stability is greatly enhanced. Note that JS
contains
the RHS of equation (3) and contributions associated with q
12
and q
12
the cross diusion terms on the LHS. To secure the above constraints on S
C
and S
P
for = u
2
, say, the source term is decomposed into the following two
parts:
JS
u
2
= JS
u
2
+J
k
(C
1
+ 2C
1w
f
x
)u
2
P
. .
A
J
k
(C
1
+ 2C
1w
f
x
)
. .
B
u
2
P
, (48)
with
S
C
= max(A, 0), S
P
=
min(A, 0)
u
2
P
+ 10
15
+ B. (49)
The addition of 10
15
in equation (49) is merely to avoid a computational
singularity in the limiting case of zero normal stress. Similar expressions relate
to v
2
, w
2
, k and .
3.5 Wall Conditions
Walls pose particular challenges in the context of turbulent-ow computations,
because spatial variations in the near-wall turbulence structure are intense due
to the combined inuence of viscosity and wall-induced anisotropy. When low-
Reynolds-number models are applied, numerical integration encompasses the
entire near-wall region including the viscous sublayer. Hence, in this case, the
implementation of wall conditions is (numerically) straightforward, consisting
simply of imposing no-slip and impermeability relations. The treatment is
more dicult when log-law-based wall laws are adopted in conjunction with
high-Re models to bridge the viscous sublayer. Particular diculties arise when
Reynolds-stress modelling is applied in conjunction with curved walls; it is this
aspect which is of particular interest here.
Attention is directed to a general near-wall volume abutting a curved wall,
as shown in Figure 4. The point P is assumed to be in the log-law region,
with the logarithmic variation assumed to prevail normal to the wall at any
tangential velocity position. The resultant shear force F
s
shear
acting on the
cells southern face Area
s
is:
F
s
shear
=
w
Area
s
(50)
168 Leschziner and Lien
Figure 4: Boundary cell at a curved wall.
where
w
=
P
k
1/2
P
C
1/4
lnE
l
n
k
1/2
P
U
t
(51)
U
t
= UU
n
, l
n
= (r
P
r
s
) n. (52)
Here, l
n
is the normal distance away from the wall, and U
t
and U
n
denote,
respectively, the tangential and normal velocity components (here treated as
vectorial quantities). The unit normal vector pertaining to the cell in Figure
4 is n = /[[, or in expanded form:
(n
1
, n
2
, n
3
) =
(
x
,
y
,
z
)
_
2
x
+
2
y
+
2
z
, (53)
where
_
2
x
+
2
y
+
2
z
= Area
s
. (54)
Hence, the components of U
t
in (52) can be written as:
(U
t
x
, U
t
y
, U
t
z
) = (U, V, W) (Un
1
+V n
2
+Wn
3
)n, (55)
with
U
t
x
= (1 n
2
1
)U n
1
n
2
V n
1
n
3
W, (56)
U
t
y
= (1 n
2
2
)V n
1
n
2
U n
2
n
3
W, (57)
U
t
z
= (1 n
2
3
)W n
1
n
3
U n
2
n
3
V. (58)
Once the tangential velocity components are resolved, the coecient A
S
in
the discretized equation pertaining to the near-wall cell is rst nullied, and
then the source S
C
is modied in such a manner as to explicitly include the
shear force imposed on the southern cell face as follows:
[5] Applying Second-Moment Closure to Complex Flows 169
for the x-momentum equation,
S
C
S
C
P
k
1/2
P
C
1/4
lnE
l
n
k
1/2
P
U
t
x
Area
s
; (59)
for the y-momentum equation,
S
C
S
C
P
k
1/2
P
C
1/4
lnE
l
n
k
1/2
P
U
t
y
Area
s
; (60)
for the z-momentum equation,
S
C
S
C
P
k
1/2
P
C
1/4
lnE
l
n
k
1/2
P
U
t
z
Area
s
. (61)
To implement (59)(61), the correct value of k
P
is needed. In the near-wall
cell, this is essentially governed by the balance between volume-averaged pro-
duction and dissipation. Both must be evaluated in a manner consistent with
the log-law variation in the cell. If the entire cell is assumed to reside within
the log-law region, the average k-production arises as:
P
k
=
ln(l
n
/l
v
)
l
n
k
1/2
P
(
w
w
), (62)
where l
v
is the viscous sublayer thickness. With = 2(k/l
2
n
) in the viscous
sublayer and = k
3/2
/l
n
in the log-law region, the average value of becomes:
=
k
3/2
P
l
n
_
2
l
v
k
1/2
P
+
ln(l
n
/l
v
)
C
l
_
. (63)
Since is unconditionally positive (but preceded by a minus sign in the
k-equation), iterative stability can be enhanced by the replacement:
S
P
S
P
P
k
P
J. (64)
The aforementioned modications in the near-wall region suce for the im-
plementation of the k- model. The extension of the above treatment to the
Reynolds-stress model is less straightforward than might seem at rst sight.
In a Cartesian framework, the average near-wall stress productions are well
approximated by
P
11
= 2P
k
, P
22
= 0, P
33
= 0, (65)
170 Leschziner and Lien
while the shear stress itself is given by
uv
w
. (66)
A problem here is, however, that additional average pressure-strain terms (
ij
)
need to be evaluated, because these contribute substantially to the balance
equations, unlike in the case of turbulence energy. The pressure-strain terms
contain products of stresses and strains, and the variation of the former across
the sublayer is both uncertain and highly inuential to the averaging process.
Considerable further complications arise in a non-Cartesian environment, be-
cause of the tensorial nature of the stresses, productions and
ij
, and the
consequent transformation involved in determining productions and contribu-
tions of
ij
in terms of wall-oriented coordinates. These diculties provide
motivation for an alternative route. In this, the k-equation, incorporating the
production and diusion terms appropriate to the second-moment closure, is
solved together with (62) to (64), rather than the equations for u
2
, v
2
, and
w
2
. Then, the near-wall values of the Reynolds-stress components, in terms
of wall-oriented coordinates, are determined from local-equilibrium forms of
the Reynolds-stress equations from which transport terms are omitted and
in which the log-law is used to approximate the shear strain, assumed to be
the only strain. The end result is a closed set of algebraic equations for the
wall-oriented stresses in terms of the turbulence energy:
(u
2
)
w
P
= 1.098k
P
, (v
2
)
w
P
= 0.247k
P
, (uv)
w
P
= 0.255k
P
. (67)
Finally, the Cartesian stress components are determined from the wall-oriented
Reynolds stresses through a local coordinate transformation as follows (see
Lien and Leschziner 1993a):
u
2
P
= (u
2
)
w
P
t
2
1
+ (v
2
)
w
P
n
2
1
+ 2(uv)
w
P
t
1
n
1
(68)
v
2
P
= (u
2
)
w
P
t
2
2
+ (v
2
)
w
P
n
2
2
+ 2(uv)
w
P
t
2
n
2
(69)
uv
P
= (u
2
)
w
P
t
1
t
2
+ (v
2
)
w
P
n
1
n
2
+ 2(uv)
w
P
(t
1
n
2
+t
2
n
1
), (70)
where t
i
and n
i
are the components of the tangential and wall-normal unit
vectors, respectively.
3.6 Approximation of Turbulence Convection
It is generally assumed that the numerical approximation of turbulence con-
vection is of subordinate importance, because the associated equations are
dominated, in most shear ows, by source and sink terms arising from genera-
tion, dissipation and redistribution. This is often put forward as a justication
for the use of the rst-order upwind scheme to approximate turbulence con-
vection, so as to increase the diagonal dominance of the discretized sets of
[5] Applying Second-Moment Closure to Complex Flows 171
equations and thus enhance the stability of the solution. Although the upwind
scheme introduces a substantial amount of articial turbulence diusion, the
argument is that this diusion only makes a small contribution to the balance
expressed by the equations.
The principal requirement for numerical stability in the solution of the
turbulence-model equations is that the numerical scheme should not intro-
duce articial minima which could cause negative values for (physically) un-
conditionally positive quantities (e.g. normal stresses and dissipation) and
hence instability. While this requirement is met by the rst-order upwind
scheme, a much less diusive approach which is entirely satisfactory is to use
a modern Total Variation Diminishing (TVD) scheme. This in a nonlinear
method, in that the approximation is sensitive to the local solution, introduc-
ing just enough articial diusion to eliminate extremes which the basic non-
TVD scheme would normally provoke. An example is the UMIST (Upstream
Monotonic Interpolation for Scalar Transport) scheme of Lien and Leschziner
(1994b), implemented in combination with second-moment closure and the
foregoing stability-promoting measures into a general 3D non-orthogonal-grid
algorithm for complex ows. The UMIST scheme is a monotonic form of the
QUICK scheme of Leonard (1979) and has been constructed along the lines of
van Leers (1979) MUSCL scheme.
Applications reported in Lien and Leschziner (1994b) conrm that many
ows are largely insensitive to the accuracy of approximating turbulence con-
vection. However, there are ows in which this accuracy is important. One
group include boundary layers undergoing bypass transition, in which the pre-
cise representation of the evolution of the turbulence quantities can have a
substantial impact on the position of transition. Figure 5, taken from Chen
et al. (1998), illustrates this sensitivity for the case of a transitional at-plate
boundary layer computed with a low-Re linear eddy-viscosity model.
The general message is that it is always advisable to apply the most accurate
discretization scheme, subject to stability constraints, to the turbulence equa-
tions. This is especially pertinent to second-moment closure, because stress
convection can become inuential in some ows in which turbulence-damping
processes (e.g. swirl and density stratication) diminish generation and redis-
tribution relative to transport.
3.7 Multigrid Acceleration
Multigrid relaxation has become a well-established method for accelerating the
convergence of elliptic and hyperbolic ows, in the wake of Brandts pioneering
work in the 1970s (Brandt 1977). In its simplest form, a two-level multigrid
scheme transfers (restricts) residuals from the (ne) working grid to a coarser
grid, and then returns (prolongates) incremental changes (corrections) result-
ing from a reduction in the coarse-grid residuals to the ner working grid. This
process is followed by a few solution (relaxation) steps on the ner grid to yield
172 Leschziner and Lien
Figure 5: Sensitivity of predicted transition of a at-plate boundary layer to
the approximation of turbulence convection (from Chen et al. 1998).
the nal solution. As residuals decay more rapidly on coarse grids than on ne
grids, this double transfer, with residual relaxation (smoothing) eected on
the coarser grid, tends to give much faster convergence than a relaxation on
the working grid only. In practice, elaborate multigrid cycles are used, in
which residuals, corrections and actual solutions are transferred through a se-
quence of coarsening and rening grids. Figure 6 shows, schematically, three
alternative cycles, including a full multigrid V-cycle.
Many applications have been reported in which variants of the multigrid
method have been exploited to accelerate the convergence of computations for
high-speed inviscid and laminar ows. In these relatively simple conditions, the
CPU costs tend to increase in proportion to N log N, where N is the number
of grid nodes. This (almost) linear increase compares with quadratic or even
cubic rates of increase for conventional single-grid relaxation schemes.
Applications to complex turbulent ows are relatively rare. Experience shows
that, with some special steps in the prolongation and restriction operations
applied to the turbulence-transport equations (see Lien and Leschziner 1994c),
the eectiveness of the multigrid scheme is generally maintained, if turbulence
eects are represented by eddy-viscosity models, provided the grid is not too
distorted and the cell-aspect ratio does not exceed O(10) to O(50). Hardly
any experience exists on the performance of multigrid schemes in computa-
tions using second-moment closure.
Lien and Leschziner (1994c) have undertaken an extensive study of the
performance of multigrid schemes in a wide range of ows, including turbu-
lent ows computed with two-equation eddy-viscosity models (both high-Re
[5] Applying Second-Moment Closure to Complex Flows 173
Figure 6: Alternative multigrid cycles.
and low-Re variants) and second-moment closure. Their experience was that
the convergence acceleration achieved with second-moment closure, although
worthwhile was well below that attained for simpler models. Figure 7 shows
results for the separated ow in a 2D plane diuser computed with the k-
model and the GibsonLaunder (1978) Reynolds-stress model. Also included
are convergence histories for the U-momentum residual, in terms of work
units (WU), for computations with a single grid of 160 40 nodes and se-
quences of 2, 3 and 4 grids. As seen, convergence is much faster with the
multigrid scheme. However, it is found that the N log N behaviour is no longer
maintained.
3.8 Density-based Scheme
Although pressure-based schemes have been used in combination with second-
moment closure to compute compressible ows, including shocks (e.g. Lien and
Leschziner 1993b, Leschziner and Ince 1995), the usual approach is to solve the
mass-conservation equation directly to yield the density. There are numerous
density-based schemes for compressible ow. Probably the simplest and most
widely used is that due to Jameson et al. (1981), which solves the conservation
equations for mass, momentum and energy in a segregated, explicit manner
(e.g. by a RungeKutta method) using centred approximations for the uxes
in conjunction with stabilizing articial second/fourth-order dissipation. Most
modern upwind schemes are now based on Riemann solvers which exploit the
174 Leschziner and Lien
Figure 7: Predicted solutions and multigrid convergence histories for turbulent
ow in a 2D plane diuser (from Lien and Leschziner 1994c).
characteristics of the conservation equations and resolve the characteristic lines
along which acoustic and contact waves propagate. These characteristics arise
upon the diagonalization of the Jacobian ux matrices A, B and C of
the conservation set, normally written in the form:
Q
t
+
(F F
v
)
+
(GG
v
)
+
(HH
v
)
=
Q
t
+A
Q
+B
Q
+C
Q
= S,
(71)
[5] Applying Second-Moment Closure to Complex Flows 175
where Q is the vector of dependent variables, F, G and H are convective
uxes, F
v
, G
v
and H
v
are diusive uxes and S is the source vector. The di-
agonalization gives the eigenvalues that dene the characteristic lines (waves)
and hence identies the upwind directions. The characteristics, representing
shock, rarefaction and contact waves, are determined as part of an exact or
approximate solution to the Riemann problem. A Riemann solver determines
the incremental change in state of the ow at a particular node, typically the
centre of a cell, as one traverses waves in a direction normal to the interface
joining the cell to its neighbour. The state of the Riemann problem at cell faces
(the interface state, determined as a function of left and right states), allows
the interface ux to be evaluated according to the directions of the waves aris-
ing within the Riemann solution. The resulting equations, including the ux
approximations in terms of dependent variables, can then be solved (marched)
in time by an explicit or implicit solution method (e.g. Euler-implicit), the lat-
ter entailing a coupled solution of the equations for example, using a Newton
linearization and a relaxation or factorization (ADI) solution.
In the case of implicit schemes, a natural and potentially very stable route
is to incorporate the turbulence-model equations into the set of conservation
equations which are then solved implicitly. Within a Riemann-solver-based
scheme, the extended Jacobian ux matrices (7 7 for 2-equation models in
3D, and 12 12 for Reynolds-stress models in 3D, per direction) have to be
linearized. If source terms are integrated based on cell-centred data, this does
not introduce any more characteristics than arise in the inviscid equation set
(eigenvalues associated with the propagation of the turbulence-model quanti-
ties are the same as that for the contact or shear wave), although the character-
istics are modied slightly by the turbulence equations for example, through
the appearance of the turbulence energy in the total-energy uxes. Then, a
full linearization of ux balances and source terms (i.e. non-ux terms), can
be handled by some implicit (e.g. Newton) method.
A fully-coupled solution is very challenging, as coupling arises at three lev-
els: point-wise coupling of the uxes, point-wise coupling via the source terms
and spatial coupling (in 3D!). In the case of two-equation models, Barakos and
Drikakis (1998) have adopted a solution of all equations within an unfactored
scheme based on a linearized Rieman solver using a point-implicit relaxation
method. In the case of Reynolds-stress closures, the task of a fully-implicit so-
lution is formidable. Morrison (1992) has presented an approximate factorized
solution process which requires the inversion of block penta-diagonal systems,
with blocks consisting of 1212 matrices for the Reynolds-stress model. How-
ever, the scheme is not fully implicit, in that source terms of the turbulence
equations associated with production, redistribution and dissipation are only
point-coupled. Further simplications introduced by Morrison led to a scheme
in which the turbulence-model equations are, in eect, decoupled from the
mean-ow equations, except for coupling eected through the presence of the
176 Leschziner and Lien
turbulence energy (normal stresses) in the uxes, via the Jacobian ux matri-
ces. Probably the most elaborate solution scheme has been employed by Vallet
(1995), who adopted a ux-vector-splitting technique (an early upwind scheme
which is somewhat more diusive than most Riemann solvers when applied
to contact discontinuities, shear waves and boundary layers). Vallet adopted a
point-implicit method, in which coupling among uxes as well as source terms
was accounted for across the entire set of 12 conservation equations. Close
examination of the presentation of the method especially the structure of
the ux and source Jacobians suggests, however, that coupling between the
turbulence-equation and mean-ow subsets is rather weak. Thus, the mean-
ow characteristics are aected by the stresses and the stresses are aected by
the contact wave, but the turbulence-model sources are not strongly coupled
to the mean-ow equations. However, coupling among the turbulence-model
equations is established through a full linearization of all sources with respect
to all turbulence quantities. The subset is then solved by sub-iterations rather
than full inversion (which would require very large amounts of storage). There
is no unambiguous evidence that this strong level of coupling is benecial to
stability and convergence. For example, Vallet reports convergence histories for
a transonic channel ow which show that convergence stalls after a reduction
in residuals by less than two orders of magnitude.
A somewhat simpler approach, adopted by Batten et al. (1997) in conjunc-
tion with Reynolds-stress modelling and a nonlinear (HLLC) Riemann solver,
is to solve the mean-ow equations by means of a block-coupled implicit scheme
and solve the set of turbulence-transport equations as a segregated set. Thus,
the ux F of any turbulence variable () is assembled by reference to the
contact-wave velocity and the (Riemann problem) interface state of the den-
sity, and then an implicit, decoupled equation is derived by including only
diagonal components of the convective, diusive and source Jacobians in the
equation, the remaining terms being treated explicitly. In eect, the ux F
in the turbulence equation is decomposed into a sum of implicit and explicit
contributions, the latter treated as a deferred correction:
F = F
Implicit
+F
Correction
. (72)
While this approach is not necessarily TVD in transients, the above treatment
is sucient to ensure positivity. The sum of the ux corrections is treated as a
source term and linearized via the following approach, due to Patankar (1980).
The uncoupled, implicit equation for any conserved scalar quantity, , may
be written as:
J
_
()
n+1
()
n
f
n+1
ds = JS
t
+S
c
, (73)
where
f
n+1
denotes the sum of all implicitly discretized uxes (including
diusion and convection terms), S
t
are the source terms arising from the tur-
bulence model, and S
c
represents the sources from any deferred-corrections.
[5] Applying Second-Moment Closure to Complex Flows 177
Linearizing convective and diusive uxes and splitting all sources into posi-
tive and negative contributions gives:
_
J
t
f
()
_
() =
f
n
+J
_
S
+
t
+S
t
_
+S
+
c
+S
c
. (74)
Positive source terms are treated explicitly, whilst negative contributions are
scaled by ()
n+1
/()
n
and moved to the left-hand side of the equations.
Introducing a small positive constant, = 10
30
, to prevent division by zero
gives:
_
J
t
f
()
_
()
_
JS
t
+S
c
()
n
+
_
()
n+1
=
f
n
+JS
+
t
+S
+
c
. (75)
To retain the form, the term
_
JS
t
+S
c
()
n
+
_
()
n
is added to both sides of equation (75) to give:
_
J
t
f
()
JS
t
+S
c
()
n
+
_
()
=
f
n
+JS
+
t
+S
+
c
+
_
JS
t
+S
c
()
n
+
_
()
n
.
(76)
In the above form, all ux-balance and source terms appear on the right-hand
side. However, the small parameter, , also appears in the denominator of
the S
or 30
and Re = 6.510
6
(based on
chord) is much more challenging, but poses signicant uncertainties due to a
complex pattern of natural transition on the windward surface. Experimental
data have been obtained by Meier et al. (1984) and Kreplin et al. (1985) for
pressure, skin friction and mean-velocity, the last with a 5-hole probe. Corre-
sponding computations have been performed by Lien and Leschziner (1995)
with the second-moment closure of Gibson and Launder (1978), coupled to a
low-Re linear eddy-viscosity model (EVM) in the viscous sublayer. A second-
order TVD scheme was used on a high-quality conformal mesh of 98 82 66
nodes, with the y
+
-value closest to the wall being kept to 0.5 to 1 across the
entire surface. Test calculations with a nonlinear EVM on a 128
3
grid have
shown only the skin friction to change slightly at this level of grid renement.
Figure 8 contains comparisons of azimuthal velocity proles at one stream-
wise location at 10
incidence, one
azimuthal pressure distribution, one skin-friction distribution and one veloc-
ity eld, the last showing the leeward separation and the associated transverse
vortex.
In Figure 8, the truncated RSTM is the GibsonLaunder model without
the wall correction
w
ij,2
to the rapid part of the pressure-strain term. This
truncation was motivated by the observation, made in computations for sepa-
rated aerofoil ows, that the correction tended to increase rather than decrease
the level of wall-normal turbulence intensity and shear stress as the boundary
layer approaches separation. In general, the nonlinear EVM and the second-
moment closure give similar results which are closer to the experimental data
that those obtained with the linear EVM. However, the improvement is not
uniformly pronounced across all ow properties, and the uncertainties associ-
ated with transition in the 30
(=
2
/2), is completely dierent from the previous model predictions. We
demonstrate the results of a critical assessment of existing
equations for
both two-equation and second-order closure models. Based on these assess-
ments, we construct a new dissipation rate equation for temperature variance,
taking into account all the budget terms in the exact
equation. Also, we
present a similarly rened k
equa-
tion, to constitute a new two-equation heat-transfer model. Comparisons of
the rened model predictions with the DNS data for a channel ow with heat
transfer are given, which shows excellent agreement for the proles of k
and
and
model are rst veried in basic ows under arbitrary wall thermal boundary
conditions and next in backward-facing-step ows at various Prandtl numbers
through a comparison of the predictions with the DNS and measurements.
1 Introduction
The turbulence model for heat transfer is a set of dierential equations which,
when solved with the mean-ow and turbulence Reynolds-stress equations,
188
[6] Modelling heat transfer in near-wall ows 189
allows calculations of relevant correlations and parameters that simulate the
behavior of thermal turbulent ows. Like the classication of turbulence mod-
els for the Reynolds stresses, phenomenological turbulent heat-transfer models
are classied into zero-equation, two-equation, and heat-ux equation mod-
els. The zero-equation heat-transfer model is a typical and most conventional
method for analyzing the turbulent heat transfer, in which the eddy diusivity
for heat
t
is prescribed via the known eddy viscosity
t
together with the most
probable turbulent Prandtl number Pr
t
, so that
t
=
t
/Pr
t
. Thus, in this
formulation the analogy is tacitly assumed between turbulent heat and mo-
mentum transfer (Launder 1988). Many previous experimental studies have,
however, revealed that there are no universal values of Pr
t
even in simple ows
(Kays 1994), e.g. at the same streamwise location, a value of Pr
t
close to the
wall is dierent from that away from the wall (Cebeci 1973; Antonia 1980). The
recent sophisticated renormalization group (RNG) theory for turbulence based
on an iterative averaging method (Nagano and Itazu 1997) indicates that the
turbulent Prandtl number changes according to the molecular Prandtl number
(Itazu and Nagano 1998). This lack of universality restricts the applicability of
a zero-equation model. On the other hand, a heat-ux equation model ought
to be more universal, at least in principle. This model, however, is still rather
primitive and extensive further research is in progress.
For the velocity eld, the linear eddy viscosity k- model of turbulence is
still regarded as a powerful tool for predicting many engineering ow problems
including jets, wakes, wall ows, reacting ows, and ows with centrifugal and
Coriolis forces (Rodi 1984). For scalar turbulence, Nagano and Kim (1988)
developed a corresponding two-equation model for heat transport (hereinafter
referred to as the NK model). They modelled the eddy diusivity for heat
t
using the temperature variance k
= (
2
/2) and the dissipation rate of tem-
perature uctuations
heat-transfer
model. Thus, a modied k
equation
is performed by taking into account all the budget terms in the exact
equation. Finally,
we verify a set of model equations using DNS data and experimental data.
In order to show the performance of the proposed two-equation heat-transfer
model, we analyze the case of a sudden change in thermal wall condition,
which is hard to measure using conventional tools; and we then investigate
the physical phenomena of the system using the results of analysis.
As mentioned above, after the rst proposal by Nagano and Kim (1988),
several k
becomes in-
nity (in other words u
k
1.4/f
t
1.3/f
t
f
t
1 + 6f
w1
f
w1
f
w
(4)
f
w2
f
w
(26)
C
1
1.45
C
2
1.9
C
3
0.005
C
4
0.5
f
2
(1 +f
2
)(1 f
w1
)1 0.6 exp[(R
t
/45)
1/2
]
f
2
exp(2 10
4
R
13
v
)[1 exp(2.2R
0.5
v
)]
R
v
(k/)[1/(1 +
t
/)](1/R
1/2
t
)f
w1
f
(1 f
w2
)1 + (60/R
3/4
t
) exp[(R
t
/55)
1/2
]
using the Kolmogorov length = (
3
/)
1/4
in place of the viscous length /u
.
2 Two-equation model of turbulence for velocity
eld
A velocity eld is described with the following continuity and momentum
equations
U
i
x
i
= 0, (2.1)
DU
i
Dt
=
1
P
x
i
+
x
j
_
U
i
x
j
u
i
u
j
_
, (2.2)
where D/Dt /t +U
j
/x
j
.
192 Nagano
In the k- model, the Reynolds stress u
i
u
j
in (2.2) (see Nagano and Tagawa
1990a) can be obtained from the following set of equations
u
i
u
j
=
t
_
U
i
x
j
+
U
j
x
i
_
2
3
ij
k, (2.3)
t
= C
k
2
, (2.4)
Dk
Dt
=
x
j
__
+
t
k
_
k
x
j
_
u
i
u
j
U
i
x
j
, (2.5)
D
Dt
=
x
j
__
+
t
_
x
j
_
C
1
k
u
i
u
j
U
i
x
j
C
2
f
2
k
. (2.6)
As indicated by Myong and Kasagi (1990), and by Nagano and Tagawa
(1990a), imposing the rigid boundary condition (i.e. no-slip) at the wall does
not necessarily lead to the correct asymptotic solutions of k y
2
, uv y
3
,
t
y
3
, and y
0
for y 0, unless the wall-limiting behavior of turbulence is
properly incorporated in the turbulence model adopted. In the present study,
for the basic formulation in the k- model, we adopt the NaganoShimada
model (Nagano and Shimada 1995a) (hereinafter referred to as the NS model),
which reproduces strictly the limiting behavior of wall. The model also repro-
duces the turbulent energy and its dissipation rate (including their budgets)
very closely for the cases of wall bounded ows:
Dk
Dt
=
x
j
__
+
t
k
_
k
x
j
_
u
i
u
j
U
i
x
j
+max
_
0.5
x
j
_
k
x
j
f
w1
_
, 0
_
, (2.7)
D
Dt
=
x
j
__
+
t
_
x
j
_
C
1
k
u
i
u
j
U
i
x
j
C
2
f
2
2
k
+f
w2
t
_
2
U
i
x
j
x
k
_
2
+C
3
k
x
k
U
i
x
j
2
U
i
x
j
x
k
+C
4
x
j
_
(1 f
w1
)
k
k
x
j
f
w1
_
. (2.8)
The NS model employed the wall friction velocity u
in this
function described as followed:
f
w
() = exp
_
_
y
_
2
_
, (2.9)
where y
= u
= ()
1/4
(or the Kolmogorov length
[6] Modelling heat transfer in near-wall ows 193
scale = (
3
/)
1/4
). This model function is more useful for analysis of various
complex ows, as conrmed by Abe et al. (1995) and Nagano et al. (1997).
The model constants and functions were optimized for the proposed function.
These are listed in Table 1. Note that, from (2.9), f
w
(4) = exp
_
(y
/4)
2
_
and f
w
(26) is similarly dened.
3 Two-equation model for heat transfer (DNS-based
modelling)
3.1 Governing Equations
The energy equation may be written
D
Dt
=
x
j
_
x
j
u
j
_
. (3.1)
In (3.1), the term on the right-hand side, the turbulent heat ux u
j
, is de-
scribed using the concept of eddy diusivity for heat
t
(see, Nagano and Kim
1988; Nagano et al. 1991),
u
i
=
t
x
i
, (3.2)
where
t
= C
k
m
. (3.3)
As a time-scale equivalent to the relative lifetime of the energy-containing
eddies or temperature uctuations, we adopt the mixed or hybrid time-scale
m
which is a function of the time-scale ratio R =
/
u
, where
u
= k/
and
= k
(k
=
2
/2) are the dynamic and scalar time-scales, respec-
tively. Obviously,
m
blends both thermal and mechanical contributions. The
characteristic length scale (i.e. the spatial extent of a uctuating temperature)
can hence be written as L
m
= k
1/2
m
, and the eddy diusivity for heat can
be modelled as
t
k
1/2
L
m
= k
m
. The present expression for
t
can be
regarded as a generalized form for the eddy diusivity introduced by Nagano
and Kim (1988). Accordingly, we incorporate near-wall eects on the thermal
eld in the model function f
t
can be expressed as a function of the state of both velocity and thermal
elds by solving the transport equations for k, , k
, and
.
The exact transport equations for k
and
Dt
= D
k
+T
k
+P
k
, (3.4)
D
Dt
= D
+T
+P
1
+P
2
+P
3
+P
4
. (3.5)
194 Nagano
The terms on the right-hand sides in (3.4) and (3.5) are identied as
Molecular diusion
D
k
=
2
k
x
j
x
j
, D
=
2
x
j
x
j
,
Turbulent diusion
T
k
=
u
j
k
x
j
, T
=
u
j
x
j
,
Mean gradient production
P
k
= u
j
x
j
, P
1
= 2
u
j
x
k
x
k
x
j
, P
2
= 2
x
k
x
j
U
j
x
k
,
Gradient production
P
3
= 2u
j
x
k
x
j
x
k
,
Turbulent production
P
4
= 2
u
j
x
k
x
k
x
j
,
Destruction
= 2
2
_
2
x
k
x
j
_
2
,
_
_
(3.6)
where k
=
2
/2 and
= (/x
k
)
2
, respectively.
3.2 Wall-Limiting Behavior of Velocity and Temperature
The behavior of the turbulent quantities of velocity and thermal elds near
the wall can be inferred from a Taylor series expansion in terms of y, together
with the continuity, momentum and energy equations, namely,
U
+
= y
+
+a
1
y
+
2
+a
2
y
+
3
+
u = b
1
y +b
2
y
2
+b
3
y
3
+
v = c
1
y
2
+c
2
y
3
+
w = d
1
y +d
2
y
2
+d
3
y
3
+
k = u
i
u
i
/2 = [(b
2
1
+d
2
1
)/2]y
2
+ (b
1
b
2
+d
1
d
2
)y
3
+
uv = b
1
c
1
y
3
+ (b
1
c
2
+c
1
b
2
)y
4
+
w
= (
2
k/y
2
)
w
= (b
2
1
+d
2
1
)
+
= Pry
+
+g
1
y
+
2
+g
2
y
+
3
+
=
w
+h
1
y +h
2
y
2
+
k
=
2
/2 =
2
w
/2 + [(h
2
1
+ 2h
2
w
)/2]y
2
+
v = c
1
w
y
2
+ (c
2
w
+c
1
h
1
)y
3
+
w
= [
2
k
/y
2
]
w
= (h
2
1
+ 2h
2
w
)
_
_
(3.7)
[6] Modelling heat transfer in near-wall ows 195
In (3.7), in view of the correspondence between k and k
/y)
w
= 0, which is exact in the case of
both uniform wall temperature and uniform wall heat ux. From (3.7), in the
vicinity of the wall, we obtain the following relations: U
+
= y
+
, u y, v y
2
,
w y,
+
= Pry
+
, and y
p/2
(where, p = 2 : without uctuations in wall
temperature,
w
; p = 0 : with
w
uctuations). These asymptotic relations
provide the representation for the wall-limiting behavior of turbulence given
as: k y
2
, uv y
3
, y
0
,
2
y
p
, v y
2+p/2
and
y
0
. Note that,
as may be seen from (3.2), v and
t
vary as the same power of y near the
wall. Consequently, from the wall-limiting behavior of turbulence, we have the
following two regimes according to the wall thermal conditions
t
y
3
for p = 2 (without
w
uctuations)
t
y
2
for p = 0 (with
w
uctuations).
_
(3.8)
As discussed later, the eddy diusivity
t
should be modelled to satisfy the
above requirements consistently.
3.3 Assessment of Modelled
Equations
3.3.1 Modelled
equations
The modelled dissipation rate equations for temperature variance used in the
current k
models for wall shear ows are written in one of two ways:
D
Dt
=
2
x
j
x
j
+T
+C
P1
f
P1
P
k
+C
P2
f
P2
k
P
k
C
D1
f
D1
2
k
C
D2
f
D2
k
+ additional term, (3.9)
D
Dt
=
2
x
j
x
j
+T
+C
P1
f
P1
P
k
+C
P2
f
P2
k
P
k
C
D1
f
D1
2
k
C
D2
f
D2
k
+ additional term, (3.10)
where P
k
= u
i
u
j
(U
i
/x
j
) is the production rate of k. The turbulent diu-
sion term T
=
_
x
j
_
x
j
_
: at a two-equation level,
x
j
_
C
s
f
R
k
u
i
u
j
x
i
_
: at a second-order closure level.
(3.11)
The T
and
equation models.
Nagano-Kim(1988) Nagano-Tagawa-Tsuji(1991)
C
0.11 0.1
C
s
C
P1
0.9 0.85
C
P2
0.72 0.64
C
D1
1.1 1.0
C
D2
0.8 0.9
1.0 1.0
m
(k/ )(2
R)
1/2
(k/)
_
(2R)
2
+ 3.4(2R)
1/2
/R
3/4
t
_
f
[1 exp(y
+
/A
)]
2
[1 exp(y
+
/A
)]
2
A
(30.5/
Pr)(C
f
/2St) 26/
Pr
f
R
f
P1
1.0 1.0
f
P2
1.0 1.0
f
D1
1.0 [1 exp(y
+
/5.8)]
2
f
D2
1.0 (1/C
D2
)(C
2
f
1) [1 exp(y
+
/6)]
2
Additional
term
t
(1 f
)(
2
/x
j
x
k
)
2
R = (k
)/(k/ ) C
2
= 1.9
f
= 1 exp[(R
t
/6.5)
2
]
In (3.10), quantities and
, (3.13)
where = 2(
k/y)
2
,
= 2(
/y)
2
, and k
= k
k
w
.
3.3.2 Assessment procedure
In (3.9) and (3.10),
and
and
0.1
C
s
0.3
C
P1
0.8 1.9
C
P2
0.3 0.6
C
D1
1.0 2.0
C
D2
0.3 0.9
1.6
m
(k/)
_
f
R
+ [3(2R)
1/2
/(R
3/4
t
Pr)]f
d
_
f
[1 exp(y
/A
)] [1 exp(y
/A
)]
A
14
A
Pr
f
R
2R/(0.7 + R) 2R/(0.5 + R)
f
P1
1.0 [1 exp(y
)]
2
f
P2
1.0 1.0
f
D1
1.0 [1 exp(y
)]
2
f
D2
1.0 (1/C
D2
)(C
2
f
1) [1 exp(y
/5.7)]
2
Additional
term
2C
w2
(k
)v
2
(
2
/y
2
)
2
/k
C
w2
= max[0.1, 0.35 0.21Pr] f
d
= exp[(R
t
/200)
2
]
C
2
= 1.9
f
= 1 exp[(R
t
/6.5)
2
]
3.3.3 Models for assessment
We assess the six temperature dissipation rate equations proposed by Nagano
and Kim (NK) (1988), Hattori, Nagano and Tagawa (HNT) (1993) and Shika-
zono and Kasagi (SK) (1996), which are the
to prevent divergence in
the calculation caused by nite values of and
term is modelled
using
t
, but the Reynolds shear stress u
i
u
j
and turbulent diusion term T
are modelled at a second-order closure level [see (3.11)]. The AKN model puts
f
P1
= f
D1
to avoid divergence in the calculation of ows with complete dissim-
ilarity between velocity and thermal elds. In the SK model, where the k
model is employed to calculate the time scale of the thermal eld, the tur-
198 Nagano
Table 2: (continued)
Sommer-So-Lai(1992) Hattori-Nagano-Tagawa(1993)
C
0.11 0.1
C
s
0.11
C
P1
0.9 0.85
C
P2
0.72 0.64
C
D1
1.1 1.0
C
D2
0.8 0.9
1.0
m
(k/)
_
(2R)
1/2
+
_
0.1(2R)
1/2
/R
1/4
t
_
(k/)
_
(2R)
1/2
+ [7.9(2R)
1/2
/R
3/4
t
]f
d
_
(f
t
/f
)
_
f
_
1 exp
_
y
+
/A
_
2
_
1 exp
_
y
+
/A
_ _
1 exp
_
y
+
/A
_
A
30/(1 + 11.8P
+
)
A
30 A
/Pr
1/3
f
R
1.0
f
P1
1.0 1.0
f
P2
1.0 1.0
f
D1
1.0
f
D2
/ (1/C
D2
)(C
2
f
1)
Additional
term
f
t
[(C
D1
2)(
/k
+ C
D2
( /k)
t
(1 f
w
)(
2
/x
j
x
k
)
2
(
2
/(2k
) + (1 C
P1
)(
/k
)P
f
t
= exp[(R
t
/80)
2
] f
d
= exp[(R
t
/120)
2
]
2(k
/y
2
) f
w
= [1 exp{y
+
/(30/Pr
1/3
)}]
2
P
= u(/x) P
+
= (dP/dx)/u
3
C
2
= 1.9
f
= 1 0.3 exp[R
2
t
]
bulent diusion and production terms are modelled at a second-order closure
level.
A characteristic time scale
m
, whose importance was demonstrated by
Nagano and Kim (1988), has been used in all the two-equation heat-transfer
models for wall shear ows. It can be shown that the eddy diusivity for heat
t
is governed near the wall by the Kolmogorov microscale in the NTT, HNT
and AKN models, and by the Taylor microscale in the SSL model.
3.3.4 Assessment results
The results of assessment for
- and
-equation modelling
is obtained from
. The resultant
characteristic time scale
and
equations: (a)
equations (
)
at a two-equation level; (b)
- and
-
and
-equation models indicate that none of the four models can reproduce
accurately the DNS behavior. Especially, predicted
: (a) in
and
equations at a
second-order closure level.
constant stress and constant heat ux layer, namely
2
/Pr
t
_
C
+
C
P1
C
D1
R
+C
P2
C
D2
= 0, (3.14)
where = 0.39 0.41, Pr
t
= 0.9, C
without
w
uctuations can be inferred from a Taylor series expansion in
terms of y as follows (Youssef et al. 1992):
= h
1
+ 4h
2
y +O(y
2
), (3.15)
where the coecients h
1
and h
2
are independent of the y coordinate. On the
other hand, from (3.4), the molecular diusion term balances the dissipation
term at y = 0:
2
k
y
2
with
2
k
y
2
= h
1
+ 6h
2
y +O(y
2
). (3.16)
From (3.15) and (3.16), the coecient h
2
[= (1/4)(
/y)[
w
] should be zero.
This can be seen, of course, in the DNS data.
In theory, the wall-limiting behavior of
must be
y
2
. In the
equation
of the NK and HNT models, however, the molecular diusion term balances
with C
D1
2
/k
becomes
y
1
. Therefore, in the
and
equation is represented by (
2
/y
2
) = (
2
/y
2
) +
2
2
(
2
[(
/y)
2
]/y
2
). Obviously, the two-equation model predictions
are not in agreement with the DNS data. As seen from Figure 3(c), the sum
total of the budget in the SK model is the closest to the DNS. This is a
consequence of the smaller model constants C
P2
and C
D2
used, which render
the production and destruction terms smaller in magnitude. In the
-equation
models, the NTT model is rather close to the DNS. This is because the NTT
model has no additional production term.
From these assessments it becomes clear that solutions for
are signi-
cantly inuenced by any additional production term, the values ascribed to
the model constants, and the formulation of the characteristic time scale.
3.4 Construction of a Rigorous k
Model
3.4.1 Modelling the eddy diusivity for heat
t
Thermal eddy diusivity
t
given by (3.3) must be adequately modelled with
the dominant characteristic velocity and time scales responsible for scalar
transfer. Thus, it is important to reect the inuence of the time scales for
both velocity and thermal elds. Previous
t
models have been based on the
concept of a single time scale, e.g., the assumption of the turbulent Prandtl
202 Nagano
Figure 3: Budgets of modelled
equations (P
1
+P
2
+P
3
+P
4
+T
)
and
equations (P
1
+P
2
+P
3
+P
4
+T
- and
equations.
number Pr
t
or of a mixed time scale, e.g.,
m
=
or
m
=
2
/
u
. How-
ever, as is frequently pointed out, the former fails to predict heat transfer in
ows with a dissimilarity between velocity and thermal elds, while the latter
compromises the accuracy of the predicted near-wall turbulence quantities be-
cause it relates
u
and
m
using the hybrid time scale
h
=
u
R/(C
m
+R) (i.e., 1/
h
= 1/
u
+C
m
/
with C
m
as a model constant):
m
=
k
_
2R
0.5 +R
+
2R
Pr
3
R
3/4
t
exp
_
_
R
t
200
_
2
__
, (3.17)
where R
t
= k
2
/() is the turbulent Reynolds number and Pr is the molecular
Prandtl number.
Using the multiple-time-scale similar to (3.17), we adopt the following rep-
resentation for
t
to satisfy the wall-limiting behavior of thermal turbulence
indicated by (3.8):
t
= C
k
m
= C
k
_
k
_
2R
0.5 +R
+
2R
Pr
4/3
26
R
3/4
t
exp
_
R
h
220
_
__
,
(3.18)
f
= [1 f
w
(A
)]
1/2
[1 f
w
(A
)]
1/2
(3.19)
where R
h
= k
h
/ = R
t
[2R/(0.5+R)] is the turbulent Reynolds number based
on the harmonic-averaged time scale
h
= (k/)[2R/(0.5 + R)]. Note that
h
becomes identical to
u
= k/ in local equilibrium ows with R = 0.5.
3.4.2 Modelling the
equation
As shown in the previous section, none of the existing
and
models at a
two-equation level give qualitative and quantitative agreement with the DNS.
Hence, we will construct an
equation.
(a) Modelling of P
1
, P
2
, P
4
and
The P
1
, P
2
, P
4
and
+P
2
+P
4
= C
P1
f
P1
u
j
x
j
C
D1
f
D1
2
k
C
P2
f
P2
k
u
i
u
j
U
i
x
j
C
D2
f
D2
k
.
(3.20)
The DNS data indicates that the P
1
and P
2
and
and P
2
terms
using an order-of-magnitude analysis, as done in modelling by Rodi and
Mansour (1993) and Nagano and Shimada (1995a). With k
and
=
k
, P
2
and P
4
terms as
P
1
= O
__
k
__
_
G
_
,
P
2
= O
__
kk
_
S
_
,
P
4
= O
__
kk
2
__
__
,
_
_
(3.21)
where G = [(/x
j
)(/x
j
)]
1/2
represents the mean temperature gradi-
ent, S = [(U
i
/x
j
)(U
i
/x
j
)]
1/2
is the mean strain rate, and =
_
k/
and
=
_
k
/P
4
)G =
G/(
/)
1/2
, P
2
/P
4
(/
k)S = S/(/)
1/2
. Consequently, we dene the
parameters R
T
and R
U
as
R
T
=
_
(/x
j
)
2
_
1/2
_
(/x
j
)
2
_
1/2
=
G
(
/)
1/2
, (3.22)
R
U
=
_
(U
i
/x
j
)
2
_
1/2
_
(u
i
/x
j
)
2
_
1/2
=
S
(/)
1/2
. (3.23)
These parameters represent the ratio of the gradient of mean ow to that of
uctuating components. Apparently, the relations P
1
/P
4
R
T
and P
2
/P
4
R
U
hold. Since the structure of turbulent shear ows near the wall is gov-
erned mainly by the gradient of the mean ow (see, e.g., Hinze (1975)),
contributions of P
1
and P
2
/( +
t
)] and the constant stress layer
[S = (
w
/)/( +
t
) = u
2
/( +
t
)]. Then, (3.22) and (3.23) lead to
R
T
=
u
(/Pr +
t
) (Pr
/)
1/2
f
w
(6), (3.24)
R
U
=
u
2
( +
t
)(/)
1/2
f
w
(6). (3.25)
[6] Modelling heat transfer in near-wall ows 205
The contributions of P
1
and P
2
P1
)f
p
,
f
P1
= exp(7 10
5
R
T
10
)[1 exp(2.2R
T
1/2
)],
_
(3.26)
f
P2
= (1 f
P2
)f
p
,
f
P2
= exp(7 10
5
R
U
10
)[1 exp(2.2R
U
1/2
)],
_
(3.27)
where f
p
is introduced for correcting overproduction near the wall, and the
wall reection function f
w
(6) in (3.24) and (3.25) is given by (2.9) with = 6.
In the following equations (3.28) and (3.29), f
w
(12) and f
w
(3) are similarly
dened.
(b) Modelling of P
3
The P
3
, P
2
, P
4
and
.
However, when compared with the sum of these terms, i.e., P
1
+P
2
+P
4
,
the P
3
is also important.
In the present model, we adopt the following form similar to (2.8) in the
k- model:
P
3
=
t
f
w
(12)
_
2
x
j
x
k
_
2
+C
P3
f
R
k
x
k
x
j
x
j
x
k
, (3.28)
where f
R
= 2R/(0.5 + R). It should be noted that the hybrid turbulent
Reynolds number R
h
and the corresponding time scale
h
in (3.18) can be
written as R
h
= f
R
R
t
and
h
= f
R
u
.
(c) Modelling of T
2(/k)k
, we write T
as
T
=
x
j
_
x
j
_
+C
x
j
_
[1 f
w
(3)]
3/2
k
k
x
j
f
w
(3)
_
. (3.29)
(d) Modelled
-equation
To sum up, the proposed
Dt
=
x
j
__
+
t
x
j
_
C
P1
f
P1
u
j
x
j
C
P2
f
P2
k
u
i
u
j
U
i
x
j
C
D1
f
D1
C
D2
f
D2
k
+
t
f
w
(12)
_
2
x
j
x
k
_
2
206 Nagano
+C
P3
f
R
k
x
k
x
j
x
j
x
k
+C
x
j
_
[1 f
w
(3)]
3/2
k
k
x
j
f
w
(3)
_
. (3.30)
The wall reection function f
w
() is given by (2.9).
(e) Model functions and constants
From (3.30), the molecular diusion term balances the dissipation terms at
y = 0:
y
2
= C
D1
f
D1
+C
D2
f
D2
k
. (3.31)
Considering the limiting behavior of wall turbulence, f
D2
y
2
and f
D1
y
2
(without
w
uctuations) or f
D1
y
n
where n > 0 (with
w
uctuations) are
required to satisfy (3.31). In free turbulence, as described next (see (3.41)),
the limiting behavior requires
C
D2
f
D2
= C
2
f
1. (3.32)
In the present model, the following equations are thus proposed to meet the
requirements for both wall and free turbulence
f
D1
= 1 exp
_
(y
/12)
2
_
= 1 f
w
(12) (3.33)
f
D2
= (1/C
D2
)(C
2
f
1) [1 f
w
(12)] (3.34)
with f
= 1 0.3 exp[(R
t
/6.5)
2
].
The constants appearing in the present two-equation heat-transfer model
are determined as follows. Firstly, we specify a value of C
in (3.18) dening
the eddy diusivity for heat,
t
. In the log-law region where the molecular
diusion is negligible, i.e. f
= f
= 1, C
= C
/ (Pr
t
f
R
) with f
R
= 2R/(0.5 +R) (3.35)
thus, substituting the typical values of C
= 0.10.
We determine the constants C
D1
and C
D2
in the equation for
, (3.30)
from the decay law of homogeneous turbulence. In a homogeneous decaying
turbulent ow, (2.5), (2.6), (3.4) and (3.30) become simply
U
dk
dx
= , (3.36)
U
d
dx
= C
2
f
2
k
, (3.37)
U
dk
dx
=
, (3.38)
U
d
dx
= C
D1
f
D1
C
D2
f
D2
k
, (3.39)
[6] Modelling heat transfer in near-wall ows 207
where the x axis is taken in the ow direction. On the other hand, it is known
that the time-scale ratio R = (k
dx
=
1
R
_
2
k
k
2
C
2
f
2
k
k
2
k
_
=
(C
2
f
1)
k
. (3.40)
Equations (3.39) and (3.40) give the following relations
C
D1
f
D1
= 1,
C
D2
f
D2
= C
2
f
1. (3.41)
Equation (3.41) is also valid for the initial period (f
= f
D1
= f
D2
= 1) in
decaying turbulent ows, and hence we have C
D1
= 1 and C
D2
= C
2
1 = 0.9.
The model constants C
P1
and C
P2
for the production terms in the
-
equation (3.30) are determined by considering the characteristics of the log-
law region (constant stress-heat-ux layer) in wall turbulence. In this region,
the convection terms in the transport equations k, , k
, and
can all be
ignored, and the production terms for k and k
y
_
k
2
f
R
y
_
C
P1
y
C
P2
k
uv
U
y
C
D1
C
D2
k
= 0.
(3.42)
With the above-mentioned characteristics of constant stress-heat-ux layer,
the following relation is obtained from (3.42)
C
P2
= (C
D1
C
P1
)/R +C
D2
(
2
/Pr
t
)/(
C
1/2
), (3.43)
where is the von K armam constant. Equation (3.43) is similar to the well-
known relation in the k- model given by
C
1
= C
2
2
/(
C
1/2
). (3.44)
The value C
P2
= 0.77 is then obtained if we substitute the foregoing values
of C
D1
, C
D2
, R, Pr
t
, and C
-equation at a two-
equation level are listed in Table 3.
208 Nagano
Table 3: Model constants and functions in the present
models.
C
0.1
C
s
C
P1
0.9
C
P2
0.77
C
P3
0.05
C
D1
1.0
C
D2
0.9
C
1.6
1.8
m
(k/)
_
f
R
+ [26(2R)
1/2
/(Pr
4/3
R
3/4
t
)]f
d
_
f
[1 f
w
(A
)]
1/2
[1 f
w
(A
)]
1/2
f
w
() exp
_
(y
/)
2
_
A
28
A
/Pr
1/3
f
R
2R/(0.5 +R)
f
P1
(1 f
P1
)f
p
[1 f
w
(12)]
f
P2
(1 f
P2
)f
p
f
D1
1 f
w
(12)
f
D2
(1/C
D2
)(C
2
f
1)[1 f
w
(12)]
Additional term (= P
3
)
f
d
exp(R
h
/220)
C
2
1.9
f
1 0.3 exp[(R
t
/6.5)
2
]
f
P1
exp(7 10
5
R
10
T
)[1 exp(2.2R
1/2
T
)]
f
P2
exp(7 10
5
R
10
U
)[1 exp(2.2R
1/2
U
)]
f
p
1 + 0.75 exp[(R
h
/40)
1/2
]
R
T
f
w
(6)u
/[( +
t
)(
/)
1/2
]
R
U
f
w
(6)u
2
/[( +
t
)(/)
1/2
]
(f ) Second-order closure modelling
In second-order closure modelling, the turbulent diusion term T
and the
production term P
3
+v
(
/Pr)
1/2
f
w
(6), (3.45)
[6] Modelling heat transfer in near-wall ows 209
Figure 4: Assessment of the proposed
-equation models (P
1
+P
2
+P
3
+P
4
+T
).
R
U
=
u
2
+uv
()
1/2
f
w
(6). (3.46)
The model functions f
P1
and f
P2
in (3.26) and (3.27) are dened by
f
P1
= exp(7 10
5
R
T
10
)[1 exp(1.1R
T
1/2
)],
f
P2
= exp(7 10
5
R
U
10
)[1 exp(1.1R
U
1/2
)].
_
(3.47)
210 Nagano
Figure 5: Budget of temperature variance k
.
The turbulent diusion term, T
, can be written as
T
=
x
k
_
C
s
k
f
R
u
j
u
k
x
j
_
+C
x
j
_
[1 f
w
(3)]
3/2
k
k
x
j
f
w
(3)
_
,
(3.48)
where C
s
= 0.11, and f
R
and C
= C
P4
u
j
u
k
k
f
R
x
k
x
x
j
+C
P3
f
R
u
j
u
x
k
x
j
x
k
, (3.49)
with C
P3
= 0.1 and C
P4
= 0.25.
(g) Assessment of proposed
equation models
Figure 4 shows the solutions obtained from new
equation
Figure 5 shows the budget of temperature variance predicted by the NTT
model (Nagano et al. 1991), which is the basis for the proposed model and the
[6] Modelling heat transfer in near-wall ows 211
Table 4: Model constants and functions in the present k
model.
C
0.1
h
1.8/[1 + 0.5f
w
(28)]
AKN model (Abe et al. 1995). Obviously, the model predictions are dierent
from DNS data near the wall because of the solution given by the
equation
and the modelling of the turbulent diusion term in the k
equation. Therefore,
in the k
that
should be modelled. We adopt the foregoing turbulent diusion modelling, and
write T
k
as
T
k
=
x
j
_
h
k
x
j
_
+C
x
j
_
u
k
u
d
k
n
e
j
[1 f
w
(28)]
1/2
k k
[f
w
(28)]
1/2
_
,
(3.50)
where d
k
, n
and e
j
are unit vectors in the streamwise, wall-normal and x
j
directions, respectively, and
u
k
u
x
=
_
1 (x 0),
1 (x < 0).
(3.51)
The nal formulation of the k
Dt
=
x
j
__
+
t
h
_
k
x
j
_
+C
x
j
_
u
k
u
d
k
n
e
j
[1 f
w
(28)]
1/2
k k
[f
w
(28)]
1/2
_
+P
k
. (3.52)
The model functions and constants in the present k
k/y)
2
,
w
= 2(
/y)
2
and
w
or q
w
are determined by experimental or DNS
data at a wall, U/y = k/y = /y = /y = k
/y =
/y = 0
at the axis for internal ows (symmetry); U = U
e
, k = = k
= 0
and =
e
at the outer edge of the boundary layer, where U
e
and
e
are
prescribed from experiments.
The criterion for convergence is
max [X
(i+1)
X
(i)
/X
(i)
[ < 10
5
, (3.54)
where X = U, k, , , k
, and
= 150)
shown in Figure 6. From Figure 6, it can be seen that the mean velocity and
turbulent energy are predicted quite successfully for both cases.
Next, we assess the constructed two-equation heat-transfer model with the
k- model in a fully developed channel ow under both constant-temperature
[6] Modelling heat transfer in near-wall ows 213
Figure 6: Channel ow predictions: (a) mean velocity; (b) turbulent energy.
(Kim and Moin 1989) (Re
, in the k
in channel ow.
al. 1992), and the present predictions also indicate good agreement with the
experimental data.
3.5.4 Constant wall temperature followed by adiabatic wall
The next test case for which calculations have been performed is concerned
with a more complex thermal eld in a boundary layer along a uniformly
heated wall followed by an adiabatic wall. Figure 11 shows a comparison of
216 Nagano
Figure 10: (a) mean temperature; (b) turbulent heat ux; (c) rms temperature.
the predicted results with the experimental data (Reynolds et al. 1958) of
temperature dierences between the wall and the free-stream (=
e
w
).
It can be seen that the proposed model gives generally good predictions for the
rapidly changing thermal eld. Also by comparison, the present model gives
no prediction inferior to the AKN model (Abe et al. 1995).
[6] Modelling heat transfer in near-wall ows 217
Figure 11: Comparison of the predicted variations of wall temperature with
the measurement.
Figure 12: Comparison of the predicted rms temperature proles and mea-
surements (sudden decrease in wall heat ux).
3.5.5 Constant heat ux followed by adiabatic wall
To further verify the eectiveness of the present model for calculating vari-
ous kinds of turbulent thermal elds, we have carried out the calculation of a
boundary layer ow along a uniform heat-ux wall followed by an adiabatic
wall, which has been reported in detail by Subramanian and Antonia (1981).
The calculated distributions of rms temperature uctuations normalized by
temperature dierence between the free stream and the wall,
c
, at a step
change in surface thermal condition, are shown in Figure 12, compared with
the experimental data (Subramanian and Antonia 1981) and the prediction
of the AKN model. Both models indicate a slight underprediction of the peak
value of rms temperature. The proposed model, however, shows the variation
of physical phenomena of rms temperature in the thermal layer along a uni-
form heat-ux followed by an adiabatic wall. In particular, the rapid decrease
218 Nagano
Figure 13: Comparison of the predicted variations of wall temperature and
Stanton number with the measurements (double-pulse heat input)
in temperature uctuations from the inner region has been captured by the
proposed model.
3.5.6 Double-pulse heat input
As a nal test case, we have calculated the more complex thermal case, where
the heat input is spatially intermittent in a double-pulse manner. Then we
have investigated the mechanism of turbulent heat transfer in such a rapidly
changing thermal layer.
The temperature dierence between the free stream and the wall, =
w
e
, and the Stanton number reported by Reynolds et al. (1958), are
shown in Figure 13 compared with the prediction of the present model. It is
indicated that both the velocity and the thermal elds are well predicted, and
the turbulent heat transfer characteristics in the thermal entrance region are
reproduced very well. Figure 14(a) shows how the turbulent near-wall thermal
layer changes when the heat input is intermittent, where =
e
is
normalized by the temperature dierence between the wall and the free stream,
c
=
wc
e
, just before the rst heat input/cuto point. It can be seen
that a very abrupt decrease and increase in mean uid temperature occurs in
the wall region, which is a consequence of the no-heat-input condition followed
by heat input, i.e., /y[
w
= 0 /y[
w
= constant. Within a short
distance from the abrupt change-over, the mean temperature prole becomes
uniform over most of the thermal layer. The following discussion deals with
how these phenomena aect other turbulent quantities. Figure 14(b) shows the
distribution of turbulent heat ux normalized by u
and
c
. Just after the
rst heat input/cuto point, with vanishing mean temperature gradient near
the wall, the turbulent heat ux, v, decreases rapidly. Just before the second
heat input point, v has greatly decayed with its maximum occurring in the
outer layer. Over the reheated wall, v again shows a rapid increase near the
[6] Modelling heat transfer in near-wall ows 219
Figure 14: Variations of turbulent quantities for double-pulse heat input: (a)
mean temperature; (b) turbulent heat ux; (c) rms temperature.
wall. This is qualitatively consistent with the experimental result (Antonia et
al. 1977) for the thermal entrance region of the boundary layer on a at plate.
Next, variations of the rms temperature are shown in Figure 14(c). Just
after the heat cuto point, distributions of the rms temperature tend to be
similar to the experimental evidence obtained by Subramanian and Antonia
220 Nagano
Figure 15: Budget of temperature variance for double-pulse heat input: (a)
x = 0.517 [m]; (b) x = 0.579 [m]; (c) x = 0.876 [m]; (d) x = 0.936 [m].
(1981) and discussed in the previous section. It can be seen that just before
the second heat input point, the rms temperature remains in the outer region
only, and that it increases very rapidly near the wall beyond that point. Fig-
ures 15(a)(d) show budgets of temperature variance at locations just before
the rst heat cuto (x = 0.517 m), just after the heat cuto (x = 0.579 m),
just before the second heat input (x = 0.876 m), and just after the second heat
input (x = 0.936 m), respectively. In these gures, each term is normalized
by the peak value of the production P
k
Model
4.1.1 Eddy diusivity
t
for k
model
As mentioned in the foregoing, the eddy diusivity for heat,
t
, is generally
given by (3.3). Features of
t
in (3.3) with (3.17) are summarized as follows:
in the near-wall region,
t
yields the relation
t
k
_
R/Pr
t
_
R/Pr,
so it is possible to adequately capture the behavior of dissipative motions and,
in the region far from the wall, the
t
model consists only of time scales of the
energy-containing eddies through the hybrid time scale
h
(see section 3.4.1).
In the present k
m
=
k
_
2R
C
m
+R
+
_
2R
Pr
B
R
t
f
_
, (4.1)
where C
m
is the weighting constant of the composite time scale, B
is the
model constant that represents the eectiveness of dissipation eddies, and f
-aected region.
As for the model function f
= 1 exp
_
A
n
y
2n
_
= 1 exp
_
A
y
2
_
,
where
A
= A
(1 +C
Pr)
n
. (4.2)
Here, A
and C
= y/
m
= (1 +C
Pr) y
,
m
=
_
1
+
C
_
1
=
1 +C
Pr
.
(4.3)
222 Nagano
The length scale = (
3
/ )
1/4
represents the Kolmogorov microscale dened
previously, and
= (
2
/ )
1/4
is the Batchelor microscale (Batchelor 1959).
y
is dened as y
becomes
dominant. [For low Prandtl number uids, the so-called Obukhov microscale
(Obukhov 1949)
o
= (
3
/ )
1/4
= Pr
3/4
could be used as the characteristic
length scale of dissipation eddies. However, it is easily understood that is
always smaller than
o
because of the relation
o
/ = Pr
3/4
.] It should also
be noted that f
and y
in (4.2) is
2 results from a restriction in the wall-limiting behavior of
t
(Youssef et al.
1992; Abe et al. 1995). We determine the power n, the constant A
, and the
weighting constant C
= A
(1 + C
Pr)
n
so that the maximum values of k
at Pr = 0.1, 0.71
and 2.0 agree with those of the DNS data, while the remaining constants and
functions are xed. As a result, we obtain n = 1/4, A
= 710
4
, and C
= 2.
The weighting constant C
m
plays an important role in giving weight to
a shorter time scale between
u
and
before determining C
m
. It is clear that, in decaying homo-
geneous ows with high Prandtl number uids, the time-scale ratio R(=
/
u
)
is greater than unity in the high R
t
region (Iida and Kasagi 1993). Also, it is
easily veried that for turbulent wall ows, R strictly equals Pr at the wall.
(Note that R = Pr is ensured for the case of
w
= 0). Hence, the characteris-
tic time scale suitable for high Prandtl number uids is
u
. In the case of low
Prandtl number uids, on the other hand, since the DNS data (Iida and Kasagi
1993; Kasagi et al. 1992; Kasagi and Ohtsubo 1992) indicate that R is always
smaller than unity irrespective of the types of ow elds,
is appropriate for
the characteristic time scale. When Pr 1 (e.g., air ow), the wavenumbers
related to peak intensities of both velocity and temperature uctuations are
close to each other, so that the inuence of both
u
and
becomes signi-
cant. For the reasons mentioned above, the weighting constant C
m
would be
expected to change with Pr, and hence we decide that C
m
= 0.2/Pr
1/4
.
As for the model function f
exerts
a signicant inuence on the prediction of high Prandtl number uids in which
the dissipation scale becomes much smaller. Thus, we write the model function
f
as f
= exp[( y
/25)
3/4
].
Finally, we consider the model constant B
t
independent of wall thermal conditions, and to reproduce the fact that the
[6] Modelling heat transfer in near-wall ows 223
ratio of the respective time scales for dissipation eddies in the velocity and
thermal elds becomes equal to
_
R/Pr (Youssef et al. 1992; Abe et al. 1995),
B
Pr)
1/4
. Here, the value of 120 is chosen so
that the wall-limiting behavior of the calculated v agrees with that of the
DNS. As a result, the near-wall behavior of
t
is proportional to
t
_
R/Pr,
similar to that given by Abe et al. (1995).
To sum up, the proposed
t
model is written as follows:
t
= C
k
2
_
2R
0.2/Pr
1/4
+R
+
_
2R
Pr
120
(1 + 2
Pr)
1/4
1
R
t
exp
_
_
y
25
_
3/4
__
,
f
= 1 exp(7 10
4
y
1/4
y
7/4
), (4.4)
y
= (1 + 2
Pr) y
.
The model constant C
-equation
The k
.
The turbulent diusion term T
k
-equation
and produces an incorrect behavior of
=
x
__
h
_
k
_
+C
k k
f
w
u
i
u
j
e
Si
e
Nj
e
_
,
h
=
h0
/f
h
,
(4.5)
where
h0
, C
, f
h
, and f
w
are the model constants and functions, respectively.
(Note that, as already mentioned, the sign function
u
i
u
j
and the unit vectors
224 Nagano
Table 5: Model constants and functions of the proposed k
model.
C
C
m
h
C
C
P1
C
P2
C
D1
C
D2
f
D2
0.10
0.2
Pr
1/4
1.6
f
h
0.1 1.8 0.825 0.9 1.0 (4.15)
C
D3
C
f
h
f
w
f
P1
f
P2
f
D1
f
D3
0.025 3.0 (4.2) (4.6) (4.7) 1.0 1.0 1.0 f
w
e
Si
, e
Nj
, e
R
t
100
__
, (4.6)
f
w
= (1 f
w
)
2
f
1/4
w
with f
w
= exp( y
/8) , (4.7)
where f
w
is the wall reection function similar to (2.9),
R
t
= k
2
/( ) is the
turbulent Reynolds number based on , and (C
m
+R)/2R =
u
/(2
h
) in (4.6)
is necessary for f
h
to ensure the balance in the order of magnitude between the
modelled turbulent diusion term through GGDH [(
t
/
h
k
/x
j
)/x
j
O(
t
/
h
2
/
2
), where ( )
O(u
2
/)] in the
k
-
equation model is written as follows:
Dk
Dt
=
x
__
+
t
h
_
k
+C
k k
f
w
u
i
u
j
e
Si
e
Nj
e
_
_
+P
k
.
(4.8)
One may note that (4.8) is almost identical to (3.52).
4.1.3 Modelling the
-equation
In the two-equation modelling of a thermal eld, the dissipation rate
of
temperature variance must be determined from its transport equation. The use
of the
Dt
=
x
j
__
+
t
x
j
_
+
(C
P1
f
P1
P
k
C
D1
f
D1
)
+
k
(C
P2
f
P2
P
k
C
D2
f
D2
) +
t
(1 f
)
_
2
x
j
x
k
_
2
, (4.9)
[6] Modelling heat transfer in near-wall ows 225
where
, C
P1
, C
P2
, C
D1
, C
D2
are model constants, and f
P1
, f
P2
, f
D1
, f
D2
are model functions. In the present study, these model constants and functions
are basically identical to those adopted by previous k
and
is dened by (3.13).
In the proposed model, we use the above
(time scale)
1
. Thus, we classify terms in the transport equation into the
following three groups:
(i) Comparatively slow motion due to the existence of mean velocity and
temperature gradients (contributing to the generation of turbulence
quantities);
(ii) Relatively rapid motion due to the eect of large (energy-containing)
eddies;
(iii) remarkably rapid motion caused by dissipation eddies.
First, we consider the time scale dominating the motion (i). After inves-
tigating the coherent structure of wall turbulence using the DNS database
(Robinson 1991), it was claried that the dissipation reaches its maximum in
the internal shear layer (or near-wall shear layer (Robinson 1991)) occurring
near the wall. This means that a close relationship exists between the produc-
tion process of (or
are expected
to be represented through those related to mean velocity and thermal elds,
e.g., mean velocity gradient
U
= 1/(U/y).
Put another way, the shear rate parameter S
_
2S
ij
S
ij
, where S
ij
(U
i
/x
j
+ U
j
/x
i
)/2 denotes the mean velocity gradient (strain) tensor,
seems to be suitable as the characteristic time scale to be considered. However,
since turbulence must be generated by the production terms P
k
and P
k
in
the k- and k
as
U
=
1
P
k
/k
=
_
u
i
u
j
k
U
i
x
j
_
1
,
=
1
P
k
/k
=
_
u
j
x
j
_
1
.
(4.10)
It is interesting to note that (4.10) is clearly identied with the time scales
of the production processes previously introduced by Nagano and Kim (1988)
[see (4.9)].
Next, let us examine uid motion (ii). It is already well known that the
motion of energy-containing eddies makes little contribution to generating
turbulence but rather acts to supply energy to smaller eddies through the
energy-cascade process. The characteristic time scales of the energy-containing
eddies for both velocity and thermal elds are closely connected with those
of eddies with wavenumbers related to the maximum values of the spectral
distributions for k and k
u
= k/,
= k
. (4.11)
Note that the eects of (4.11) are already explicitly included in (4.9).
Finally, we discuss the motion (iii) with rapid change. It is known that
almost all the energy fed by the mean eld is accumulated in the energy-
containing eddies and, subsequently, that it is successively supplied to smaller
eddies (or dissipation eddies) through the deformation work of larger eddies.
The dissipation eddies have vorticity much higher than that of the energy-
containing eddies, so that the characteristic time scale of the former eddies
becomes shorter compared with that of the latter. Almost all the modelled
-
or
u
=
_
/. (4.12)
In a similar way, the representative time scale of the dissipation eddies in a
thermal eld can be dened as follows:
=
_
/
_
k
/k =
_
R/Pr
u
. (4.13)
[6] Modelling heat transfer in near-wall ows 227
It should be noted that the Taylor microscales =
_
k/ and
=
_
k
for velocity and thermal elds, which are often used as the scales of the smallest
eddies, are strongly linked to the characteristic time scales of dissipation eddies
as =
k
u
and
.
In the present study, as mentioned previously, the characteristic time scales
of dissipation eddies [(4.12) and (4.13)] are written through the following com-
posite time scale
m
:
1
m
=
1
u
+
C
=
1
u
_
1 +C
_
Pr
R
_
. (4.14)
where C
D2
f
D2
_
1 +C
D3
f
D3
_
R
t
_
1 +C
_
Pr
R
__
, (4.15)
with
C
D2
f
D2
= (C
2
f
2
1)
_
1 exp
_
y
2
_
. (4.16)
Here, C
D3
, C
, and f
D3
are C
D3
= 0.025, C
= 3.0, and f
D3
= f
w
, respectively.
[We would like to emphasize that the proposed
/k
y
2
(Kawamura and Kawashima 1994; Shikazono and Kasagi 1996).
4.2 Model Performance in Thermal Fields
In this section, we examine the validity of the proposed k
m
o
d
e
l
f
o
r
v
a
r
i
o
u
s
o
w
e
l
d
s
.
E
E
C
a
s
e
A
B
C
D
(
w
=
c
o
n
s
t
)
(
q
w
=
c
o
n
s
t
)
F
w
c
o
n
s
t
y
_
w
=
q
w
c
p
w
|
s
u
c
o
n
s
u
c
t
i
o
n
s
i
d
e
c
o
n
s
t
c
o
n
s
t
y
_
w
=
q
w
c
p
y
_
w
=
q
w
c
p
w
|
i
n
j
o
n
i
n
j
e
c
t
i
o
n
s
i
d
e
k
w
0
0
0
0
0
0
a
n
d
y
=
0
0
w
0
0
0
0
0
0
0
[6] Modelling heat transfer in near-wall ows 229
Case F. Backward-facing step ow for various Prandtl number uids.
The numerical scheme and the grid system are the same as for the k-
model, as already mentioned. In the analysis of case D, however, 201 grid
points were used for the calculation of relatively low Reynolds and/or Prandtl
number uids, but 1001 ner grid points were used to calculate thermal elds
in high Reynolds and/or Prandtl number uids with reference to the recent
work by Sato et al. (1994). The wall-boundary conditions are collectively listed
in Table 6. (Symmetric boundary conditions are imposed on various thermal
turbulence quantities at the centerline of internal ows, while free-stream con-
ditions are imposed for external ows, Hattori et al. (1993).)
4.2.1 Channel ow with internal heat source
First of all, we conrm the validity of the proposed k
model in channel
ows with internal heat source. Figures 16(a)16(c) show the thermal-eld
turbulence quantities, compared with the DNS data of Kim and Moin (1989)
at Re
= 180. These gures reveal the excellent agreement between the pre-
dictions and the DNS data, and the Prandtl number dependence is adequately
captured. Predicted budget proles in various Prandtl number uids are given
in Figure 17. From these gures, we immediately notice the dierent contribu-
tions of each term in the k
, which is given by
+ 2(
/y)
2
, can
[6] Modelling heat transfer in near-wall ows 231
Figure 17: Budget proles of the k
232 Nagano
Figure 18: Thermal turbulence quantities in air and mercury channel ows with
uniform wall heat ux: (a) mean temperature; (b) temperature uctuation
intensities; (c) turbulent heat ux.
w
[
inj
and turbulent heat ux v
+
normalized by u
and
of each wall,
together with the DNS data (Sumitani and Kasagi 1995). It is readily seen from
the gures that, though the present k
+
, turbulent heat ux v
+
, temperature uctuation
/(
w
0
), and tur-
bulent Prandtl number Pr
t
in pipe ow are presented in Figures 24(a)24(d).
Owing to the lack of experimental data for various turbulence quantities at the
corresponding Reynolds number, simply the results obtained by the proposed
model under varying Pr are discussed. It is readily found from those gures
that there are systematic variations with varying Pr; for example,
+
shown
in Figure 24(a) reveals that the conduction sublayer becomes increasingly thick
236 Nagano
(a) (b)
(c)
(d)
Figure 24: Prandtl number dependence of various thermal turbulence quanti-
ties in pipe ow at Re
m
= 10000: (a) mean temperature; (b) turbulent heat
ux; (c) temperature uctuation; (d) turbulent Prandtl number.
as Pr decreases (e.g., the outer edge of the conduction sublayer is y
+
8 at
Pr = 0.71 while y
+
150 at Pr = 0.025), and for Pr > 0.1, the logarithmic
temperature distributions can clearly be recognized. It is also clear from Fig-
ure 24(b) of v
+
that the peak location of v
+
shifts toward the center of the
pipe with decreasing Pr and that the peak at Pr < 0.1 is around y
+
150 in
close accordance with the peak of temperature uctuation mentioned below. A
look at the temperature uctuation proles [Figure 24(c)] immediately reveals
some distinct features. For example, its attened distributions over the whole
ow region can be seen for Pr 0.1. When Pr increases from 0.1, on the other
hand, the temperature uctuation, reaches its maximum in the vicinity of the
wall. As for the turbulent Prandtl number Pr
t
, some interesting trends can
be seen from Figure 24(d). It is apparent that the predicted Pr
t
at Pr = 0.71
can correctly capture the tendency of the DNS obtained for air channel ow
at a similar Reynolds number. It is also clear that the predicted Pr
t
tends
to vary systematically as Pr decreases: Pr
t
gradually decreases in the near-
wall region (1 < y
+
< 10) and increases in the logarithmic region (y
+
> 40)
from the standard value at Pr = 0.71. This can be recognized by investigating
the molecular and turbulent transport terms of the energy equation. For fully
developed turbulent ows, the energy equation reduces to
[6] Modelling heat transfer in near-wall ows 237
Figure 25: The mean temperature prole at a high Prandtl number uid (Pr =
95, channel ow at Re
m
= 10000).
Figure 26: Turbulent heat and mass transfer at various Prandtl and Schmidt
numbers (pipe ow at Re
m
= 10000).
q
q
w
=
_
1
Pr
+
1
Pr
t
_
d
+
dy
+
. (4.17)
The rst term in brackets on the right-hand side denotes the molecular con-
duction and the second denotes the turbulent diusion. The mean temperature
distribution near the wall can be expressed as
+
= Pry
+
. (4.18)
This equation signies that the heat transport near the wall is dominated by
the molecular conduction, as conrmed by many experiments. In low Prandtl
number uids, it is conrmed from Figures 21 and 24 that the region where the
molecular conduction predominates is extended to the logarithmic region of
the velocity eld. In this region,
t
/ in (4.17) is large, so the turbulent Prandtl
number should become larger to render the eect of the turbulent diusion
negligible. In the viscous sublayer, on the other hand,
t
/ is much smaller
than unity so the value of Pr
t
may be on the order of unity. It should be noted
that the fact that the obtained Pr
t
at any Pr in the immediate vicinity of the
238 Nagano
wall approaches a constant value (about 1.21) is due to the near-wall behavior
of the proposed
t
and
t
models irrespective of the variation of Pr, i.e.,
Pr
t
C
_
2R/Pr
1.21,
as y
+
0 (with R Pr). (4.19)
Next, before evaluating the performance of the proposed model in higher
Prandtl number uids (Pr 1.0), we discuss the predictive accuracy at ex-
tremely high Prandtl numbers. Owing to the diculty of carrying out turbu-
lent heat transfer experiments in higher Prandtl number uids, the measure-
ments reported up to now have been very few, and generally limited to mean
temperature proles; hence, the predictive accuracy of the present model is
veried through a comparison of the prediction with the following measure-
ments of mean temperature proles. The predicted mean temperature prole
in a fully developed channel ow (Re
m
= 10000) of engine oil (Pr = 95) is
shown in Figure 25 with the corresponding experimental data. It can be seen
from the gure that the mean temperature proles for both the prediction
and the experiment show abrupt changes in the region adjacent to the wall
(y
+
< 10). It is also clear that, though the beginning of the predicted almost
constant mean temperature region is displaced slightly away from the wall in
comparison with the measurement, the value obtained at the center of the
channel agrees well with the experimental one. This means that either the
wall friction temperature
/(
w
0
) proles shown in Figures 24(b) and 24(c) also exhibit sys-
tematic variations with increasing Pr: the constant heat ux layer (v
+
1)
emerges in the wall region; and the maximum k
)/(
w
) , turbulent heat ux v
+
, and Stanton
number St = q
w
/[c
p
U
0
(
w
0
)] (where c
p
denotes specic heat at con-
stant pressure) are illustrated in Figures 27 and 28, in comparison with the
experimental data (Vogel and Eaton 1984) of
0
/H = 1.1, where H is the step
240 Nagano
height and
0
is the upstream boundary-layer thickness at the step location.
Note that the predicted results for the corresponding velocity eld are given
in Figure 29. Again, in each gure, X
= (X X
R
)/X
R
. As seen
from the mean temperature proles in Figure 27(a), there is good agreement
between the present prediction and the measurement at various streamwise
locations, though the Pr
t
-constant model gives consistently high values of St.
This indicates that the conventional approach using a Pr
t
-constant model has
serious problems which cannot be overlooked in calculating a thermal eld in
complex ows of industrial interest. From the proles of v in Figure 27(b),
we immediately notice quite dierent characteristics between the recirculation
(X
-equation
through the turbulent diusion model. Thus, the present model can properly
reproduce the same tendency of v as in the experiment, as already known
from the agreement between the predicted and measured proles. Within
the redeveloping region (X
) in Figure 32 in
which the contour value is normalized by the respective maximum tempera-
ture (
max
- and
-equation, all of
the production and destruction terms are modelled to reproduce the correct
behavior of
near the wall. Note that, in order to obtain the correct wall-
limiting behavior of
prole
near the wall. It is also shown that the present model works very well for
calculating the heat transfer under dierent thermal conditions. Furthermore,
the present model reproduces the budget proles of turbulence quantities as
accurately as DNS. Thus, we anticipate a practical application of the present
model in revealing the underlying physics of turbulent heat transfer in complex
ows of technological interest.
In k
model in channel
244 Nagano
ows under arbitrary wall thermal boundary conditions show that the Prandtl
number dependence identied by DNS is satisfactorily captured. The Reynolds
and Prandtl (or Schmidt) number dependence of the proposed k
model
is thoroughly investigated in a pipe ow, and we have demonstrated that
the predicted heat (and mass) transfer rate coincides with reliable empirical
formulas and experimental data with high accuracy. Of course, this agreement
cannot be achieved without a high-performance k- model.
References
Abe, K., Kondoh, T. and Nagano, Y. (1994). A new turbulence model for predict-
ing uid ow and heat transfer in separating and reattaching ows -I. Flow eld
calculations, Int. J. Heat Mass Transfer 37, 139151.
Abe, K., Kondoh, T. and Nagano, Y. (1995). A new turbulence model for predicting
uid ow and heat transfer in separating and reattaching ows -II. Thermal eld
calculations, Int. J. Heat Mass Transfer 38, 14671481.
Antonia, R.A. (1980). Behavior of the turbulent Prandtl number near the wall, Int.
J. Heat Mass Transfer 23, 906908.
Antonia, R.A., Danh, H.Q. and Prabhu, A. (1977). Response of a turbulent boundary
layer to a step change in surface heat ux, J. Fluid Mech. 80, 153177.
Azer, N.Z. and Chao, B.T. (1961). Turbulent heat transfer in liquid metals fully
developed pipe ow with constant wall temperature, Int. J. Heat Mass Transfer
3, 7783.
Batchelor, G.K. (1959). Small-scale variation of convected quantities like temperature
in turbulent uid, Part 1. General discussion and the case of small conductivity,
J. Fluid Mech. 5, 113133.
Bhatti, M.S. and Shah, R.K. (1987). Turbulent and transition ow convective heat
transfer in ducts. In Handbook of Single-Phase Convective Heat Transfer, S. Kaka c
et al. (eds.), Sec. 4.1, Wiley.
Borishansky, V.M., Gelman, L.I., Zablotskaya, T.V., Ivashchenko, N.I. and Kopp, I.Z.
(1964). Investigation of heat transfer in mercury ows in horizontal and vertical
pipes, Convective Heat Transfer in Two- and One Phase Flows, 340. Energiya,
Moscow (Collected Papers).
Cebeci, T. (1973). A model for eddy conductivity and turbulent Prandtl number,
Trans. ASME, J. Heat Transfer 95, 227234.
Chieng, C.C. and Launder, B.E. (1980). On the calculation of turbulent heat trans-
port downstream from an abrupt pipe expansion, Numerical Heat Transfer 3,
189207.
Gibson, M.M., Verriopoulos, C.A. and Nagano, Y. (1982). Measurements in the
heated turbulent boundary layer on a mildly curved convex surface. In Turbu-
lent Shear Flows 3, L.J.S. Bradbury et al. (eds.), Springer, 8089.
Harriott, P. and Hamilton, R.M. (1965). Solid-liquid mass transfer in turbulent pipe
ow, Chem. Eng. Sci. 20, 10731078.
Hattori, H. and Nagano, Y. (1995). Calculation of turbulent ows with pressure
gradients using a k- model, JSME Int. J. II 38, 518524.
Hattori, H. and Nagano, Y. (1998). Rigorous formulation of two-equation heat trans-
fer model of turbulence using direct simulations, Numerical Heat Transfer B 33,
153180.
[6] Modelling heat transfer in near-wall ows 245
Hattori, H., Nagano, Y. and Tagawa, M. (1993). Analysis of turbulent heat transfer
under various thermal conditions with two-equation models. In Engineering Tur-
bulence Modelling and Experiments 2, W. Rodi and F. Martelli (eds.), Elsevier,
4352.
Hinze, J.O. (1975). Turbulence, 2nd ed., McGraw-Hill.
Hochreiter L.E. and Sesonske, A. (1974). Turbulent structure of isothermal and non-
isothermal liquid metal pipe ow, Int. J. Heat Mass Transfer 17, 113123.
Iida, O. and Kasagi, N. (1993). Direct numerical simulation of homogeneous isotropic
turbulence with heat transport (Prandtl number eects), Trans. Jpn. Soc. Mech.
Eng. B 59, 33593364.
Itazu, Y. and Nagano, Y. (1998). RNG modeling of turbulent heat ux and its ap-
plication to wall shear ows, JSME Int. J. B 41, 657665.
Jones, W.P. and Musonge, P. (1988). Closure of the Reynolds stress and scalar ux
equations, Phys. Fluids 31, 35893604.
Kasagi, N. and Ohtsubo, Y. (1992). Direct numerical simulation of low Prandtl num-
ber thermal eld in a turbulent channel ow. In Turbulent Shear Flows 8, F. Durst
et al. (eds.), Springer, 97119.
Kasagi, N., Tomita, Y. and Kuroda, A. (1992). Direct numerical simulation of passive
scalar eld in a turbulent channel ow, Trans. ASME, J. Heat Transfer 114, 598
606.
Kawamura, H. and Kawashima, N. (1994). A proposal of k- model with relevance to
the near wall turbulence. In Proc. Int. Symposium on Turbulence, Heat and Mass
Transfer, Lisbon, Portugal, I.1.1I.1.4.
Kays, W.M. (1994). The 1992 Max Jakob Memorial Award Lecture: Turbulent Prandtl
number where are we? Trans. ASME, J. Heat Transfer 116, 284295.
Kim, J. and Moin, P. (1989). Transport of passive scalars in a turbulent channel
ow. In Turbulent Shear Flows 6, J.-C. Andre et al. (eds.), Springer, 8596.
Kim, J., Moin, P. and Moser, R. (1987). Turbulence statistics in fully developed
channel ow at low Reynolds number, J. Fluid Mech. 177, 133166.
Kline, S.J. (19801981) AFOSRHTTMStanford Conference Complex Turbulent
Flows.
Kondoh, T. Nagano, Y. and Tsuji, T. (1993). Computational study of laminar heat
transfer downstream of a backward-facing step, Int. J. Heat Mass Transfer 36,
577591.
Launder, B.E. (1988). On the computation of convective heat transfer in complex
turbulent ows, Trans. ASME, J. Heat Transfer 110, 11121128.
Moser, R-D, Kim, J. and Mansour, N.N. 1999 Direct numerical simulation of turbu-
lent channel ow up to Re
_
1/4
. (1)
A connection with ow Reynolds number can be made if we make some fur-
ther assumptions. For a ow in equilibrium we may take production equal
to dissipation. The production can be assumed to scale as U
3
/L where U is
a reference bulk velocity and L a length scale of the problem, usually xed
[7] Introduction to direct numerical simulation 251
by the geometry. Both U and L are characteristic of the largest scales of the
turbulence. Thus we can write
L
Re
3/4
(2)
where Re = UL/ is the bulk Reynolds number of the ow. The number of grid
points N that we will require for a given simulation will be proportional to L/
and hence to Re
3/4
. For dierent denitions of Reynolds number appropriate
for dierent ows the exponent changes in the range 0.75 to 0.9 (Reynolds
1989). We can now extend our prediction of increases in N over time from the
previous section to increases in Re over time. A doubling of Reynolds number
will require a factor of 2.2 to 2.5 increase in N at a total cost increase of a
factor of 37 to 67. At the current rate of increase of computer performance,
and assuming that algorithms can be made to continue scaling eciently, this
leads to a potential doubling of Reynolds number every 6 or 7 years.
It is important to appreciate the practical implications of this Reynolds
number scaling. Suppose that with present computers we can carry out a
simulation of a turbulent boundary layer with displacement thickness Reynolds
number of 4000 (double that of Spalart who did his simulations over a decade
ago). At this rate it will be another 20 years before we can simulate the
Reynolds numbers typical of the ow over the wings of large aircraft, and
then only in small computational domains corresponding to perhaps a few
centimetres in each direction. The external aerodynamics application is one
of the most extreme, but timescales upwards of 20 years appear in a majority
of technological application areas. Thus for design purposes one will rely on
closure at the RANS and LES levels for many years to come. That said, there
are already niche areas where DNS can contribute at Reynolds numbers that
are already relevant; examples of these relate to transition in attached or
separated ows as discussed elsewhere in this volume. What will be important
for further development of RANS and LES is that it will be feasible to provide
DNS data over a range of low to medium Reynolds numbers (covering perhaps
a factor of four variation), such that Reynolds number trends can be assessed.
As we have seen, the number of grid points required depends on the compu-
tational box size and the microscale of turbulence . However this microscale
was derived purely on dimensional grounds. From practical calculations of
ows away from solid boundaries it appears that the actual resolution re-
quired is approximately h = 5 where h is the grid spacing. For ow near
walls the turbulence is highly anisotropic. Guidelines for the required resolu-
tion are based on wall units using a reference length of /u
where u
=
_
w
/
is the friction velocity,
w
being the wall shear stress. These units are given a
superscript
+
and based on this we have guidelines x
+
< 15 (streamwise grid
spacing) z
+
< 8 (spanwise) and 10 points in the region y
+
< 10 (normal).
The normal direction is given dierently as the grid is usually stretched in this
direction. To develop a more general criterion Manhart (2000) has suggested
the use of a directional dissipation scale.
252 Sandham
A corresponding Kolmogorov micro-timescale for the small eddies, , is given
by
=
_
_
1/2
. (3)
The time step must be selected so that the smallest timescales of turbulence are
accurately computed. For the majority of algorithms (fully explicit and mixed
explicit/implicit) used for DNS the time step required for stability is already
signicantly smaller than this timescale, so this should be already resolved. Of
course generally this should be checked as each new ow is attempted.
3 Numerical methods
The governing equations for incompressible ow are the continuity equation,
which reduces to a solenoidal constraint on the velocity
u
i
x
i
= 0, (4)
and the three components of the momentum equation
u
i
t
+u
j
u
i
x
j
+
1
p
x
i
=
2
u
i
x
j
x
j
. (5)
It can be seen that the pressure only appears as a derivative and thus an arbi-
trary constant can be added to it. A reference pressure must be decided upon
for each calculation. The nonlinear terms, shown here in a non-conservative
form, may be rearranged before discretisation as discussed later. It is from
these terms that turbulence develops its characteristic wide range of spatial
scales. To solve the equations on a computer we need to consider methods for
time advancement and spatial discretisation. For the time advancement, meth-
ods suitable for ordinary dierential equations (Runge-Kutta, CrankNicolson,
AdamsBashforth) may be readily obtained from numerical methods texts and
will not be discussed further. For many applications an implicit treatment of
the viscous terms, in addition to pressure, is desirable. It is rare to nd fully
implicit methods used, since the cost per time step is high with these meth-
ods and to resolve the small time scales of turbulence many time steps are
required. For the remainder of this section we focus on issues relating to the
spatial discretisation and the calculation of derivatives.
With the vector computers that dominated scientic computing in the 1980s
and the rst half of the 1990s computer memory was a major limiting factor
and much eort was spent developing ecient methods for simulating a range
of scales accurately. A useful idea of the resolving power of a numerical dis-
cretisation can be obtained by a simple modied wavenumber analysis which
asks how well a given method calculates single Fourier modes e
ix
.
[7] Introduction to direct numerical simulation 253
hk
h
k
0 1 2 3
0
0.5
1
1.5
2
2.5
3
exact (spectral)
2nd
4th
6th (compact)
Figure 2: Graph of modied against actual wavenumber for dierent schemes,
including the exact spectral solution.
For example on substitution into a standard second-order accurate nite
dierence method
f
(x) =
f(x +h) f(x h)
2h
(6)
we have the result
f
(x) = i
sin(h)
h
e
ix
. (7)
This can be compared to the exact result f
= ie
ix
. Hence we can think of
a modied wavenumber
=
sin(h)
h
. (8)
This is shown on Figure 2 together with a fourth-order central scheme
f
(x) =
2 (f(x +h) f(x h))
3h
f(x + 2h) f(x 2h)
12h
, (9)
and a compact 5-point scheme (Lele 1988) requiring a tridiagonal matrix in-
version
f
(x+h)+3f
(x)+f
(xh) =
7 (f(x +h) f(x h))
3h
+
f(x + 2h) f(x 2h)
12h
.
(10)
We can see that these schemes have an increasing range of wavenumbers over
which they compare well to the exact result shown by the straight line. In this
case a Fourier spectral method gives the exact result directly.
254 Sandham
In fact a great deal of eort has been spent on exploiting the properties of
spectral methods for turbulence simulation. Subtleties of the methods concern
mainly the nonlinear terms. One can make use of fast transforms (FFTs) to
reduce the order N
2
convolution to something of order N log N, where N is
the number of modes. In these pseudo spectral methods the time advance
takes place in wave (Fourier transformed) space, while the nonlinear terms
are computed in real space, using FFTs for the transformations back and
forth. Aliasing errors are a source of concern with spectral methods since the
high wavenumber components interact during the calculation of the nonlinear
terms producing wavenumbers that are not resolved. These can then reect
back and corrupt wavenumbers that are carried by the computation. Solutions
to the problem involve various degrees of additional expense, for example
mode truncation, mode extension (the 3/2 rule is perhaps the most commonly
used approach) and random phase shifting (Rogallo 1981). A subject that has
been less well explored is the inuence of the mathematical formulation of the
convective terms. In particular skew-symmetric formulations (Zang 1989) have
attractive properties. The latest study on the combined eect of convective
term formulation and de-aliasing method is by Kravchenko and Moin (1997).
For incompressible isotropic turbulence and related ows where all three di-
rection are assumed periodic, spectral methods are well developed. The reader
is referred to the book by Canuto et al. (1987) for details. For inhomoge-
neous incompressible ows the numerical issues relate to the simultaneous
satisfaction of the incompressibility (divergence free velocity) condition and
the boundary conditions, for example no-slip walls. The inuence matrix tech-
nique of Kleiser and Schumann is described in Canuto et al. Another elegant
solution is to use divergence free variables (Spalart et al. 1991). Parallel im-
plementation of the Kleiser-Schumann method is discussed in Sandham and
Howard (1998).
The limitation of spectral methods is the geometry. Fourier methods neces-
sitate periodic boundary conditions. Ecient schemes using Chebyshev poly-
nomials are available for problems with one inhomogeneous direction, allowing
channel and boundary layer problems to be computed eciently. However with
more than one inhomogeneous direction spectral methods become rather cum-
bersome. Various tricks with buer regions (e.g. Spalart and Coleman 1998)
extend the range of ows that can be computed with periodic boundary condi-
tions, including separation bubble ows, but these zones need tuning for each
case. Spectral element methods (see for example the book by Karniadakis and
Sherwin 1999) set out to combine the excellent resolution properties of spec-
tral methods with the geometric exibility of nite element methods. Cost has
been a limiting factor thus far in the use of such methods and they have not
yet found wide application to DNS of turbulence.
For compressible ow, compact and non-compact high-order nite dierence
schemes are popular and explicit time advance schemes are usually applied.
[7] Introduction to direct numerical simulation 255
Severe numerical problems are present when shock waves coexist with turbu-
lence. For particular examples, the reader is referred to the compressible ow
chapter of this book.
With recent MPP architectures the limitation of computer memory for DNS
has eased, leaving run time the critical factor. Ecient spectral turbulence
simulations that use all the memory of a 512 processor Cray T3E, for exam-
ple, would take many months to run on that computer. There is nowadays a
perceptible trend to return to lower- (second-) order schemes to allow more
complex geometries to be calculated. The extra memory available is used to
add extra grid points to compensate, at least in part, for the reduced accuracy
of the scheme. Examples are the Le and Moin (1987) backward-facing step
calculation, the Wu et al. (1999) bypass transition calculation and the Yao et
al. (2000, 2001) trailing-edge ow. For such methods an early rule of thumb
was that a factor of 1.5 to 2 more resolution in all directions was required.
This requirement gets worse for higher-order statistics but with 50% more
resolution in all directions, compared to spectral methods, it is possible to get
statistical data from nite dierence simulations that is suciently accurate
for modelling at the second moment level.
4 Validation Issues
Producing quality simulation data is by no means a trivial task. Computer
programs containing several thousands of lines of code need to be debugged
initially, but even a working code needs to be applied properly. Users of the
data need to be familiar with the methods by which data has been validated
so that they can use it sensibly. Here we list some of the major issues.
A. Basic numerical method validation. This might be calculation of exact
solutions of the NavierStokes equations or of known stability charac-
teristics. For example in a boundary layer ow one might x a Blasius
base ow and compare the growth rates and mode shapes of uctua-
tions with the eigenvalues and eigenvectors from a separate solution of
the Orr-Sommerfeld equation.
B. Internal consistency checks on the results. It may be veried for ex-
ample that the output statistics from a turbulence simulation conserve
mass and momentum and balance the mean and turbulence kinetic en-
ergy equations to an acceptable accuracy. If instantaneous owelds are
available it can be veried that the ow locally satises the governing
equations.
C. Resolution, domain size and sample size. This checks the convergence of
the solution. Put simply a good simulation should give the same answers
to within an acceptable error bound when recomputed with double the
number of points in each direction separately, or with half the time
256 Sandham
step, or with double the domain size in each direction, or with twice the
statistical sample time. With spectral methods it is often convenient to
monitor an energy spectrum during a simulation and thus ensure that
there is sucient resolution for the energy in the highest wavenumbers
to be small (for example six to eight orders of magnitude smaller than
the peak).
D. Comparisons between dierent codes. This is rarely done before publi-
cation, but a published DNS will often become a test case that is re-
computed by other researchers using dierent codes. This provides an
accumulated condence in the data.
E. Comparison with experimental data. One might expect this to be the
most popular validation approach but it is rare to nd examples of DNS
where the main source of condence in the results is a demonstrated
agreement with experiments. This is due primarily to the simple geome-
tries of most DNS to date and the diculty of reproducing (or indeed
measuring at all) the exact inow or initial conditions in laboratory ex-
periments. In two cases of the listed milestone simulations (Kim, Moin
and Moser 1987, and Gilbert 1988) there was actually a severe disagree-
ment with existing data (Gilbert shows the experiments overestimate the
mean ow mean ow, plotted in wall units, by 10% and the spanwise tur-
bulence uctuations by up to 50%). This was in fact due to measurement
diculties in the near-wall region. Later experiments have been in much
better agreement with the DNS (Nishino and Kasagi 1989). Another
problem is the diculty of obtaining genuinely two-dimensional ows in
experiments for comparison with simulations employing periodic span-
wise boundary conditions. A majority of supposedly 2D experiments
may not in fact be suitable for any sort of CFD validation (Spalart 2000).
None of the above methods are sucient in themselves, and a combination is
usually employed. Of most interest to users of DNS databases are the internal
checks that can be done on the data after it has been accumulated and stored.
5 Applications
Despite restrictions of Reynolds number and the variety of ows available, the
use of DNS data is changing the way turbulence models are built and tested. In
particular the completeness of information available from simulation, including
terms such as pressure-velocity correlation that have not been available from
experiments, is enabling the use of DNS data to test closure models at several
dierent levels. Poor models can now be rejected very quickly. The sources of
errors in existing models can be more readily identied and detailed budgets
can be used to identify missing physics and suggest new modelling strategies. In
the following subsections we will highlight a range of uses for DNS, using data
[7] Introduction to direct numerical simulation 257
Figure 3: Sketch of plane channel ow. The ow between two innite plates is
driven by a mean pressure gradient in the x direction.
from two ow simulations as examples. Data from a wider range of simulations
is contained in Jimenez et al. (1998).
The rst case is a standard incompressible channel ow, the basic ow ge-
ometry of which is shown on Figure 3. The streamwise and spanwise directions
are periodic and resolved using Fourier modes. In the other direction there are
two no-slip walls and a Chebyshev discretisation is used. Results for Reynolds
number equal to 180, based on friction velocity and channel half width, will
be used, corresponding to the original Kim, Moin and Moser (1987) simula-
tion. The computational domain for that calculation was 4 in the streamwise
direction and 2 in the spanwise direction, with lengths normalised by the
channel half width. This test case, although relatively simple and of limited
use in developing more widely applicable closure models, provides a useful
demonstration of the potential of simulations.
The second case is a blunt trailing-edge conguration, shown on Figure
4. Here a precursor simulation of an incompressible zero pressure gradient
turbulent boundary layer was run to generate the inow data for the main
incompressible ow simulation. The inow boundary, with Reynolds number
1000 based on the inow displacement thickness, is located upstream of the
trailing edge. The main simulation Reynolds number is 1000 based on trailing
edge thickness. The downstream boundary uses a convective outow treat-
ment, while a slip-wall condition is used for the upper and lower boundaries.
The spatial discretisation is second-order on a staggered grid. Details of the
simulation are given in Yao et al. (2000, 2001). Validation included a series of
tests on grids up to 1024 256 128, together with calculation of turbulence
kinetic energy and Reynolds stress budgets. This test case provides a more
stringent test of turbulence models, and is used to illustrate some of the issues
that arise as more complicated ows are simulated.
258 Sandham
X
Y
Z
O
5h +15h
+8.5h
7.5h
0
+3h
3h
Figure 4: Schematic of the arrangement for the trailing edge simulation. A
precursor simulation provides inow data. Dimensions are given in terms of
the trailing edge thickness.
5.1 Flow visualisation
A rst use of DNS is direct interrogation of the instantaneous oweld to ob-
tain a qualitative understanding of the characteristics of the ow. Computer
visualisation allows the full three-dimensional oweld to be studied, and the
time dependence to be animated. Phenomena such as vortex shedding, sep-
aration and ow impingement on solid surfaces, which may have a decisive
inuence on the best approach to closure for a whole class of problems, can
be identied from an early stage.
There is no agreed denition of a vortex in turbulent ow, and dierent
measures have been used to reveal dierent aspects of instantaneous ow-
elds. Iso-surfaces of constant pressure, enclosing low pressure regions, iden-
tify important coherent structures (Robinson 1994). The second invariant of
the velocity gradient tensor
=
u
i
x
j
u
j
x
i
(11)
is also often used to identify swirling motions as regions where rotation rate
dominates over strain rate. This is the same as using Q from the PQR scheme
described in Chong et al. (1990). This measure tends to highlight smaller scale
structures. Another measure recently used by Jeong and Hussain (1995) is to
set a threshold on the second eigenvalue of a particular tensor which corre-
sponds to the pressure hessian
2
p/x
i
x
j
when viscous eects are ignored.
In practice for channel ow this measure turns out to be very similar to the
-criterion. The setting of a threshold for all the above measures implies some
degree of arbitrariness as to what constitutes a vortex. Kida and Miura (1998)
[7] Introduction to direct numerical simulation 259
(a)
(b)
Figure 5: Visualisation of the trailing edge ow. The ow is from left to right.
Red surfaces enclose regions of low pressure while blue regions enclose vortices
detected by the criterion.
have instead proposed the construction of a vortex skeleton, based on the
existence of closed streamlines in a plane normal to the axis of the cores of
vortices detected as low pressure regions.
On Figure 5 we illustrate two of these measures for the trailing edge ow.
Enclosed low-pressure regions are shown by the red surfaces. These clearly
pick out the large-scale shedding of predominantly spanwise vortices from the
trailing edge. Further downstream these can be seen to break up and lose
their spanwise coherence. Also shown are -vortices detected from the second
invariant of the velocity gradient eld. This measure picks out smaller scale
structures, such as the vortices that are subject to a large strain rate in the
region between successive spanwise rollers. The threshold that was set for
both these measures is larger than the strongest structures in the incoming
260 Sandham
Figure 6: Mean velocity of plane channel ow predicted by ve turbulence
models (Launder-Sharma, Chien, Kawamura-Kawashima, K-g and Durbin),
compared with DNS. For more details of the models, see Howard and Sandham
(2000).
boundary layer ow, indicating that the structures in the near wake region
are more intense. The upstream turbulent boundary layers, near wake and far
wake are thus distinct structural regions, which has implications for selecting a
closure approach. Initial results from comparisons with closure methods (Yao
et al. 2000, 2001) suggest serious limitations of RANS models for this ow.
5.2 A posteriori model testing for LES and RANS
The rst method of model testing consists of using the DNS data as an exten-
sion of the experimental database. Boundary conditions and initial or inow
conditions are matched to the DNS and then the model is run and predictions
for mean ow and turbulence kinetic energy are compared with DNS. This
standard method has become known in the LES community as a posteriori
testing (Piomelli et al. 1988) to distinguish it from a second a priori method
of using simulation data. As an example of this approach, on Figures 6 and 7
we reproduce from Howard and Sandham (2000) the mean velocity and tur-
bulent kinetic energy computed from the set of two equation models identied
in the gure caption, compared with the DNS of the turbulent channel ow.
It can be seen that variations in the mean ow prediction are of the order of
10%, while the turbulence kinetic energy varies by as much as a factor of two
between models.
In this usage, the DNS results are used in the same way as experimental
data: models have to be changed and constants tuned to improve the mean ow
predictions. This has to be done over a range of ows for the resulting model
to be practically useful. An advantage of the DNS-based approach is that it is
[7] Introduction to direct numerical simulation 261
Figure 7: Turbulence kinetic energy in plane channel ow predicted by ve
turbulence models, compared with DNS.
much easier to ensure that simulations and model calculations have exactly the
same boundary, initial and inow conditions, where appropriate. This is much
more dicult with experiments where, for example, inow dissipation proles
are not measured. Also automatic optimisation of constants to minimise the
error between DNS and model prediction can be readily carried out. This saves
time, and means that the modeller can concentrate on nding the right model
formulation to predict a certain range of turbulent ows. This requires some
judgement as to what a model is expected to predict, whether mean ow alone,
or mean ow plus turbulence kinetic energy, or other turbulence quantities.
5.3 A priori model testing for LES and RANS
A priori testing involves testing the basic assumptions and detailed term by
term construction of a turbulence model. For application to LES, the DNS
eld can also be ltered and used to test the local accuracy of the sub-grid
model. For RANS applications we can take for example the exact k equation,
which can be derived from the NavierStokes equations as
k
t
+U
j
k
x
j
= u
i
u
j
U
i
x
j
i
x
j
_
u
i
x
j
+
u
i
x
j
_
x
j
_
J
u
j
+J
p
j
+J
j
_
, (12)
where the rst and second terms on the right are the production and dissi-
pation rate of turbulence and the last three terms are turbulence transport
terms. The triple moment term J
u
j
= u
i
u
i
u
j
/2 is usually modelled as gradient
diusion with an eddy viscosity. The pressure transport term J
p
j
= p
j
is
usually ignored, while the viscous term J
j
= k/x
j
can be included as is
262 Sandham
Figure 8: Kinetic energy budgets for the trailing edge ow problem. (P is
production, dissipation, J transport and R convection.)
and need not be modelled. What can done using DNS data is to compare the
modelling of each term with the actual term measured from DNS.
As an example of a kinetic energy budget we take the trailing edge ow
problem. Figure 8 shows budgets at two locations (a) upstream of the trailing
edge and (b) downstream of the trailing edge near the end of the recirculation
[7] Introduction to direct numerical simulation 263
region. The various terms are labelled according to the above equation. We
can see that in the outer region of the boundary layer an approximate bal-
ance of production and dissipation is valid, while in the near wall region and
downstream of the wake other terms are signicant. Particular features of this
ow that require attention from a modelling point of view are the negative
production region immediately downstream of the trailing edge (not shown on
gure) and the pressure transport term, normally neglected, but which in this
ow appears as one of the most important terms in the near-wake region.
Another example of a priori testing is in the development of damping func-
tions for the eddy viscosity. Assuming an eddy viscosity relation of the form
t
= c
k
2
, (13)
Rodi and Mansour (1993) extracted the exact form of f
equation is a
more signicant source of error than the usually blamed equation.
264 Sandham
6 Conclusions
Databases of statistical quantities from DNS are already an essential part of
the turbulence modellers toolkit, whether used as a supplement to experi-
mental data, or as a guide to rational modelling methods. Further examples
of the use of DNS data will appear throughout this volume. In the future it
is expected that the relevance of DNS to the closure problem will increase
as simulations increase in Reynolds number and geometric complexity. While
not providing very high Reynolds number data, there will certainly be data
available over a sucient range of Reynolds numbers so that the correct trends
can be built into future models.
References
Canuto, C., Hussaini, M.Y., Quarteroni, A. and Zang, T.A. (1987). Spectral Methods
in Fluid Dynamics, Springer-Verlag.
Chong, M.S., Perry, A.E., Cantwell, B.J. (1990).A general classication of 3-dimen-
sional ow-elds, Phys. Fluids A-Fluid Dynamics 2, 765777.
Coleman G.N., Kim J. and Spalart, P.R. (2000). A numerical study of strained three-
dimensional wall-bounded turbulence, J. Fluid Mech. 416, 75116.
Gilbert, N. (1988). Numerische Simulation der Transition von der laminaren in die
turbulente Kanalstr omung. DFVLR-FB 88-55. DFVLR, G ottingen, Germany.
Howard, R.J.A. and Sandham, N.D. (2000). Simulation and modelling of a skewed
turbulent channel ow. Flow, Turbulence and Combustion. 65(1), 83109.
Jeong, J. and Hussain, F. (1995). On the identication of a vortex, J. Fluid Mech. 285,
6994.
Jimenez, J. et al. (1998). A selection of test cases for the validation of large-eddy sim-
ulations of turbulent ows. AGARD-AR-345, North Atlantic Treaty Organisation,
April 1998.
Karniadakis, G.E.M. and Sherwin, S.J. (1999). Spectral/hp Element Methods for CFD,
Oxford University Press.
Kida, S. and Miura, H. (1998). Identication and analysis of vortical structures, Eur.
J. of Mech. B-Fluids 17, 471488.
Kim, J., Moin, P. and Moser, R.W. (1987). Turbulence statistics in fully-developed
channel ow at low Reynolds number, J. Fluid Mech. 177, 133166.
Kravchenko, A.G. and Moin, P. (1997). On the eect of numerical errors in large
eddy simulations of turbulent ows, J. Comput. Phys. 131, 310322.
Lele, S.K. (1991). Compact nite-dierence schemes with spectral-like resolution,
J. Comput. Phys. 103, 1642.
Manhart, M. (2000). The directional dissipation scale: a criterion for grid resolu-
tion in direct numerical simulations. In Advances in Turbulence VIII: Proceedings
[7] Introduction to direct numerical simulation 265
of the Eighth European Turbulence Conference, C. Dopazo et al. (eds.), CIMNE,
Barcelona.
Moin, P. and Mahesh, K. (1998). Direct numerical simulation: a tool for turbulence
research, Ann. Rev. Fluid Mech. 30, 539578.
Moser R.D., Kim J. and Mansour N.N. (1999). Direct numerical simulation of tur-
bulent channel ow up to Re-tau=590, Phys. Fluids 11, 943945.
Nishino, K. and Kasagi, N. (1989). Turbulence statistics measurement in a two-
dimensional channel ow using a three-dimensional particle tracking velocimeter.
In Proceedings of Seventh Symposium on Turbulent Shear Flows, Stanford Univer-
sity, Stanford, CA, August 1989.
Orszag, S.A. and Patterson, G.S. (1972). Numerical simulation of three-dimensional
homogeneous isotropic turbulence, Phys. Rev. Lett. 28, 7679.
Parneix, S., Laurence, D. and Durbin, P.A. (1998). A procedure for using DNS
databases, J. Fluids Eng. 120, 4047.
Piomelli, U., Moin, P. and Ferziger, J.H. (1988). Model consistency in large eddy
simulation of turbulent channel ows, Phys. Fluids 31, 18841891.
Reynolds, W.C. (1989). The potential and limitations of direct and large eddy simu-
lation. In Whither Turbulence?, J.L. Lumley (ed.), Lecture Notes in Physics 357,
Springer, 313343.
Robinson, S.K. (1992). Coherent motions in the turbulent boundary-layer, Ann. Rev.
Fluid Mech. 23, 601639.
Rodi, W. and Mansour, N.N. (1993). Low-Reynolds-number k- modeling with the
aid of direct simulation data, J. Fluid Mech. 250, 509529.
Rogallo, R. (1981). Numerical experiments in homogeneous turbulence, NASA
TM81315.
Sandham, N.D. and Howard, R.J.A. (199). Direct simulation of turbulence using
massively parallel computers. In Parallel Computational Fluid Dynamics, D.R.
Emerson et al. (eds.), Elsever.
Spalart, P. (1988). Direct simulation of a turbulent boundary-layer up to R
= 1410,
J. Fluid Mech. 187, 6198.
Spalart, P. (2000). Trends in turbulence treatments, AIAA paper 00-2306.
Spalart, P. and Coleman, G.N. (1998). Numerical study of a separation bubble with
heat transfer, Eur. J. Mech. B-Fluids 16, 169189.
Spalart, P.R., Moser, R.D. and Rogers, M.M. (1991). Spectral methods for the
NavierStokes equations with one innite and 2 periodic directions, J. Com-
put. Phys. 96(2), 297324.
Spalart, P. and Watmu, J.H. (1993). Experimental and numerical study of a turbu-
lent boundary-layer with pressure-gradients, J. Fluid Mech. 249, 337371.
Wu, X., Jacobs, R.G., Hunt, J.C.R. and Durbin, P.A. (1999). Simulation of boundary
layer transition induced by periodically passing wakes, J. Fluid Mech. 398, 109
153.
Yao, Y, Sandham, N.D, Savill, A.M and Dawes, W.N. (2000). Simulation of a turbu-
lent trailing-edge ow using unsteady RANS and DNS. In Proc. 3rd Intl. Symp. on
Turbulence, Heat and Mass Transfer, Nagoya, January 2000.
266 Sandham
Yao, Y., Sandham, N.D., Thomas, T.G. and Williams, J.J.R. (2001). Direct nu-
merical simulation of turbulent ow over a rectangular trailing edge. Theor. and
Comput. Fluid Dyn. 14(5), 337358.
Zang, T.A. (1989). On the rotation and skew-symmetrical forms for incompressible-
ow simulations, Appl. Num. Math. 7, 2740.
8
Introduction to Large Eddy Simulation of
Turbulent Flows
J. Frohlich and W. Rodi
1 Introduction
This chapter is meant as an introduction to Large-Eddy Simulation (LES) for
readers not familiar with it. It therefore presents some classical material in a
concise way and supplements it with pointers to recent trends and literature.
For the same reason we shall focus on issues of methodology rather than
applications. The latter are covered elsewhere in this volume. Furthermore,
LES is closely related to direct numerical simulation (DNS) which is also
widely discussed in this volume. Hence, we concentrate as much as possible
on those features which are particular to LES and which distinguish it from
other computational methods.
For the present text we have assembled material from research papers, earlier
introductions and reviews (Ferziger 1996, H artel 1996, Piomelli 1998), and our
own results. The selection and presentation is of course biased by the authors
own point of view. Supplementary material is available in the cited references.
1.1 Resolution requirements of DNS
The principal diculty of computing and modelling turbulent ows resides in
the dominance of nonlinear eects and the continuous and wide spectrum of
observed scales. Without going into details (the reader might consult classical
text books such as Tennekes and Lumley (1972)) we just recall here that the
ratio of the size of the largest turbulent eddies in a ow, L, to that of the
smallest ones determined by viscosity, , behaves like L/ Re
3/4
u
. Here,
Re
u
= u
L/ with u
u c
1/2
f
,
c
f
Re
0.2
, where Re is based on the center line velocity and the channel
height. In a DNS no turbulence model is applied so that motions of all size
have to be resolved numerically by a grid which is suciently ne. Hence,
the computational requirements increase rapidly with Re. According to this
estimate a DNS of channel ow at Re = 10
6
for example would take around
hundred years on a computer running at several GFLOPS. This is obviously
not feasible. Moreover, in an expensive DNS a huge amount of information
267
268 Fr ohlich and Rodi
would be generated which is mostly not required by the practical user. He or
she would mostly be content with knowing the average ow and some lower
moments to a precision of a few percent. Hence, for many applications a DNS
which is of great value for theoretical investigations and model testing is not
only unaordable but would also result in computational overkill.
1.2 The basic idea of LES
Suppose somebody wants to perform a DNS but the grid that would be re-
quired exceeds the capacity of the available computer; so a coarser grid is used.
This coarser grid is able to resolve the larger eddies in the ow but not the
ones which are smaller than one or two cells. From a physical point of view,
however, there is an interaction between the motions on all scales so that the
result for the large scales would generally be wrong without taking into ac-
count the inuence of the ne scales on the large ones. This requires a so-called
subgrid-scale model as discussed below. Hence, LES can be viewed as a poor
mans DNS. The poor man, however, has to compensate by cleverness in that
a model for the unresolved motion has to be devised and an intricate coupling
between physical and numerical modelling is generated. On the other hand,
the resolution of the large scales of the ow while modelling only the small ones
not the entire spectrum is an advantage of the LES approach compared to
methods based on the Reynolds-averaged NavierStokes equations (RANS).
The latter methods often have diculties when applied to complex ows with
pronounced vortex shedding or special inuences of buoyancy, curvature, rota-
tion or compression. Finally, LES gives access to the dominant unsteady mo-
tion so that it can, for example, be used to study aero-acoustics, uid-structure
coupling or the control of turbulence by an appropriate unsteady forcing.
2 Governing Equations and Filtering
The NavierStokes equations (NSE) constitute the starting point for any tur-
bulence simulation. Here, we consider incompressible, constant-density uids
for which these equations read
u
i
x
i
= 0 (1)
u
i
t
+
(u
i
u
j
)
x
j
+
x
i
=
( 2S
ij
)
x
j
, (2)
where S
ij
= (u
j
/x
i
+u
i
/x
j
)/2 is the strain-rate tensor and = p/. For
later reference we introduce Reynolds averaging which is used in statistical
turbulence modelling (RANS) as time averaging: u) = lim
T
1
T
_
T
0
u dt.
Reynolds averaging has the properties
u)) = u), uv)) = u)v). (3)
[8] Introduction to large eddy simulation of turbulent ows 269
Figure 1: Illustration of Schumanns approach to LES as discussed in the text.
According to the idea of LES a means is required to distinguish between
small, unresolved, and larger, resolved structures. This is accomplished by the
operation u u, dened below. Unlike the above Reynolds time averaging,
this is an operation in space. The fact that RANS and LES methods employ
averaging in dierent dimensions inhibits an easy link between them. Several
attempts have been made to put both in a common framework (Speziale 1998,
Germano 1999) but they will not be discussed here. We now turn to the ways
of dening u and illustrate them in the one-dimensional case.
2.1 Schumanns approach
The volume-balance approach of Schumann (1975) starts from a given nite-
volume mesh. The integral of a continuous unknown u(x) in (1), (2) over one
cell is denoted
V
u =
1
x
_
V
u(x)dx as illustrated in Figure 1 (indices referring
to cells are dropped). Integrating the NSE over a cell and using Gauss theorem
relates these values to surface-averaged quantities denoted
j
, such as
j
uv.
These need to be expressed in terms of the cell-averages, which is done in
two steps. If the discretization is suciently ne, it is possible to replace
j
uv
by
j
u
j
v, with only a minor approximation error, as is usual in nite-volume
methods. This is done in DNS. If the grid is not ne enough, however, the
dierence can be signicant and the unresolved momentum ux
j
uv
j
u
j
v
270 Fr ohlich and Rodi
has to be accounted for by a model, the so-called subgrid-scale (SGS) model.
Subsequently, the
j
u are related to the
V
u either by setting them equal to
cell averaged quantities if a staggered arrangement is used or by interpolating
from neighbouring values. The nal SGS contribution to be modelled therefore
also depends on the expressions used for
j
u, i.e. on the discretization scheme.
To sum up, the equations are discretized and thereby the split into large and
small scales is performed, since the latter cannot be resolved by the discretized
system. Note that the operations u
V
u and u
j
u map an integrable
function onto discrete values; a continuous function u(x) is not constructed.
Thus, with Schumanns approach, scale separation, discretization, and the SGS
model are not separated conceptually but are intimately tied together. This
has advantages in that anisotropies and inhomogeneities of the grid can easily
be incorporated. However, it renders the analysis of the various contributions
to the solution relatively dicult and hence is considered too restrictive by
many workers in the eld.
2.2 Filtering
Leonard (1974) proposed dening u by
u(x) =
_
+
G
_
x x
_
u
_
x
_
dx
. (4)
An integral of this kind is called a convolution. Here, G is a compactly
supported, or at least rapidly decaying, lter function with
_
G(x) dx = 1
and width . The latter can be dened by the second moment of G as
=
_
12
_
x
2
G(x) dx. Figure 2 displays the Gaussian Filter, G
G
=
_
6/ (1/)
exp(6x
2
/
2
), and the box lter dened by G
B
= 1/ if [x[ /2 and
G
B
= 0 elsewhere. In fact, Deardor (1966) had already used (4) in the spe-
cial case G = G
B
. Figures 3a and 3b illustrate the ltering with smaller or
larger lter width: the larger , the smoother is u. According to (4), u is a
continuous smooth function as displayed in Figure 3 which can subsequently
be discretized by any numerical method. This has the advantage that one can
separate conceptually the ltering from the discretization issue.
It is helpful to transfer equation (4) to Fourier space by means of the deni-
tion u() =
_
u(x) e
ix
dx, since in Fourier space, where the spatial frequency
is the independent variable, a convolution integral turns into a simple prod-
uct. Equation (4) then reads
u() =
G() u() . (5)
Figure 4 illustrates the ltering in Fourier space. Equation (5) allows the def-
inition of another lter, the Fourier cuto lter with
G
F
() = 1 if [[ /
and 0 elsewhere. From (5) it is obvious that only this lter yields u = u
since (
G
F
)
2
=
G
F
. In all other cases the identity is not fullled. This can
[8] Introduction to large eddy simulation of turbulent ows 271
G(x)
`
x
>
>
G
G
>
>
G
B
Figure 2: Gaussian lter G
G
and box lter G
B
as dened in the text, both
plotted for the same lter width .
be appreciated by comparing u and u for the box lter in Figures 3 and 4.
The second relation in (3) is never fullled except in trivial cases, so that for
general ltering we have
u ,= u, uv ,= u v (6)
which distinguishes clearly the ltering in LES from Reynolds averaging (see
Germano (1992) for a detailed discussion).
The vertical line in Figure 4 represents the nominal cuto at / related
to the grid. The Fourier cuto lter
G
F
would yield a spectrum of u which
is equal to that of u left of this line, and zero, right of it. Equation (5) and
Figure 4 therefore demonstrate that when a general lter, such as the box
lter, is applied, this does not yield a neat cut through the energy spectrum
but rather some smoother decay to zero. This is important since SGS modelling
often assumes that the spectrum of the resolved scales near the cuto follows
an inertial spectrum with a particular slope and a particular amount of energy
transported from the coarse to the ne scales on the average. We see that even
if u fullls this property this can be altered by ltering (for further remarks
see Section 4.3). Nevertheless, it is convenient and common to use the notion
of a simple cuto as a model in qualitative discussions. Equation (5) is also
helpful for illustrating the fact that derivative and lter operations commute,
i.e. u/x = (u/x). Any convolution lter (4) can be written as in (5)
regardless of the choice of G. Dierentiation appears as multiplication by i
in Fourier space (see equation (28) below), which is commutative.
Applying the three-dimensional equivalent of the lter (4) to the NSE (1)
and (2), the following equations for the ltered velocity components u
i
result:
u
i
x
i
= 0 (7)
272 Fr ohlich and Rodi
(a)
x
u
u
u
`
(b)
x
u
u
u
`
Figure 3: Filtered functions u and u obtained from u(x) by applying a box
lter: (a) narrow lter, (b) wide lter.
u
i
t
+
(u
i
u
j
)
x
j
+
x
i
=
( 2S
ij
)
x
j
ij
x
j
, (8)
where S
ij
and are dened analogously to the unltered case. The term
ij
= u
i
u
j
u
i
u
j
(9)
represents the impact of the unresolved velocity components on the resolved
ones and has to be modelled. In mathematical terms it arises from the nonlin-
earity of the convection term, which does not commute with the linear ltering
operation.
An important property of u
i
is that it depends on time. Hence, an LES
necessarily is an unsteady computation. Furthermore, u
i
always depends on
all three space-dimensions (except for very special cases). Symmetries of the
boundary conditions generally produce the same symmetries for the RANS
[8] Introduction to large eddy simulation of turbulent ows 273
u
u
u
log E()
log
Figure 4: Eect of ltering on the spectrum. Here the box lter employed in
Figure 3 is used as well, but the curves are similar for other lters such as
the Gauss Filter. u
and u
), or if the prolongation of u
274 Fr ohlich and Rodi
from a nite domain to the real axis induces discontinuities, the commutation
property is lost and additional commutator terms arise in (7), (8) (Ghosal and
Moin 1995). In contrast to the usual SGS term
ij
, which is generated by the
nonlinearity of the convection term, the commutator also appears for linear
expressions (see the discussion by Geurts (1999) and Section 4.3). This issue
is relevant for pronounced grid stretching in the interior of the domain and
close to walls but has been disregarded until recently. Studies for a channel
ow are reported in Fr ohlich et al. (1998, 2000).
2.4 Implicit versus explicit ltering
The ltering approach relaxes the link between the size of the computed scales
and the size of the grid since the lter can be coarser than the employed
grid. Consequently, the modelled motion should be called sublter- rather than
subgrid-scale motion. The latter labelling results from the Schumann-type ap-
proach and is frequently used for historical reasons to designate the former.
In practice, however, the lter G does not appear explicitly at all in many
LES codes
1
so that in fact the Schumann approach is followed. Due to the
conceptual advantages of the ltering approach, reconciliation of both is gen-
erally attempted in two ways. The rst observation is that a nite-dierence
method for (7), (8) with a box lter employs the same discrete unknowns as
Schumanns approach; for example u(x
k
) =
V
k
u with k referring to a grid
point. Choosing appropriate nite-dierence formulae, the same or very sim-
ilar discretization matrices are obtained in both cases. Another argument is
that the denition of discrete unknowns amounts to an implicit ltering i.e.
ltering with some unknown lter (but one that in principle exists) since any
scale smaller than the grid is automatically discarded. In this way the lter is
used more or less symbolically only to make the eect of a later discretization
appear in the continuous equations. This is easier in terms of notation and
stimulates physical reasoning for the subsequent SGS modelling.
In contrast to implicit ltering one can use a computational grid ner than
the width of G and only retain the largest scales by some (explicit) lter-
ing operation. This explicit ltering has recently been advocated by several
authors such as Moin (1997) since it considerably reduces numerical discretiza-
tion errors as the retained motion is always well resolved. On the other hand it
increases the modelling demands since for the same number of grid points more
scales of turbulent motion have to be modelled and it is not yet completely
clear which approach is more advantageous (Lund and Kaltenbach 1995). The
ltering approach of Leonard is now almost exclusively used in papers on LES
and has triggered substantial development, e.g. in subgrid-scale modelling. In
practice, however, it is most often used for conceptual reasons rather than as
a precise algorithmic construction.
1
apart from some ltering operations for the dynamic model, discussed below, which is
of a somewhat dierent nature
[8] Introduction to large eddy simulation of turbulent ows 275
3 Subgrid-scale modelling
3.1 Introduction
Subgrid-scale modelling is a particular feature of LES and distinguishes it
from all other approaches. It is well-known that in three-dimensional turbu-
lent ows energy cascades, in the mean, from large to small scales. The primary
task of the SGS model therefore is to ensure that the energy drain in the LES
is the same as that obtained with the cascade fully resolved, as in a DNS.
The cascading, however, is an average process. Locally and instantaneously
the transfer of energy can be much larger or much smaller than the average
and can also occur in the opposite direction (backscatter see Piomelli et
al. 1996). Hence, ideally, the SGS model should also account for this local,
instantaneous transfer. If the grid scale is much ner than the dominant scales
of the ow, even a crude model will suce to yield the right behaviour of
the dominant scales. This is for two reasons. First, the larger the distance,
in wavenumber space, between dierent contributions, the looser is their cou-
pling. Second, as a consequence of this, as well as from the energy cascading,
the ner scales exhibit a more universal character which is more amenable
to modelling. On the other hand, if the grid scale is coarse and close to the
most energetic, anisotropic, and inhomogeneous scales, the SGS model should
be of better quality. Obviously, there exist two possible approaches; one is
to improve the SGS model and the other is to rene the grid. In the limit,
the SGS contribution vanishes and the LES turns into a DNS. Rening the
grid, however is restricted due to rapidly increasing computational cost. The
alternative strategy, for example solving an additional transport equation in
a more elaborate SGS model, can be comparatively inexpensive.
Another aspect results from the numerical discretization scheme which intro-
duces a dierence between the continuous dierential operators and their dis-
crete equivalents. This dierence is particularly large close to the cuto scale.
For DNS this is not so disturbing, but with LES we will later see that it is pre-
cisely these scales which have a substantial inuence on the modelled SGS con-
tribution. Hence, in LES, the discretization scheme and the SGS model have to
be viewed together. Indeed, some schemes such as low-order upwind discretiza-
tions generate a considerable amount of numerical dissipation as discussed in
Section 5.2. Therefore certain authors perform LES without any explicit SGS
model (Tamura, Ohta and Kuwahara 1990, Meinke et al. 1998). The grid is re-
ned as much as possible to decrease the importance of the SGS terms, and the
energy drain is in one way or another accomplished by the numerical scheme.
Although yielding valuable results in some cases, this kind of modelling can
barely be evaluated or controlled. Hence, in most LES, central or spectral
schemes are used and the SGS term is represented by an explicit model.
We shall now turn to the description of some basic SGS models before giving
a summary at the end of this section.
276 Fr ohlich and Rodi
3.2 Smagorinsky model
The Smagorinsky model, SM, (Smagorinsky 1963) was the rst SGS model
and is still widely used. Like most of the current SGS models, it employs the
concept of an eddy viscosity, relating the traceless part of the SGS stresses,
a
ij
, to the strain rate S
ij
of the resolved velocity eld:
a
ij
=
ij
1
3
ij
rk
= 2
t
S
ij
. (10)
The advantage of (10) is that the resulting equation for u
i
to be solved looks
like (2) with u
i
instead of u
i
, +
1
3
ij
kk
instead of , and +
t
instead
of . Hence, it is very easy to incorporate this into an existing solver for the
unsteady NSE.
The second part of this model involves the determination of the eddy vis-
cosity
t
. Dimensional analysis yields
t
l q
SGS
(11)
where l is the length scale of the unresolved motion and q
SGS
its velocity scale.
From the above discussion it is natural to use the lter size as the length
scale, hence we set l = C
s
. Similarly to Prandtls mixing length model, the
velocity scale is related to the gradients of u
i
expressed by
q
SGS
= l [S[ [S[ =
_
2S
ij
S
ij
(12)
which yields
t
= (C
s
)
2
[S[. (13)
This amounts to assuming local equilibrium between the production of the SGS
kinetic energy, P =
a
ij
S
ij
and dissipation expressed by q
3
SGS
/l. Introducing
(10) and (12) in P = gives (13). The constant C
s
can be determined assuming
an inertial-range Kolmogorov spectrum for isotropic turbulence which yields
C
s
= 0.18. This value has turned out to be too large for most ows so that
often C
s
= 0.1 or even lower values are employed. Close to walls
t
has to
be reduced to account for the anisotropy of the turbulence. This is generally
accomplished by replacing C
s
in (13) with C
s
D(y
+
). Most often the van Driest
damping is used
D(y
+
) = 1 e
y
+
/A
+
, A
+
= 25, (14)
which is known from statistical models. However, this yields
t
(y
+
)
2
for
small y
+
while
t
should behave like y
+3
. The correct behaviour is achieved
by the alternative damping function (Piomelli, Moin and Ferziger 1993)
D(y
+
) =
_
1 e
(y
+
/A
+
)
3
_
1/2
. (15)
[8] Introduction to large eddy simulation of turbulent ows 277
The main reason for the frequent use of the SM is its simplicity. Its drawbacks
are that the parameter C
s
has to be calibrated and its optimal value may vary
with the type of ow, the Reynolds number, or the discretization scheme. The
kind of damping to be applied near a wall is a further point of uncertainty.
Also, the SM, like any other model based on (10) with
t
0, is strictly dis-
sipative and does not allow for backscatter. It is furthermore not appropriate
for simulating transition since it yields
t
0 even in laminar ows.
3.3 Dynamic procedure
From the previous section it is apparent that for physical reasons one would
prefer to replace the constant value C
s
by a value changing in space and time.
The dynamic procedure has been developed by Germano et al. (1991) in order
to determine such a value from the information provided by the resolved scales,
in particular the ones close to the cuto scale. In fact this procedure can be
applied with any model
mod
ij
(C, , u) for
ij
or
a
ij
containing a parameter C
2
.
The basic idea is to employ this model not only on the grid scale, or lter
scale, but also on a coarser scale
as illustrated in Figure 5. This is the
so-called test scale with, e.g.,
= 2:
sub-grid scale stresses (-level) :
ij
= u
i
u
j
u
i
u
j
mod
ij
(C, , u) (16)
sub-test scale stresses (
-level) : T
ij
=
u
i
u
j
u
i
u
j
mod
ij
(C,
,
u). (17)
From the known resolved velocities u
i
the velocities
u
i
can be computed by
applying the lter . . . to u
i
using an appropriate function
G. Similarly, the term
L
ij
=
u
i
u
j
u
i
u
j
can be evaluated. It is this part of the sub-test stresses T
ij
which is resolved on the grid as sketched in Figure 5: The total stresses
u
i
u
j
in the expression for T
ij
can be decomposed into the contribution u
i
u
j
resolved on the grid and the remainder
ij
. Inserting this in (17) gives
T
ij
= L
ij
+
ij
(18)
known as Germanos identity. Hence, on one hand L
ij
can be computed, on the
other hand the SGS model yields a model expression when inserting (16),(17)
in (18):
L
mod
ij
=
mod
ij
(C,
,
u)
mod
ij
(C, , u). (19)
Ideally, C would be chosen to yield
L
ij
L
mod
ij
= 0, (20)
but this is a tensor equation and can only be fullled in some average sense,
minimizing, e.g., the root mean square of the left-hand side as proposed by
Lilly (1991). Principally, the consecutive application of
G and G to obtain
u
2
In this subsection we distinguish between exact and modelled SGS stresses for clarity.
278 Fr ohlich and Rodi
log E()
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
/
T
ij
/
ij
log
Figure 5: Illustration of the dynamic modelling idea as discussed in the text.
yields an eective lter of width
,=
, which generally is even of dierent
type to
G and G (e.g. when the box lter is used). This issue is generally
neglected in the literature. For that reason and since
=
, with the Fourier
cuto lter presently used for illustration, we write
instead of
in this
section. Eectively, it is the ratio
/ which is required by the dynamic
models.
We now apply the dynamic procedure to the SM (10),(13) to get
L
mod
ij
= 2 C
2
[
S[
S
ij
+ 2 C
2
[S[S
ij
(21)
with C = C
2
s
for convenience. Classically, the model is developed by extracting
C from the ltered expression in the second term although in fact C will vary
in space. The right-hand side of (21) can then be written as 2CM
ij
so that
inserting into (20) with the least-squares minimization mentioned above yields
C =
1
2
L
ij
M
ij
M
ij
M
ij
. (22)
The advantage of (22) or a similar equation is that now the parameter of
the SM is no longer required from the user but is determined by the model
[8] Introduction to large eddy simulation of turbulent ows 279
itself. In fact, it is automatically reduced close to walls and vanishes for well-
resolved laminar ows. Negative values of C are possible and can be viewed as
a way of modelling backscatter. The resulting backward diusion can however
generate numerical instability so that +
t
0 is often imposed. Furthermore,
C, determined by (22) as it is, exhibits very large oscillations which generally
need to be regularized in some way. Most often L
ij
and M
ij
are averaged in
spatially homogeneous directions in space before being used in (22). However,
this requires the ow to have at least one homogeneous direction. Another way
is to relax the value in time according to C
n+1
= C + (1 )C
n
using C
n
from the previous time step (Breuer and Rodi 1994). Yet another way is to
use the known value C
n
in the rightmost term of (21) so that it need not be
extracted from the test lter (Piomelli and Liu 1995). This yields smoothing
in space without any homogeneous direction required.
3.4 Scale similarity models
Scale similarity models (SSM) were created to overcome the drawbacks of
eddy viscosity-type models. Filtering the decomposition u
i
= u
i
+ u
i
yields
the (exact) relation
u
i
= u
i
u
i
. (23)
This can be interpreted as equality between the largest contributions of u
i
and
the smallest contributions of u
i
(see Figure 4). Furthermore, it is computable
from u
i
. Introducing the decomposition of u
i
into (9) and modelling u
i
u
j
u
i
u
j
and u
i
u
j
u
i
u
j
, respectively, yields the model
a
ij
= L
m,a
ij
with
L
m
ij
= u
i
u
j
u
i
u
j
, (24)
where . . .
a
indicates the traceless part of a tensor. A model constant is not
introduced as this would destroy the Galilean invariance of the expression. For
a spectral cuto lter u is replaced by
u with
=
2 since for this lter
u = u as discussed above.
The SSM allows backscatter, i.e. transfer of energy from ne to coarse scales,
and does not impose alignment between the SGS stress tensor
ij
and the strain
rate S
ij
. On the other hand, (24) turns out to be not dissipative enough so that
it is generally combined with a Smagorinsky model. Horiuti (1997) subsumes
some current SSMs in the model
a
ij
= C
L
L
m,a
ij
+C
B
L
R,a
ij
2C
2
[ S [ S
ij
(25)
with L
R
ij
= u
i
u
j
u
i
u
j
, evaluated using (23). A further step is to combine
(25) with the dynamic procedure for the determination of the constants:
(a) C
with C
L
= 1, C
B
= 0 (Zang, Street and Kose 1993);
(b) C
L
and C
with C
B
= 0 (Salvetti and Banerjee 1995);
280 Fr ohlich and Rodi
(c) C
B
and C
with C
L
= 1 (Horiuti 1997).
Dierent tests in the cited references as well as by Piomelli, Yu and Adrian
(1996) show that SSMs, in conjunction with the dynamic procedure, perform
quite well for low-order nite-dierence or nite-volume methods. Apart from
the ability to represent backscatter this may also be due to the fact that
no spatial derivatives are involved in the SSM which reduces the impact of
numerical discretization errors.
3.5 Further models and comparative discussion
Let us sum up a few strategies or concepts which are currently followed in
SGS modelling. One, already mentioned in the beginning of this section, is to
employ a crude model and to compensate by grid renement, which decreases
the impact of the model.
Another strategy is to employ the same approaches as in RANS mod-
elling. The Smagorinsky model, based on an eddy-viscosity and an algebraic
mixing-length expression, is the most prominent example. But as with RANS,
more elaborate methods can be used to compute the turbulent viscosity,
such as a model employing a transport equation for the SGS kinetic energy,
k
SGS
= 1/2
kk
, which furnishes a velocity scale, q
SGS
= k
SGS
1/2
, (Schumann
1975, Davidson 1997). Obviously, the lter width constitutes an adequate
reference length so that according to (11)
t
= C k
1/2
SGS
is a reasonable model
and no second length-scale-determining transport equation is required. Spalart
et al. (1997) have developed an approach called Detached Eddy Simulation
(DES). They start from a one-equation RANS turbulence model (Spalart and
Allmaras 1994) based on a transport equation for
t
. In this equation the dis-
tance from the wall is introduced as a length scale in the destruction term.
Replacing this physical length scale by a resolution-based scale, C
D
(where
C
D
is a parameter), turns the model into a SGS model. This method fur-
thermore oers a particular way of wall modelling which is discussed below.
Still more complex approaches have been carried over from RANS. Fureby et
al. (1997) employ the SGS equivalent of a Reynolds-stress model and obtain
satisfactory results in some tests. The cost increase is claimed to be moderate,
as solving the pressure equation requires most of the work.
A third strategy, applied with SSM and the dynamic procedure, is based
on the multiscale nature of turbulence. It could only be developed with scale
separation dened independently from the discretization according to (4) since
ltering is used as an individual operation. By analyzing experimental ow
elds along these lines Liu, Meneveau and Katz (1994) propose
ij
= C
L
L
ij
(26)
with L
ij
dened in (18). This diers from (24) since two lters of dierent size
are used.
[8] Introduction to large eddy simulation of turbulent ows 281
A fourth strategy is to relate SGS models to classical theories of turbulence.
An elementary example is the determination of the Smagorinsky constant as-
suming a Kolmogorov spectrum. This strategy is also pursued when den-
ing a wave-number-dependent eddy viscosity to be employed with a spectral
Fourier discretization and using EDQNM theory to determine
t
(k) (Chollet
and Lesieur 1981). The spectral eddy viscosity model has also been reformu-
lated in physical space for application in complex ows yielding the structure
function model (Metais and Lesieur 1992):
t
= 0.063
_
F
2
(), F
2
(r) = (u
i
(x +r) u
i
(x))
2
, (27)
with F
2
spatially averaged in an appropriate way. Dierent variants have been
developed (Lesieur and Metais 1996). It can be shown that when implemented
in a nite-dierence context, this yields a Smagorinsky-type model with [S[ in
(13) replaced by [u
i
/x
j
[.
The last strategy we mention concerns the testing of SGS models. Of course,
as with other turbulence models, prototype ows can be computed and the
results then compared with experimental data or DNS results. This is still the
ultimate test to pass. However, another kind of testing particular to LES has
been developed, namely the so-called a priori test: a fully resolved velocity eld
from a DNS is used to explicitly compute the terms which have to be modelled
in an LES on a coarser grid. The large-scale velocity on that grid is extracted to
determine the SGS stresses by means of a SGS model. The dierence between
exact and modelled stresses reects the quality of the model. This information
should however be taken with some caution as the test involves discretization
eects in a substantially dierent way than in the actual LES.
Finally, one has to bear in mind that a perfect SGS model is impossible.
Assume the exact grid-scale velocity u is known at all points. The perfect SGS
model would then amount to inferring from u on the exact instantaneous SGS
velocity u
_
u
x
_
() = i
G() u(). (28)
The numerical evaluation of u/x by a nite-dierence formula corresponds
to replacing the factor in (28) by a modied wavenumber
e
() which de-
pends on the particular scheme employed. Derivation and formulas are given,
e.g., by Ferziger and Peric (1996). For symmetric schemes this is a real quan-
tity, otherwise it is complex. Starting from
e
(0) = 0, [
e
[ increases and
then drops down to zero again. The point
so that
the solution to be computed is located entirely at /
0. If however
(b)
log
for a
second-order scheme in a three-dimensional computation.
also computed a plane channel with Re
) = W(u
1
), y
1
). (29)
Here, y
1
is the distance of the rst grid point from the wall,
) the average
wall shear stress, u
1
) the average tangential velocity at y
1
, and W a functional
dependence. Note that any relation u)
+
= f(y
+
) can be converted to the form
(29). This can be the log-law, the 1/7-power law, a linear viscous law or even
a numerical t to DNS data. An appropriate blending is generally used so that
W is dened from y
+
1
0 to y
+
of several hundreds.
Even if many physical properties such as the low-order moments, and hence
W, are well-known for a certain ow this is the case for the developed ow
in a plane channel, for example it is a delicate task to introduce this knowl-
edge in the context of LES. The available information is of a statistical nature
whereas the ltered velocity is an instantaneous, uctuating quantity. On the
other hand it has been demonstrated that the inner and outer regions of a tur-
bulent boundary layer are only loosely coupled (Brooke and Hanratty (1993))
so that an artical boundary condition bridging the inner layer has a chance
of being successful (Piomelli 1998). One of these wall-function methods (Schu-
mann 1975) employs the mean velocity u(y
1
)), which is successively computed
during the LES, to determine the average wall shear stress
)
u(y
1
))
u(y
1
), (30)
which is used as a boundary condition in the LES. Werner and Wengle (1993)
employed the 1/7-power law instead of the log-law to avoid an iterative eval-
uation of W. Furthermore, they replaced (29) by
= W(u
1
, y
1
) so that the
average velocity is no longer needed. Other combinations and variants are pos-
sible as well. A generalization of the approach for the subcritical ow around
a cylinder is described by Fr ohlich (2001). In the technically relevant case of a
rough surface, using the wall-function approach is unavoidable since resolving
the ow around each roughness element is impossible. The roughness eect is
brought in by the roughness parameter in the log-law (Gr otzbach 1977).
288 Fr ohlich and Rodi
Wall function boundary conditions work reasonably well in simple ows and
save substantial CPU time due to the reduced resolution requirements. They
have also been applied to some complex ows around obstacles (Rodi et al.
1997) which, however, were found not to be very sensitive to variations in the
boundary conditions.
5.3 Other approaches
Wall functions establish a relation between the local wall shear stress and
the velocity at the wall-adjacent grid point. This can be generalized to the
case where information on a line or a whole plane at some distance parallel
to the wall is used to generate the wall-shear stress at a certain point. Such
information can be introduced as a boundary condition in unsteady turbu-
lent boundary layer equations which are solved along the wall within the wall
cell using an embedded grid (Balaras, Benocci and Piomelli 1996) (cf. Figure
7d). In these equations turbulence is modelled with an eddy viscosity depend-
ing on the wall distance. Similar work has been done by Cabot (1995,1996)
where dierent models of this type were devised and applied to ow in a plane
channel and over a backward facing step. Although yielding better results in
some computations than wall functions, these methods have not found wide
application yet due to their implementational complexity.
Another approach that can be introduced more easily in an LES is based
on using the no-slip condition which in turn requires renement in the wall-
normal direction. Parallel to the wall, however, the step size of the outer region
is maintained leading to substantial savings. The idea then is to replace the un-
resolved near-wall structures by elements from RANS simulations. Schumann
(1975) decomposed
ij
into an isotropic part for which the Smagorinsky model
is used and an anisotropic part resulting from the mean ow gradient. The
latter part is modelled with an eddy viscosity
t,an
= min(c
x,z
, d)du)/dy
This is a RANS-like model in which close to the wall the size of the grid
x,z
is replaced by the distance from the wall d as a length scale. A similar switch
is used in the DES approach of Spalart et al. (1997) mentioned above so that
close to a wall the original RANS model is employed (see Figure 7b). DES,
although conceived for dierent applications, has been tested for channel ow
by Nikitin et al. (2000). The authors observe a spurious buer layer reecting
diculties in connecting the quasi-steady RANS layer close to the wall to the
outer unsteady computation. Further adjustments need to be introduced to
apply DES in such cases. Bagett (1998) discussed the issue of blending RANS
with LES turbulence models and points out the requirements for adequate
spanwise resolution in the near-wall region in order to capture the energeti-
cally dominant features. If these are not captured, unphysical structures are
generated which degrade the result.
[8] Introduction to large eddy simulation of turbulent ows 289
Figure 7: Schematical pictures for the dierent approaches close to solid walls:
(a) resolving the nearwall structure, (b) blending with a RANS model, (c)
application of a wall function, (d) determination of wall stress by boundary
layer equation solved along the wall on an imbedded grid.
5.4 Inow and outow conditions
After discussing the boundary conditions at solid walls we briey mention the
conditions at articial boundaries, an issue shared with DNS. Turbulent out-
ow boundaries are relatively uncritical. Here, damping zones or convective
conditions are generally applied which allow vortices to leave the computa-
tional domain with only small perturbations of the ow in its interior. A
convective condition for a quantity reads
t
+ U
conv
n
= 0 (31)
applied on the outlet boundary with n the outward normal coordinate and
U
conv
an appropriate convection velocity such as the bulk velocity.
The diculty posed by turbulent inow conditions stems from the fact that
LES computes a substantial part of the spectrum and hence requires specica-
tion of the inow conditions in all this spectral range, not just the mean ow.
The need for this information can be avoided by imposing streamwise periodic-
ity with a sucient periodic length, but this is inapplicable in many practical
ows. Imposing the mean ow plus random perturbations is generally not
successful since these perturbations are unphysical so that a large upstream
distance must be computed to produce the correct turbulence statistics. With
290 Fr ohlich and Rodi
more sophisticated perturbations the distance can be shortened. This is a sub-
ject of current research. If feasible the best solution is to impose some fully
developed ow at the inlet. A separate companion LES, e.g. with streamwise
periodicity, can then be performed to generate velocity signals at the grid
points in the inow plane of the main LES. An example is the ow around a
single cube investigated in Rodi et al. (1997).
5.5 Sample computations
In order to illustrate the above discussion, we present results from a standard
test case for LES calculations, namely fully developed plane channel ow. For
this ow between two innitely extended plates, periodic conditions can be
imposed in the streamwise direction x and the spanwise direction z, with typ-
ical domain sizes of L
x
= 2 and L
z
= , respectively. Reference quantities
are the channel half-width and the bulk velocity U
b
. DNS of this ow has
been performed for low and medium Reynolds number of which currently the
highest is Re
b
= U
b
/ = 10935 (Moser, Kim and Mansour 1999) employing
a high-precision spectral method with 384 257 384 points. This Reynolds
number has been used in the computations below. The results have been ob-
tained with the structured collocated nite-volume code LESOCC developed
by Breuer and Rodi (1994). The dynamic Smagorinsky model was used with
test-ltering and averaging in planes parallel to the walls. The bulk Reynolds
number was xed and an external pressure gradient adjusted so as to yield
the desired ow rate.
Figure 8 shows a computation where resolving the near-wall ow has been
attempted, i.e. no wall-function was used. The number of points in the y
direction is 65 and a stretching of 11% has been applied to cluster them close
to the walls. The gure shows the average streamwise velocity and the rms-
uctuations. The computed shear stress
w
yields Re
=
_
w
/ determines the
scaling of the axes, and in particular the u)
+
= f(y
+
)-curve is quite sensitive
to it. If the v- and w-uctuations are plotted in outer scaling, i.e. not by
u
, they are even further below the DNS curves. The observed failure occurs
because resolving the ow near the wall requires adequate discretization in all
three directions, not just normal to the wall. Here in particular the spanwise
resolution is too coarse.
In Figure 9 the wall-normal resolution has been improved using 159 points
in y with clustering in the buer layer accompanied by a substantially better
resolution in spanwise direction. The computed shear stress yields Re
= 598.5
and the Reynolds stresses compare quite satisfactorily with the DNS data. It
is obvious that with a structured discretization the grid in the interior of the
channel is ner than it really needs to be, due to the requirements near the
wall. To avoid this, a method with local renement is benecial as discussed
above.
[8] Introduction to large eddy simulation of turbulent ows 291
Figure 8: Computation without wall function using x
+
= 62, z
+
=
30, y
+
1
= 1.8. Left: u
+
, right: u
+
rms
, v
+
rms
, w
+
rms
, uv)
+
. Continuous lines are
DNS data, symbols LES.
Figure 10 presents a computation using the wall function of Schumann
(1975). In the wall-normal direction 39 equidistant volumes are used. In this
case the rst cell centre is still located in the buer layer, but the viscous
sublayer is well bridged. The computed wall shear stress gives Re
= 589.5
which compares very well with Re
i
u
j
x
j
. .
Reynolds stress term
(2.1)
U
i
x
i
= 0 (2.2)
[9] Introduction to two-point closures 301
and the equations for the uctuating component
u
i
t
+U
j
u
i
x
j
+u
j
U
i
x
j
+
x
j
(u
i
u
j
u
i
u
j
)
. .
Nonlinear term
=
p
x
i
. .
Pressure term
+
2
u
i
x
j
x
j
. .
Viscous term
(2.3)
and
u
i
x
i
= 0 (2.4)
Here, U
i
and p are the mean velocity and pressure (pressure divided by
density), while u
i
and p
i
, about
a basic ow, U
i
, with an additional forcing term, u
i
u
j
/x
j
, in the inhomo-
geneous case. Although the aims of stability theory (to characterise growth of
the perturbation) and of the theory of turbulence (to determine the statistics
of u
i
) are dierent, we believe it is nonetheless valuable to draw parallels be-
tween the two elds of study. It is our hope that in so doing we will encourage
workers in both areas to become more conversant with each others work.
Equation (2.3) is now used to derive equations for the time evolution of
velocity moments, i.e. averages of products of u
i
with itself at one or more
points in space. Setting up the equations for the nth-order velocity moments
at n points, one discovers that there are two main diculties. Firstly, the
term in (2.3) which is nonlinear in the uctuations leads to the appearance of
(n+1)th-order moments in the evolution equation at nth-order. Secondly, the
pressure term introduces pressure-velocity moments.
The pressure eld is intimately connected with the incompressibility condi-
tion. Indeed, taking the divergence of (2.3) leads to a Poisson equation
2
p
=
2
x
j
x
j
(u
i
U
j
+U
i
u
j
+u
i
u
j
u
i
u
j
) (2.5)
for the pressure uctuations. Solution of this equation by Greens functions
expresses p
i
in a PDF approach
P(
i
, x, t) = (u
i
(x, t)
i
))
allows removal of the problem of closure due to nonlinearity, which is present
in any description in terms of statistical moments. Finally, any problem of
closure is removed using the Hopf equation but the price to pay is an incred-
ibly complicated conguration-space! The probabilistic description, which is
of practical interest regarding a concentration scalar eld rather than a veloc-
ity eld, is extensively addressed elsewhere in this volume in the context of
combustion modelling, and we will no longer consider it here.
[9] Introduction to two-point closures 305
3 Review of [2,1] and [3,2] models
3.1 Second-order, one-point [2,1] models
In addition to simple closure models for the Reynolds-averaged NavierStokes
equations, such as models of turbulent viscosity using a mixing length as-
sumption, [2,1] models oer both a dynamical and a statistical description of
the turbulent eld, since the governing equations for the Reynolds stress ten-
sor, turbulent kinetic energy, and for its dissipation rate can reect the eects
of convection, diusion, distortion, pressure and viscous stresses, which are
present in the equations that govern the uctuating eld u
i
.
The exact evolution equation for the Reynolds stress tensor R
ij
= u
i
u
j
,
derived from (2.3), has the form
R
ij
t
+U
k
R
ij
x
k
. .
R
ij
= T
ij
+
ij
ij
x
k
(D
ijk
) (3.1)
where T
ij
=
U
i
x
k
R
kj
U
j
x
k
R
ki
is usually referred to as the production tensor
and is the only term on the right of (3.1) which does not require modelling,
since it is given in terms of the basic one-point variables U
i
and R
ij
of the
model. The remaining terms are not exactly expressible in terms of the ba-
sic one-point variables and heuristic approximations, forming the core of the
model, are introduced to close the equations.
The second term on the right of (3.1) is associated with the uctuating
pressure and is given by
ij
= p
(
u
i
x
j
+
u
j
x
i
), consisting of one-point correla-
tions between the uctuating pressure and rate of strain tensor. As discussed
in the introduction, p
ij
=
(r)
ij
+
(s)
ij
+
(w)
ij
(3.2)
corresponding to the three components of the Greens function solution of
(2.5). The rst is known as the rapid pressure component and arises from
the source terms in (2.5) which are linear in the velocity uctuation, stemming
from u
i
U
j
+u
j
U
i
. Being linear, this component is present in RDT, hence the
term rapid component. The second term in (3.2) is the slow component
and comes from the nonlinear source term in (2.5). Finally,
(w)
ij
is the wall
component and corresponds to a surface integral over the boundaries of the
ow in the Greens function solution for p
ij
have zero trace, and are conceived of as representing physically distinct
mechanisms of turbulent evolution. Hence, they are modelled dierently.
306 Cambon
The terms present in the rate equations for Reynolds stress models in homo-
geneous turbulence can be exactly expressed as integrals over Fourier space
of spectral contributions derived from the second-order spectral tensor
ij
,
which is the Fourier transform of double correlations at two points, and from
the third-order transfer spectral tensor T
ij
(details in section 4). All one-point
quantities in (3.1) can be expressed as integrals over wavenumber space. R
ij
is given by integrating
ij
, while
(r)
ij
= 2
U
m
x
l
_ _
2
lj
+
j
2
il
_
d
3
(3.3)
and
(s)
ij
=
_
T
ij
d
3
(3.4)
express the rapid and slow parts of
ij
, and
ij
= 2
_
2
ij
d
3
(3.5)
gives the viscous term (see the next subsection for details). In order to avoid
confusion the classical turbulent kinetic energy is denoted k with the wavevec-
tor denoted by , having for its modulus.
The equation for the dissipation rate =
i
) (in quasi-homogeneous and
quasi-incompressible turbulence), can be derived from the exact equation that
governs the uctuating vorticity eld
i
. However, the practical procedure for
deriving the -equation hardly uses the latter exact equation and consists of
basing the equation for / on the equation for
k/k with adjustable constants.
Advantages and drawbacks of single versus multi-point closure techniques
can be briey discussed as follows. Single-point closure models are much more
economical and exible, and can currently address anisotropic and inhomo-
geneous ows. Nevertheless, they can easily be questioned in the presence of
complex anisotropising mechanisms, and in the presence of a modied cascade
with spectral imbalance, even if one restricts the comparison to the pure homo-
geneous incompressible case (see [26] for more general ows). These weaknesses
appear in predicting the dynamics of R
ij
, when looking at the rapid pressure-
strain correlation for complex anisotropisation processes, and when looking at
the -equation and slow pressure-strain tensor for the cascade, sophisticated
though the single-point modelling of (3.3), (3.4), (3.5) may be. For the sake of
brevity, only the anisotropy problem will be illustrated, by comparing RDT
and single-point closures in the same rapid limit of homogeneous turbulence
in the presence of uniform mean velocity gradients. In this limit, R
ij
- mod-
els seem to work satisfactorily in the presence of an irrotational mean ow,
even for time-dependent pure straining processes, such as the successive plane
strains addressed by Gence and Mathieu (1979), or, more recently, for cyclic
irrotational compression (Le Penven, and Hadzic, Hanjalic and Laurence, pri-
vate communications). In the same situation, k- models give wrong results
[9] Introduction to two-point closures 307
because of the instantaneous relationship between the deviatoric part of R
ij
and the mean strain-rate tensor, a relationship that is usually known as the
Boussinesq approximation.
The same contrast between k- and full R
ij
-, with only the latter work-
ing satisfactorily, is found when looking at stabilising-destabilising eects of
rotation in a plane channel (only the trends induced by terms present in homo-
geneous turbulence are analysed). A clue to understand why R
ij
- can roughly
mimic RDT in the cases mentioned above, is their ability to take into account
the production term, in a way much more realistic than in k-. For instance,
the RDT solution for pure irrotational mean ow exhibits the dominant role of
the time-accumulated strain, which is a particular case of the Cauchy matrix,
so that correct trends can also be captured by R
ij
- models even if the rapid
pressure-strain is only roughly modelled, e.g. as proportional to the deviatoric
part of the production tensor. Regarding the relevance of rotational Bradshaw
(or Richarson) numbers for stabilisingdestabilising eects of rotation in a
shear ow, Leblanc and Cambon (1997) have explained why an apparently 2D
and pressure-less analysis (Bradshaw 1969) gave the same criterion as an ex-
act linear stability analysis (Pedley 1969). The reason is the dominant role of
pure spanwise modes, which are naturally unaected by pressure uctuations,
yielding again a production dominated mechanism. Things are completely
dierent when rotation interplays with the straining process in a more subtle
way, for instance by inducing inertial waves for which anisotropic dispersion
relationships aect (3.3), not to mention (3.4) and (3.5) beyond the RDT
limit. For instance, Townsends equations for homogeneous RDT were shown
(Cambon 1982; Cambon et al. 1985) to develop angular peaks of instability in
Fourier space if the rotation rate (half the vorticity) of the mean ow is strictly
larger than the strain rate, a fact which was recovered by Bayly (1986) in the
dierent context of incisively revisiting the gloried elliptical ow instability.
The reader is referred to Pierrehumbert (1986) for the literature on elliptical
ow instability, to Cambon and Scott (1999) for the linkage of stability analy-
sis to RDT, and nally to Salhi et al. (1997) for more details on the discussion
which is touched upon here.
Even without additional mean strain, pure rotation induces complex rapid
and slow eects, for which even the basic principles of single-point closures
are questionable (see subsection 5.2). Single-point closures look particularly
poor since there is no production by the Coriolis force, whereas the dynamics
is dominated by waves whose anisotropic dispersivity is induced by uctuating
pressure. This suggests discriminating turbulence dominated by production
eects from turbulence dominated by wavy eects. In short, single-point
closures are well adapted to simple turbulent ow patterns of the rst class in
rather complex geometry, whereas two-point closures are more convenient for
complex turbulent ows in simplied geometry, as illustrated by the second
class.
308 Cambon
3.2 Third-order, two-point [3,2] models
Although several dierent approaches exist in the literature, the simplest way
to introduce two-point closure models is to look at the governing equations for
double velocity correlations at two points and for triple correlations at three
points. This level of information removes from consideration non-local eects
(gure 1), so that the exact relationship between pressure and velocity is ac-
counted for. However, it is possible to derive a tractable formalism only for
quasi-homogeneous ows where Fourier space is relevant. Accordingly, wave-
space is an invaluable tool for these approaches, but it is only a mathemati-
cal convenience for treating non-local operators and multi-point correlations.
Hence the double correlations u
i
(x, t)u
j
(x +r, t)) at two points (or [2, 2] in
gure 1) are treated through their Fourier transform with respect to r, de-
noted
ij
(, t); this second-order spectral tensor is also proportional to the
covariance matrix u
i
u
j
). The equation that governs the second-order spec-
tral tensor, possibly in the presence of mean gradients uniform in space (Craya
1958), involves the transfer terms linked to triple correlations at two points
(or [3, 2]), which need to be closed. Rather than using an equation for the [3, 2]
term, it is more consistent (avoiding again the closure due to nonlocality) to
derive the latter from correlations at three points (or [3, 3]), which appear as
u
i
(, t) u
j
(p, t) u
l
(q, t)) with +p +q = 0 (3.6)
in agreement with triadic interactions caused by the quadratic nonlinearity
seen in Fourier space. Looking at the equations that govern (3.6), or
uuu)
t
+ exact linear uuu) terms = uuuu), (3.7)
a closure can nally be made by assuming a linear relationship between fourth-
and third-order cumulants, or
uuuu)
uu)uu) =
1
uuu). (3.8)
Regarding the structure of the equation given above, it is important to give
some preliminary remarks as follows:
The latter two equations are abridged and symbolic, in the sense that
the Fourier velocity components u
i
at dierent wave vectors are actually
involved in place of u, so that uuu) in (3.7) and (3.8) would repre-
sent (3.6) and uu) would represent
ij
. Accordingly, the true equation
abridged by (3.8) allows one to close the equation (3.7) that governs
(3.6), and then to close the equation for the second-order spectral ten-
sor: the innite hierarchy of open equations ([n, n] [n +1, n] in gure
1) is broken at the fourth order.
[9] Introduction to two-point closures 309
Equation (3.8) is formally consistent with nearly linear and nearly
Gaussian assumptions: since linear operators conserve the gaussianity,
pure linear dynamics (reected for instance by the so-called RDT) con-
serve the Gaussian properties if present in the initial data. Accordingly,
all cumulants remain zero in this situation and both right-hand side
and left-hand side of (3.8) identically vanish. Hence a linear relationship
such as (3.8) is consistent with considering both right-hand side and
left-hand side as formal weak departures from gaussianity, caused by
formal weak nonlinearity.
The impact of a closure relationship such as (3.8) on the dynamics of triple
correlations, that carry the energy cascade, is conventionally seen as follows:
since (3.7) can be rewritten as
uuu)
t
+ exact linearuuu)terms +
1
uuu) =
uu)uu),
the term
uu)uu) in the left-hand side acts as a source term for increasing
triple correlations, so that the contribution to the right-hand side term, which
comes from the closure relationship (3.8), appears as a damping term which
exhibits the characteristic time denoted . An ad hoc eddy-damping term is
chosen using the EDQNM-type model, but the structure of more sophisticated
two-point closure theories (DIA, TFM) is not fundamentally dierent. The
originality of the two-point closure models reported in the following sections
mainly lies in the straightforward treatment of anisotropising linear operators,
for both second- and third-order moments.
4 From RDT to anisotropic two-point closures
The simplest multi-point closure consists of the drastic measure of dropping all
nonlinear terms in (2.3) before averaging. If one also drops the viscous term,
in keeping with the high Reynolds number associated with the large scales of
turbulence, the result is known as rapid distortion theory (RDT), introduced
by Batchelor and Proudman (1954) (see Townsend 1976; Hunt and Carruthers
1990; and especially Cambon and Scott 1999, sections 2 and 5, for recent
reviews). In neglecting nonlinearity entirely, the eects of the interaction of
turbulence with itself are supposed to be small compared with those resulting
from mean-ow distortion of turbulence. One often has in mind ows such
as weak turbulence encountering a sudden contraction in a channel or ows
around an aerofoil. Implicit is the idea that the time required for signicant
distortion by the mean ow is short compared with that for turbulent evolution
in the absence of distortion. Linear theory can also be envisaged, at least over
short enough intervals of time, whenever physical inuences leading to linear
terms in the uctuation equations dominate turbulent ows, such as strongly
stratied or rotating uid or a conducting uid in a strong magnetic eld. For
such cases, the term rapid distortion theory is probably a little misleading.
310 Cambon
Thanks to linearity, time evolution of u
i
may be formally written as
u
i
(x, t) =
_
(
ij
(x, x
, t, t
)u
j
(x
, t
)d
3
x
(4.1)
where (
ij
(x, x
, t, t
to time t. Whereas u
i
is a random quantity, varying from realisation to
realisation of the ow, (
ij
is deterministic and can, in principle, be calculated
for a given U
i
(x, t). From (
ij
and the initial turbulence, (4.1) may be used to
determine later time behaviour.
Another simplifying assumption which is often made is that the size of tur-
bulent eddies, , is small compared with the overall length scales of the ow,
L, which might be the size of a body encountering ne-scale free-stream tur-
bulence (see e.g. Hunt 1973). In that case, one uses a local frame of reference
convected with the mean velocity and approximates the mean velocity gradi-
ents as uniform, but time-varying. Thus, the mean velocity is approximated
by
U
i
=
ij
(t)x
j
(4.2)
in the moving frame of reference. In the example of ne-scale turbulence en-
countering a body, one may imagine following a particle convected by the mean
velocity, which sees a varying mean velocity gradient,
ij
(t), even when the
mean ow is steady. This velocity gradient distorts the upstream turbulence
in a manner one would like to determine. Since the time history is dierent
depending on the particle considered, separate calculations are needed for the
dierent streamlines of a steady mean ow. In this case, linear solutions of
equation (2.3) are found as Lagrangian Fourier modes
u
i
(x, t) =
_
u
i
exp( x) d
3
(4.4)
the RDT solution is
u
i
[(t), t] = G
ij
(, t, t
) u
j
[(t
0
), t
0
] (4.5)
and (t
0
) is given in terms of (t) by a Cauchy matrix. Changes in the
wavenumber due to mean velocity gradients are reected as dependence of
u
i
(, t) on a dierent wavevector K = (t
0
) at time t
0
, a process of spectral
transfer in wavenumber space.
Recalling that the objective is to calculate statistical properties for a random
u
i
representing turbulence, one can use the above solution in terms of Fourier
transforms to that eect. This is straightforward if the turbulence is assumed
statistically homogeneous. In that case, the second-order, two-point moments
[9] Introduction to two-point closures 311
of velocity can be derived from a spectral tensor
ij
(, t) (Batchelor 1953),
related to the Fourier transform u
i
(k, t) in individual realisations by
u
i
(p, t) u
j
(, t) = u
i
(p, t) u
j
(, t) =
ij
(, t)( p) (4.6)
from which the two-point moments may be obtained via the Fourier transform
u
i
(x, t)u
j
(x +r, t) =
_
ij
(, t) exp(.r) d
3
(4.7)
This relation shows that, for homogeneous turbulence, the second-order mo-
ments in physical space and the spectral tensor in wavenumber space contain
essentially the same information. The one-point moment may be obtained by
setting r = 0 as
u
i
u
j
=
_
ij
(, t)d
3
. (4.8)
From (4.5), (4.6), and the fact that G
ij
is real, it follows that spectral
evolution takes the form
ij
[(t), t] = G
ik
(, t, t
0
)G
jl
(, t, t
0
)
kl
[(t
0
), t
0
]. (4.9)
Given an initial
ij
at t = t
0
, for instance isotropic, one can calculate it at
later times using (4.9), provided the Greens function G
ij
(, t, t
) is known.
The determination of G
ij
is thus the main problem in applying homogeneous
RDT in practice. The Greens function also will appear in models allowing for
nonlinearity through formal solutions of the moment equations, in which the
nonlinear terms are treated as forcing of the linear part.
Although purely linear theory closes the equations without further ado and
simplies mathematical analysis, it is rather limited in its domain of applica-
bility, ignoring as it does all interactions of turbulence with itself, including the
physically important cascade process. Multi-point turbulence models which ac-
count for nonlinearity via closure lead to moment equations with a well-dened
linear operator and nonlinear source terms. The view taken in this chapter is
that, even when nonlinearity is signicant, the behaviour of the linear part
of the model often still has a signicant inuence. Thus, it is important to
rst understand the properties of the linearised model, an undertaking which
is, moreover, mathematically more tractable than attacking the full model di-
rectly. As a bonus, linearised analysis often allows a simplied formulation of
the nonlinear model using more appropriate variables.
For the sake of brevity, we do not report here various calculations and ex-
periments in the area of homogeneous turbulence subjected to a mean ow
of the kind given by (4.2) (Cambon 1982; Cambon et al. 1985; Gence 1983;
Cambon and Scott 1999; Leuchter et al. 1992; Leuchter and Dupeuble 1993),
including elliptic and hyperbolic cases, nor stability analyses which use essen-
tially the same relationships as (4.2) and (4.5) (Cambon 2001a, and references
312 Cambon
therein). It is important, however, to recall that the feedback of the Reynolds
stress tensor in (2.1) vanishes due to statistical homogeneity (zero gradient of
any averaged quantity), so that the mean ow (4.2) has to be a particular so-
lution of the Euler equations and can be considered as a base ow for stability
analysis. In turn, the form (4.2) is consistent with maintaining homogeneity of
the uctuating ow governed by (2.3) and (2.4), provided homogeneity holds
for the initial data. This explains why homogeneous RDT can have the same
starting point as a rigorous and complete linear stability analysis in this case,
before the random initialisation of the uctuating velocity eld is considered
in (4.5).
In general, RDT operators break statistical isotropy at any scale, even if
the initial data are strictly isotropic. It should be borne in mind that isotropy
imposes a very special form
ij
(, t) =
E(, t)
4
2
_
ij
2
_
(4.10)
on the spectral tensor, where E(, t), with = [[, is the usual energy spec-
trum, representing the distribution of turbulent energy over dierent scales and
the quantity in brackets will be recognised as the projection matrix, P
ij
().
Thus,
ij
is determined by a single real scalar quantity, E, which is a function
of the magnitude of alone. Both the form of
ij
at a single point and its
distribution over -space are strongly constrained by isotropy. Given the, in
our view, importance of allowing for anisotropy and the associated eects of
mean ow gradients, we will not discuss the many isotropic models of spec-
tral evolution which have been proposed (see, for instance, Monin and Yaglom
(1975), section 17). Instead we concentrate on a small number of models ca-
pable of handling the anisotropic case and which illustrate the way in which
linear theory combines with nonlinear closures.
Independently of closure, the spectral tensor
ij
is not a general complex
matrix, but has a number of special properties, including the fact that it
is Hermitian, positive-denite, as follows from (4.6), and satises
ij
j
=
0, obtained from (4.6) and the incompressibility condition
j
u
j
= 0. Taken
together, these properties mean that, instead of the 18 real degrees of freedom
of a general complex tensor,
ij
has only four. Indeed, using a spherical polar
coordinate system in -space, the tensor takes the form (see Cambon et al.
1997 for details).
=
_
_
0 0 0
0 e +Z
r
Z
i
H/
0 Z
i
+H/ e Z
r
_
_
(4.11)
where the scalars e(, t) and H(, t) are real, and Z(, t) = Z
r
+ Z
i
is com-
plex. The quantity e(, t) =
1
2
ii
is the energy density in -space, whereas
H(, t) =
l
lij
ij
is the helicity spectrum and, along with Z, is zero in the
[9] Introduction to two-point closures 313
isotropic case. Anisotropy is expressed through variation of these scalars with
the direction of , as well as departures of H and Z from zero at a given
wavenumber. Whatever spectral closure is used, the number of real unknowns
may be reduced to the above four when carrying out numerical calculations,
and presentation of the results can be simplied using these variables, partic-
ularly when the turbulence is axisymmetric.
Our starting point for closure is the equation for the Fourier transform of
the velocity uctuation, which takes the form
u
i
+M
ij
u
j
+
2
u
i
= s
i
(4.12)
where
u
i
= u
i
/t
lm
l
u
i
/
m
corresponds to linear advection by the
mean ow (4.2), and M
ij
=
mj
(
im
2
i
m
/
2
) gathers linear distortion and
pressure terms. Once nonlinear and viscous terms are added, (4.12) generalises
the linear inviscid equation for which the RDT solution is (4.5). The nonlinear
term s
i
is given by
s
i
(, t) = P
ijk
()
_
p+q=
u
j
(p, t) u
k
(q, t)d
3
p (4.13)
in terms of a convolution integral, the usual expression of a quadratic non-
linearity, and P
ijk
=
1
2
(P
ij
k
+ P
ik
j
) which arises from the elimination of
pressure using the incompressibility condition
i
u
i
(, t) = 0.
The evolution equation for
ij
(Craya 1958), derived from (4.6) and (4.12),
is
ij
+M
ik
kj
+M
jk
ik
+ 2
2
ij
= T
ij
(4.14)
where the left-hand side arises from the linear part of (4.12), consisting of
the term
ij
, which, as in (4.12), is a convective time derivative in -space,
together with RDT and viscous components. A system of equations for the set
(e, Z, H), using (4.11), can readily be derived from (4.14); it is particularly use-
ful in the presence of solid body rotation (Cambon et al. 1992, 1997; Reynolds
and Kassinos 1995). Of course, this equation with a zero right-hand side has
the RDT solution (4.9). The detailed expression for the right-hand side of
(4.14), which represents nonlinear triadic interactions, consists of an integral
over the third-order spectral moments and requires closure. The quasi-normal
assumption gives the typical relationship between T
ij
and
ij
, as recalled be-
low. When the result is employed in the forcing term of the evolution equation
for the third-order moments and the latter solved by Greens function tech-
niques, one obtains
T
ij
(, t) =
ij
(, t) +
ji
(, t) (4.15)
where
ij
(, t)
= P
jkl
()
_
t
_
+p+q=0
G
im
(, t, t
)G
kp
(p, t, t
)G
lq
(q, t, t
)
314 Cambon
qn
(q
, t
)
_
1
2
P
mnr
(
)
pr
(p
, t
) +P
pnr
(p
)
mr
(
, t
)
_
d
3
p dt
, (4.16)
in which the triple product of Greens functions arises from the Greens func-
tion solution for the third-order moments and the notation +p +q = 0 on
the integral sign means that q should be replaced by p throughout the
integrand, representing interacting triads of wavenumbers , p, q which form
triangles. Equation (4.16) can be seen as a solution of the last symbolic equa-
tion of section 3. It gives the generic anisotropic structure of most generalised
classical theories dealing with two-point closure, provided the basic Greens
function is replaced by a slightly modied version, for instance including vis-
cous terms and eddy damping as in EDQNM.
We should perhaps say a few words about DIA (Kraichnan 1959), which is
more complicated than EDQNM, since it is based on spectral tensors involv-
ing two times, rather than
ij
(, t). Furthermore, it introduces an additional
response tensor for which, like the spectral tensor, an evolution equation is
formulated and closed using heuristic approximations. These approximations
are similar in nature to the quasi-normal one introduced above, supposing
as they do that the uctuating velocities have properties similar to those of
Gaussian variables, although such assumptions are less explicit in DIA. The
nal evolution equation for the two-time spectral tensor contains an integral
whose structure is much the same as the quasi-normal expression (4.16), with
terms such as G
lq
(q, t, t
)
qn
(q
, t
ln
(q
, t, t
ij
(, t), but they are nonetheless usually among the rst quantities to be
calculated following an anisotropic spectral calculation, along with correlation
lengths in dierent directions. Given
ij
, for instance obtained using RDT,
evaluation of the integral over 3D Fourier space can be a nontrivial task. For
example, the RDT Greens function can be determined analytically in the case
of simple shear, but the integrals in (4.8) are not straightforward and must be
evaluated numerically or asymptotically (Rogers 1991; D.J. Bodony and G.A.
Blaisdell, unpublished results).
5 Application to stably stratied and rotating uid
5.1 General features
In this section we consider application of the methods to turbulent ows in
stably stratied conditions and to turbulence subjected to rotation. These
cases illustrate the ows dominated by wavy eects with zero production
introduced at the end of subsection 3.1. For pure rotating turbulence there is
zero production of kinetic energy and for stratied rotating turbulence there is
[9] Introduction to two-point closures 315
zero production of total (kinetic plus potential) energy. The reader is referred
to Cambon et al. (1997) and Cambon (2001c) for more details. Linearised
solutions of the NavierStokes equations, with buoyancy force b within the
Bousinesq assumption, are easily obtained in the presence of a uniform mean
density gradient and in a rotating frame. For the sake of simplicity, the mean
ow is restricted to a uniform vertical gradient of density and to a solid body
rotation in the horizontal direction, with typical parameters N (the Brumt
Waisala frequency) and (the angular velocity). No additional mean velocity
gradients are considered in the rotating frame. Pressure uctuations are re-
moved from consideration in the Fourier-transformed equations by using a
local frame in the plane normal to the wave vector (Craya 1958), taking ad-
vantage of (2.4), so that the problem in ve components (u
1
, u
2
, u
3
, p, b) in
physical space is reduced to a problem in three components, two solenoidal
velocity components and a component for b, in Fourier space. For mathemati-
cal convenience, as in Cambon (1989) and Godeferd and Cambon (1994), the
velocity-temperature eld in three components is nally gathered into a single
vector v, whose 3D Fourier transform, denoted by a hat (), can be written
as
v = u +
1
N
. (5.1)
(A similar three-component term, denoted W
K
, is extensively used in Riley
and Lelong (2000), p. 626, but it is not a true vector, in contrast to v). The
scaling of the contribution from the buoyancy force allows one to dene twice
the total energy spectral density as
v
i
v
i
= u
i
u
i
+N
2
b. (5.2)
Without stratication, N
1
has to be replaced by another time scale
0
in
(5.2), but coupling between the velocity and buoyancy (or temperature) elds
vanishes in this case.
The linear equation for v is similar to (4.5), but with a constant wavevector
(t) = (t
0
), since the advection by a mean ow with antisymmetric gradient
(solid body rotation) amounts to the addition of a Coriolis force when the
motion is seen in the rotating frame, only modifying M
ij
in (4.12). A similar
Greens function can be expressed as
G
ij
(, t, t
0
) =
=0,1
N
i
()N
j
() exp[
(t t
0
)], (5.3)
in which N
0
and N
1
are the eigenmodes, related to quasi-geostrophic motion
and inertio-gravity waves respectively, whereas
=
_
N
2
(
/)
2
+ 4
2
(
/)
2
(5.4)
holds for the absolute value of the frequency given by the dispersion law of
inertio-gravity internal waves. Because of the form of the eigenvectors and of
316 Cambon
the dispersion law, the structure of G in (5.3) is consistent with axisymmetry
around the axis of reference (chosen vertical here), without mirror symmetry,
and
and
hold for axial (along the axis) and transverse (normal to the
axis) components of .
Looking only at inviscid RDT (e.g. Cambon 1989; Hanazaki and Hunt 1996;
van Haren 1993; van Haren et al. 1996 for pure stratied turbulence) the
following results can be predicted without detailed calculations.
Inviscid RDT solutions are derived from (5.3) and (5.4) for second-order
spectral tensors, which are related to u
i
u
j
) or v
i
v
j
). They consist of
sums of steady and oscillating terms, the frequency of oscillations being
directly connected to the dispersion law of internal waves. Such solutions
are completely reversible.
After integration over -space, such as (4.8), including averaging over all
directions, oscillating terms for spectral tensors yield damped oscillations.
This damping eect, called phase-mixing in Kaneda and Ishida (1999),
physically reects the anisotropic dispersivity of inertio-gravity waves.
It cannot appear for N = 2, or for particular initial data at N ,= 2,
such as the equipartition case considered for nonlinear applications in
Godeferd and Cambon 1994). The fact that some terms in u
i
u
j
are
conserved after integration, whereas other ones are damped, explains
the change of anisotropy over time for all single-point correlations. This
change is completely determined by the initial distribution in terms of
steady and wavy modes.
Except for the anisotropisation of two-time single-point correlations (see at
the end of [26]), the linear limit exhibits no interesting creation of structural
anisotropy. However in practice there is two-dimensionalisation in rotating
turbulence and a horizontal layering tendency in the stably stratied case.
In other words, RDT only alters phase dynamics, and conserves exactly the
spectral density of typical modes (full kinetic energy for the rotating case, to-
tal energy and vortex, or potential vorticity, energy for the stably stratied
case), so that two-dimensionalisation or two-componentalization (horizontal
layering), which aect the distribution of this energy, are typically nonlinear
eects. In Godeferd and Cambon (1999), RDT results have begun to be com-
pared with the results of a full DNS, in order to have the denite answer as to
what is linear (given by RDT) and what is nonlinear (only given by DNS) in
the rotating and stratied cases. In addition to such numerical comparisons,
the eigenmodes of the linear regime, derived from RDT, form a useful basis for
expanding the uctuating velocity-temperature eld, even when nonlinearity
is present, and nonlinear interactions can be evaluated and discussed in terms
of triadic interactions between these eigenmodes. Accordingly, the complete
anisotropic description of two-point second-order correlations, e.g. (4.11), can
be related to spectra and cospectra of these eigenmodes.
[9] Introduction to two-point closures 317
5.2 Pure rotation
Rotation of the reference frame is an important factor in certain mechanisms
of ow instability, and the study of rotating ows is interesting from the point
of view of turbulence modelling in elds as diverse as engineering (e.g. turbo-
machinery and reciprocating engines with swirl and tumble), geophysics and
astrophysics. Eects of mean curvature or of advection by a large eddy can be
tackled using similar approaches.
From several experimental, theoretical and numerical studies, in which rota-
tion is suddenly applied to homogeneous turbulence, some agreed statements
are summarised as follows (Bardina et al. 1985; Jacquin et al. 1990; Cambon
et al. 1992,1997; Cambon 2001c).
Rotation inhibits the energy cascade, so that the dissipation rate is re-
duced.
The initial 3D isotropy is broken through nonlinear interactions modied
by rotation, so that a moderate anisotropy, consistent with a transition
from a 3D to a 2D state, can develop.
Both previous eects involve nonlinear or slow dynamics, and the sec-
ond is relevant only in an intermediate range of Rossby numbers as
found by Jacquin et al. (1990). This intermediate range is delineated by
Ro
L
= u
rms
/(2L) < 1 and Ro
= u
rms
/(2) > 1, in which u
rms
is an
axial rms velocity uctuation, whereas L and denotes a typical integral
lengthscale (macroscale) and a typical Taylor microscale respectively.
If the turbulence is initially anisotropic, the rapid eects of rotation (lin-
ear dynamics tackled in a RDT fashion) conserve a part of the anisotropy
(called directional) and damp the other part (called polarization aniso-
tropy), resulting in a spectacular change of the anisotropy of R
ij
.
These eects, which are not at all taken into account by current one-point
second-order closure models (from k- to R
ij
- models), have motivated new
modelling approaches by Cambon et al. (1992, 1997), and to a lesser ex-
tent by Reynolds and Kassinos (1995) for linear (or rapid) eects only. It
is worth noticing that the modication of the dynamics by the rotation ulti-
mately comes from the presence of inertial waves (Greenspan 1968), having an
anisotropic dispersion law, which are capable of changing the initial anisotropy
of the turbulent ow and also can aect the nonlinear dynamics. Contrary to
a well-known interpretation, the Proudman theorem shows only that the slow
manifold (the stationary modes unaected by the inertial waves) is the 2D
manifold at small Rossby number, but cannot predict the transition from 3D
to 2D turbulence, which is a nonlinear mechanism of transfer from all possible
modes towards the 2D ones. In Fourier space, the slow and 2D mani-
fold corresponds to the wave plane normal to the rotation axis, or
= 0. In
318 Cambon
Figure 2: Correlation coecient of vertical velocity and temperature in decay-
ing stably-stratied turbulence. Comparison between some single-point closure
models and generalised EDQNM (gure courtesy van Haren 1993).
(5.3), only the wavy modes N
1
, which reduce to the Walee (1993) helical
modes, are present, and therefore form a complete basis for the velocity eld.
Accordingly, the resonant triads
k
p
q
0, with + p + q = 0, with
Energy
containing
cone
Energy
containing
cone
g
g
Figure 3: Representation of the spectral angular dependence occuring in strat-
ied (top left) and rotating (top right) turbulence, with the corresponding
schematic physical structures: layered ow (bottom left) for stratication;
columnar vertically correlated shapes (bottom right) for rotation.
The mode related to vertical wave-vectors appears to be very important,
since the concentration of spectral energy on it is the best identication of the
development of vertical collapse and layering. It corresponds to the limit of the
wavy mode, when the dispersion frequency tends to zero. Strictly speaking,
this mode is a slow mode, which cannot be incorporated in the wave-vortex
decomposition of Riley et al. (1981), and more generally is not present in a
classical poloidaltoroidal decomposition. It is absorbed in any decomposition
based on the CrayaHerring frame (see Cambon 2001c, Appendix), provided
that some care is taken to extend by continuity the denition of the unit
vectors (e
(1)
, e
(2)
) towards aligned with the polar (vertical here) axis of the
frame of reference (Cambon 1982, 2001c). In so doing, the mode related to e
(1)
coincides with a toroidal, or horizontal vortex, mode, but for vertical wave
vectors, where it includes half the energy of the vertical slow mode. In the
same way, the mode related to e
(2)
coincides with a poloidal mode, aected by
the wavy motion, but for vertical wave vectors, where it includes the other half
of the energy of the vertical mode. In the dierent context of weakly nonlin-
ear and weakly inhomogeneous RDT approach, and related DNS of Galmiche
et al. (2001), the vertical mode is considered as part of the mean ow and
is called the mean vertical shear mode. In fact, the DNS use Fourier modes
and periodic boundary conditions in all directions, with initial injection of
energy onto the largest vertical mode (
1
=
2
= 0,
3
=
min
), the so-called
[9] Introduction to two-point closures 321
Figure 4: a) Isovalues of u
2
/x
3
from a snapshot of 256
3
DNS (an illustration
of the layering phenomenon). (b) Integral length scales L
(1)
11
, with horizontal
separation (top) and L
(3)
11
, with vertical separation (bottom), from 256
3
DNS.
(c) Same quantities from EDQNM2 model. (Courtesy Godeferd and Staquet
2000).
mean shear mode, so that their interpretation in terms of inhomogeneous tur-
bulence and mean-uctuating interaction is only one possible interpretation.
More consistently, these results could be reinterpreted as purely homogeneous
and strongly anisotropic, illustrating concentration of energy towards verti-
cal wave-vectors as in the theoretical and numerical works by Godeferd and
Cambon (1994), Godeferd and Staquet (2000) and Smith (2000).
322 Cambon
6 Concluding remarks
We hope that this chapter has made clear the importance of linear, anisotropis-
ing processes even in turbulence which is too strong for strict validity of the
rapid distortion approximation. In principle, multi-point closures allow exact
treatment of linear terms.
On the other hand, single-point closures, being less computationally de-
manding and having no diculty with inhomogeneity, currently dominate in-
dustrial ow calculations, but involve many more heuristic assumptions. The
innite number of degrees of freedom of the spectral tensor are reduced to
just k and , or R
ij
and , in standard single-point models. Together with
a large enough number of adjustable constants, this may be sucient to de-
scribe the limited class of ows for which the model has been experimentally
parameterised, but one is always likely to encounter surprises in new types of
ows, as we have seen for the case of rotating turbulence. The latter ow is an
interesting example, not only because rotation is important in many practical
ows, but because it illustrates the subtle interplay between linear and nonlin-
ear processes and the signicance of spectral anisotropy. Anisotropy appears
in standard one-point models only via departure of u
i
u
j
from (2k/3)
ij
. The
quantity u
i
u
j
contains limited overall information on the spectral distribution,
while anisotropic structuring of the turbulence, leading to axially elongated
structures in the rotating case, is not captured at all, despite its physical
importance. The good behaviour of the TCL model, however, ought to be un-
derlined for stratied turbulence. This illustrates that the two-component and
2D limits have to be carefully discriminated.
The techniques and models presented in this chapter can be extended to
compressible ows, for which linear descriptions retain their importance. In
the case of turbulence subject to compression at small Mach number in ow
volumes of limited size, for example in the cylinders of piston engines, the mean
velocity has nonzero divergence, reecting the eects of compression, whereas
the uctuating velocity may be taken to be solenoidal, as in the incompress-
ible case (Mansour and Lundgren 1990). This description neglects thermal
boundary and acoustic eects, but allows the straightforward extension of in-
compressible models. Indeed, simple spherical compression can be taken into
account, including nonlinearity, by transformations of the time, turbulent ve-
locity and position variables without any need for additional modelling (Cam-
bon et al. 1992).
More generally, if the turbulent Mach number is not small, the eects of com-
pressibility are much more complicated, since both acoustic and entropy modes
are called into play, as well as the vortical mode inherited from the incompress-
ible case (Lele 1994). Irrotational ows have been studied by Goldstein (1978)
using an inhomogeneous RDT formulation (which will be discussed again in
[26]), while homogeneous RDT for rotational mean ows has shown the im-
[9] Introduction to two-point closures 323
portance of the gradient Mach number, S/c, where c is the sound speed and
S a measure of the mean velocity gradient (Simone et al. 1997). The latter
study helps explain the systematic changes in energy production rate with
gradient Mach number found in numerical simulations (Sarkar 1995). They
suggest that compressibility mainly alters the one-point properties of turbu-
lence through the pressure-velocity correlation tensor, rather than via bulk
viscous dissipation or other explicit compressible terms (e.g. the pressure-
dilatation term) usually considered in compressible one-point modelling. In
the absence of mean shear, interactions between solenoidal, dilatational and
pressure modes are purely nonlinear and can be analysed and modelled in
pure isotropic homogeneous turbulence. In this context, the spectral model
by Fauchet et al. (1997) gave very promising spectral information, as shown
in gure 5. Nonlinear transfer terms have a structure close to (4.16), with
a Greens function, or response tensor, which gathers both a classic linear
acoustic wave propagator and a nonlinear damping factor which ensures a
decorrelation time for triple correlations shorter than in classical EDQNM.
The two-point anisotropic description is more powerful, even if homogeneity
is assumed, than is generally recognized. In rotating and stratied turbulence
the anisotropic spectral description, with angular dependence of spectra and
cospectra in Fourier space, allows quantication of columnar or pancake struc-
turing in physical space. Among various indicators of the thickness and width
of pancakes, which can be readily derived from anisotropic spectra, integral
length scales L
(l)
ij
related to dierent components and orientations are the
most useful. It is also worth noting that a possible confusion can be made
between inhomogeneity and anisotropy, especially when considering vertically
stratied ows. For instance, DNS and RDT by Galmiche et al. (2001), which
are presented as being inhomogeneous, can be reinterpreted in the area of
strictly homogeneous strongly anisotropic turbulence. As another illustration
(Lee et al. 1990; Salhi and Cambon 1997), the streak-like tendency in shear
ows can be easily found in calculating both the L
(1)
11
component, which gives
the streamwise length of the streaks, and L
(3)
11
, which gives the spanwise sepa-
ration length of the streaks (as usual, 1 and 3 refer to streamwise and spanwise
coordinates, respectively). In pure homogeneous RDT at constant shear rate,
both length scales can be calculated analytically and their ratio (elongation
parameter) is found to increase as (St)
2
, S = U
1
/x
2
being the shear rate.
Finally, we would like to underline that a fully anisotropic spectral (or two-
point) description carries a very large amount of information, even if it only
concerns second-order statistics. In the inhomogeneous case, the POD (proper
orthogonal decomposition, Lumley 1967) has renewed interest in second-order
two-point statistics, but this technique is never completely applied to the ho-
mogeneous anisotropic case. It is only said that POD spatial modes are Fourier
modes in the homogeneous case, but the true spectral eigenvectors correspond-
ing to POD modes are not considered. These can in fact be easily obtained
324 Cambon
Figure 5: Sketch of the spectra obtained by the model of Fauchet et al. (1997).
From top to bottom, at larger wavenumber, the gure shows the solenoidal
(given) energy spectrum E
ss
, the incompressible part of pressure variance
spectrum E
pp
inc
, the dilatational energy spectrum E
dd
and the pressure variance
spectrum E
pp
. It can be seen that E
pp
collapses with E
dd
(acoustic equilib-
rium) only at smaller wavenumber, whereas it collapses with E
pp
inc
at larger
wavenumber, with E
pp
E
dd
.
by diagonalising the tensor
ij
, using the above e Z decomposition (4.11),
and noting that the angular position of the principal axes, associated with the
nonzero principal components e +[Z[ and e [Z[, is xed by the phase of Z,
at each .
References
Bardina, J., Ferziger, J.M., Rogallo, R.S. (1985). Eect of rotation on isotropic tur-
bulence: computation and modelling, J. Fluid Mech. 154, 321326.
Batchelor, G.K. (1953). The Theory of Homogeneous Turbulence. Cambridge Univer-
sity Press.
Batchelor, G.K., Proudman, I. (1954). The eect of rapid distortion in a uid in
[9] Introduction to two-point closures 325
turbulent motion, Q. J. Mech. Appl. Maths 7, 83103.
Bayly, B.J. (1986). Three-dimensional instability of elliptical ow, Phys. Rev. Lett.
57, 2160.
Benney, D.J., Saman, P.G. (1966). Nonlinear interactions of random waves in a
dispersive medium, Proc. R. Soc. London, Ser. A 289, 301320.
Bradshaw, P. (1969). The analogy between streamline curvature and buoyancy in
turbulent ows, J. Fluid Mech. 36, 177.
Caillol, P., Zeitlin, W. (2000). Kinetic equations and stationary energy spectra of
weakly nonlinear internal gravity waves, Dyn. Atm. Oceans 32, 81112.
Cambon, C. (1982). Etude spectrale dun champ turbulent incompressible soumis `a
des eets couples de deformation et rotation imposes exterieurement. Th`ese de
Doctorat d
t
+
U
i
x
i
= 0, (1)
and conservation of momentum (here in the absence of external forces) by the
NavierStokes equation,
t
U
i
+
x
j
U
j
U
i
=
p
x
i
x
j
T
ij
, (2)
328
[10] Reacting ows and probability density function methods 329
in which p is pressure and T
ij
the viscous stress tensor. For a Newtonian uid
the viscous stress tensor contains only simple shear eects and can be written
as
T
ij
=
_
U
i
x
j
+
U
j
x
i
_
+
2
3
U
k
x
k
ij
, (3)
in which is the dynamic viscosity.
The conservation equations for species mass fractions, denoted here by the
scalar vector , read
+
x
j
U
j
=
x
j
J
j
+S
(), (4)
where J
j
is the diusion ux and S
j
= ID
x
j
(5)
in which ID is the mass diusion coecient which here for simplicity is as-
sumed to be equal for all species. The enthalpy conservation equation which
in general is also needed, is also of the form (4) with the enthalpy source term
in particular containing eects of radiative heat transfer. The description is
completed with the thermodynamic equation of state and the caloric equation
of state (relating enthalpy and composition to temperature).
2 Basic statistical concepts
2.1 One-point statistics
Consider the variable which is a function of space x and time t (denoted by
(x, t)). The distribution function F of is dened as
F
(; x, t) =
(; x, t) (7)
The probability that a realization of (x, t) lies between
1
and
2
is given by
P
1
< (x, t) <
2
=
_
2
1
f
(; x, t)dt = F
(
2
; x, t) F
(
1
; x, t). (8)
The end values of F are appropriately dened as F() = 0 and F(+) = 1.
Field values and probability functions in general are considered to be functions
330 Roekaerts
of space and time but to simplify the notation the variables x and t are usually
omitted without further notication.
The average or expectation of is denoted by E, ), or by and can
be expressed in terms of the PDF by
) =
_
+
()d. (9)
In the same way, the expectation of any function Q of can be dened by
Q) =
_
+
Q()f
()d. (10)
With the denition of the ensemble average or Reynolds average the variable
can be decomposed into its mean ) and uctuation
according to
= ) +
. (11)
To clarify the notion of one-point statistics further, an example of two-
point statistics is given. Consider the two-point one-time statistics of the eld
variable (x, t) dened by F
(
1
,
2
; x, x+r, t) which describes the probability
that (x, t) <
1
, and (x +r, t) <
2
. This function cannot be expressed in
terms of the one-point statistics of . In the limit of [r[ 0 this description
reduces to the combined statistics of and its gradients. This means that
gradient statistics cannot be described in terms of the one-point statistics of
a variable and that such terms are unclosed in any one-point closure.
2.2 Joint probabilities
We now consider combined statistics of several variables (e.g. velocities and
species concentrations). This provides information about correlations between
variables.
Consider the n stochastic variables =
1
,
2
, . . . ,
n
which have a joint
probability function dened by:
F
() P
1
<
1
,
2
<
2
, . . . ,
n
<
n
. (12)
The joint PDF is dened by
f
() =
n
2
n
F
() (13)
From the joint PDF of the statistics of one variable
can be obtained by
integration over the other n 1 directions of the space:
f
) =
_
+
_
+
() d
1
d
1
d
+1
d
n
(14)
This one-variable PDF is also called the marginal PDF of
. As in the uni-
variate case (see equation (10)) the expectation of any function of the variables
i
can be expressed as an integral over the phase space of this function times
the joint PDF.
[10] Reacting ows and probability density function methods 331
2.3 Conditional probability
The conditional probability PA[B is dened as the probability that A
occurs given that B occurs, and is given by
PA[B =
PA, B
PB
. (15)
In the same manner the conditional PDF of
1
[
2
can be dened as
f
1
|
2
(
1
[
2
) =
f
1
,
2
(
1
,
2
)
f
2
(
2
)
. (16)
The expression f
1
|
2
(
1
[
2
)d
2
now denes the PDF of
1
given the fact that
2
<
2
<
2
+ d
2
.
Conditional statistics are an important concept in PDF modeling. It turns
out that all unclosed terms in the PDF equations can be written in terms of
conditional averages (e.g. of the uctuating velocity given a value of the scalar).
In other words, these terms can in general not be expressed as a pure function
of the describing variables. These conditional averages Q(
1
,
2
)[
2
=
2
)
can be written in terms of the conditional PDF f
1
|
2
by
Q(
1
,
2
)[
2
=
2
) =
_
+
Q(
1
,
2
)f
1
|
2
(
1
[
2
)d
1
. (17)
2.4 Favre averaging
In turbulent ames, density can vary by a factor of ve or more and density
uctuations can have large eects on the turbulent ow eld. To simplify the
equations describing variable density ow it is common to use density-weighted
(Favre) averaging.
The main advantage of using Favre-averaging instead of ensemble- or Rey-
nolds-averaging is the fact that explicit density correlations are avoided in the
equations (Jones 1980).
Let the density be denoted as , and let the vector denote the thermo-
chemical scalar variables (e.g. species concentrations, temperature). In the low
Mach-number limit, where it is assumed that pressure variation does not aect
the density, the density is a pure function of the scalar vector .
Favre averages are dened by
Q =
Q)
)
, (18)
and Favre-decomposition into mean and uctuation is dened as
Q =
Q+Q
, (19)
332 Roekaerts
in which Q
= 0 Q
= 0,
and in general,
,= 0 Q
,= 0.
It is also useful to dene the Favre-probability density function by
() = f
()
[ = )
)
= f
()
()
)
(20)
Because the density is a pure function of the scalar space variables, the con-
ditional average reduces to a function in and the second equality holds.
Favre-averages can be expressed in terms of the Favre-PDF according to
Q() =
_
+
Q()
()d. (21)
The Favre-PDF has the same properties as a standard probability density
function and all properties discussed in the previous sections (multivariate
statistics, conditional statistics) can be applied in a straightforward manner.
3 Averaged equations
By averaging the ow equations a set of equations describing the mean ow
properties is obtained. Dening
D
Dt
=
t
+
U
i
x
i
, (22)
and Favre-averaging the momentum and species equations (2) and (4) gives
)
D
Dt
U
i
=
p)
x
i
x
j
T
ij
)
x
j
)
u
j
u
i
(23)
)
D
Dt
=
x
j
J
j
) +)
x
j
)
u
, (24)
in which the last terms of the RHSs of the equations represent the Reynolds
stress and the Reynolds ux which occur in unclosed form. These terms contain
second moments of the velocity distribution and joint velocity-scalar distribu-
tion respectively, and cannot be expressed in terms of the rst moments or
means. Another unclosed term in the equations is the averaged reaction term.
Reaction rates are in general highly non-linear functions of composition and
temperature and the averaged reaction rate cannot be expressed as a function
of mean concentrations.
As a framework for turbulence modelling we now consider second moment
closure (SMC; Launder et al. 1975, Lumley 1980). The simpler eddy-viscosity
[10] Reacting ows and probability density function methods 333
models are still widely used for reacting ow computations and after some
modications they perform reasonably well for simple jet ows. However, in
recent years the increased computer performance has made the use of SMC
also for reacting ow computations (Jones 1994) more tractable. Also, using
a hybrid Monte Carlo method to compute the joint velocity-scalar PDF (see
[21]), it is preferable to use SMC from a theoretical point of view.
The conservation equations for the Reynolds stresses and Reynolds uxes
can be derived by standard methods from equations (2) and (23). A common
approach to modeling of variable-density ows is to apply the constant-density
second-moment closure models to variable-density ows simply by recast-
ing the Reynolds-averaged terms into Favre averages. However, the variable-
density second-moment equations contain additional terms, containing
U
k
/x
k
, u
i
or
D
Dt
i
u
j
=
u
i
u
U
j
x
k
j
u
U
i
x
k
(25a)
_
u
i
p
x
j
+u
j
p
x
i
2
3
ij
u
k
p
x
k
_
(25b)
ij
(25c)
x
k
_
i
u
j
u
k
+
2
3
ij
u
k
p
_
(25d)
u
i
p
x
j
u
j
p
x
i
(25e)
+
2
3
ij
p
k
x
k
. (25f)
Here the Reynolds average is denoted by an overbar and the combined use of
an overbar and tilde, as in
x, is used to denote the Favre average of a long
expression. The terms on the RHS are: (25a) the production by mean shear
P
ij
, (25b) the pressure-strain correlation
ij
, (25c) the viscous dissipation
ij
,
(25d) the turbulent ux T
ij
, and two terms which are zero in constant density
ows containing (25e) a mean pressure gradient, and (25f) the trace of the
uctuating strain tensor.
The production terms are in closed form whereas the uctuating pressure,
dissipation, turbulent ux, and uctuating density terms have to be modeled.
334 Roekaerts
The viscous dissipation
ij
is modeled by assuming local isotropy at the small-
est scales where viscous dissipation takes place. The dissipation model then
reads:
ij
=
2
3
ij
. (26)
For the dissipation of turbulent kinetic energy the standard dissipation equa-
tion is solved. The triple correlation terms
u
i
u
j
u
k
present in the ux terms
can be modelled by a generalized gradient diusion model that reads
i
u
j
u
k
= C
s
k
k
u
i
u
j
x
l
, (27)
where the constant C
s
has a value around 0.25. Modeling of the uctuating
density terms can be found in Jones (1980), Jones (1994). The nal unclosed
term is the pressure-strain redistribution term which has been, and probably
will remain, the focal point of Reynolds-stress modeling. This term does not
produce or destroy turbulent kinetic energy but only redistributes energy over
the components of the stress tensor. Its modelling will be discussed in [21].
3.2 Reynolds-ux and scalar-variance equations
Assuming only one scalar variable, the equations for the turbulent scalar ux
or Reynolds ux
u
respectively read
D
Dt
=
u
U
i
x
j
i
u
x
j
(28a)
x
i
(28b)
x
j
j
u
(28c)
p
x
i
(28d)
+
u
i
S(), (28e)
and,
D
Dt
= 2
u
x
j
(29a)
(29b)
x
j
(29c)
+2
S(). (29d)
[10] Reacting ows and probability density function methods 335
The terms on the right-hand sides of these equations are, in analogy with the
Reynolds-stress equations, the production terms (28a) and (29a), the pres-
sure scrambling term
i
(28b), the viscous dissipation of scalar variance
(29b), the turbulent uxes (28c) and (29c) and an additional mean pressure-
gradient term which is zero in constant density ows (28d). Furthermore, for
a reacting scalar, unclosed reaction source terms (28e) and (29d) appear in
the equations.
The dissipation rate of scalar variance g =
g
= C
k
. (30)
The empirical constant C
t
+U
i
x
i
=
x
j
_
ID
x
j
_
, (31)
with zero chemical source term. In other words, mixture fraction is a conserved
scalar. The relation between the physical scalar variables and mixture fraction
is given by the specic conserved-scalar chemistry model used (mixed-is-burnt
model, equilibrium model). In a more rened description the inuence of local
336 Roekaerts
ow conditions (hence nite Da eects) are taken into account by assuming
that the chemical composition can be retrieved from that of a strained laminar
amelet. Then scalar dissipation rate enters as a second independent variable.
A detailed explanation can be found in Peters (1984) and Peters (2000).
Even when chemistry can be described fully by mixture fraction, the thermo-
chemical variables are still non-linear functions of mixture fraction. To accu-
rately describe the mean thermo-chemistry, the turbulent mixture fraction
uctuations have to be known. A common way to model these uctuations is
by the assumed-shape PDF method. In this method, the scalar PDF of is
modeled as a known function of several of its lower moments.
For a detailed study of assumed shape PDF methods including an overview
of possible assumed PDF models the reader is referred to Peeters (1995).
Nowadays in most studies the -function PDF model is used. The -function
is selected because it can, as a function of its parameters, take various forms
that resemble physically realistic scalar PDFs (e.g. single-delta function PDFs
in fuel or oxidizer streams, or Gaussian-like PDFs in well mixed situations).
Assumed-shape PDF methods are useful for modeling turbulent reacting
ows with single-conserved-scalar chemistry and amelet models. In a situa-
tion with multiple reacting scalars, chemistry can inuence their joint PDF
shapes dramatically and scalar correlations can have large inuence on the re-
action rates. Attempts to model the joint scalar PDF in a functional assumed-
shape form have not been very successful. Use of multi-scalar chemistry models
requires an accurate description of the joint scalar distribution. This detailed
statistical information can be obtained by solving the joint scalar PDF trans-
port equation by means of a Monte Carlo method. This approach will be
addressed in [20].
An alternative method for closure of the mean reaction rate equations re-
lated to both PDF and laminar amelet methods is conditional moment closure
(CMC). It predicts the conditional averages and higher moments of scalar vari-
ables, with the condition being the value of the mixture fraction. A complete
description can be found in Klimenko and Bilger (1999).
5 One point joint velocity-scalar PDF closure
Knowledge of the joint scalar PDF is sucient to close the mean reaction
rate but to model the joint scalar PDF equation an assumption has to be
introduced on the correlation between velocity uctuations and uctuations
in the scalar PDF. This can be seen as a generalisation of the closure problem
of the Reynolds ux appearing in equation (24). This closure problem would
be absent if the joint velocity-scalar PDF would be known. Indeed knowledge
of the joint velocity-scalar PDF would at once imply a closure of all unknown
terms in the mean transport equations: the Reynolds stress, the Reynolds ux
and the mean chemical source term. This observation and the fact that a new
class of elegant Lagrangian solution algorithms can be used make calculation of
[10] Reacting ows and probability density function methods 337
the velocity-scalar PDF an attractive alternative that will be further explored
in [21].
References
Dopazo, C. (1994) Recent developments in PDF methods. In Turbulent Reacting
Fows, P. Libby and F. Williams (eds.), Academic Press, 375474.
Fox, R.O. (1996) Computational methods for turbulent reacting ows in the chemical
process industry, Revue de lInstitut Francais du Petrole, 51(2), 215243.
Jones, W.P. (1980) Models for turbulent ows with variable density and combustion.
In Prediction Methods for Turbulent Flows, W. Kollmann (ed.), Hemisphere, 379
421.
Jones, W.P. (1994) Turbulence modelling and numerical solution methods for vari-
able density and combusting ows;. In Turbulent Reacting Flows, P. Libby and
F. Williams (eds.), Academic Press, 309374.
Klimenko, A.Y. and Bilger, R.W. (1999) Conditional moment closure for turbulent
combustion, Prog. Energy Comb. Sci., 25, 595687.
Launder, B.E., Reece, G.J. and Rodi, W. (1975) Progress in the development of a
Reynolds-stress turbulence closure, J. Fluid Mech., 68, 537566.
Lumley, J.L. (1980) Second-order modeling of turbulent ows. In Prediction Methods
for Turbulent Flows, W. Kollmann (ed.), Hemisphere, 131.
Peeters, T.W.J. (1995) Numerical Modeling of Turbulent Natural-Gas Diusion
Flames. PhD thesis, Delft Universitity of Technology.
Peters, N. (1984) Laminar diusion amelet models in non-premixed turbulent com-
bustion, Prog. Energy Comb. Sci., 10, 319339.
Peters, Norbert (2000) Turbulent Combustion. Cambridge University Press.
Pope, S.B. (1985) PDF methods for turbulent reactive ows, Prog. Energy Comb.
Sci., 11, 119192.
Pope, S.B. (2000) Turbulent Flows. Cambridge University Press.
Williams, F.A. (1985) Combustion Theory, second edition. Benjamin/Cummings.
Part B.
Flow Types and Processes and
Strategies for Modelling them
11
Modelling of Separating and Impinging
Flows
T.J. Craft
1 Introduction
Flows involving separation, reattachment and impingement occur widely in
many diverse engineering applications. A number of cooling and drying pro-
cesses rely on the high heat transfer rates that can be obtained by impinging
uid onto a solid surface. Separation and reattachment are, of course, found
in numerous situations, including external aerodynamics, ow over obstacles
and internal ow through ducts and pipes with rapidly varying cross-section
or ow direction. Many of these internal ows are also associated with heat
transfer, such as internal cooling passages for gas-turbine blades.
Since heat transfer rates are predominantly determined by the ow be-
haviour in the immediate vicinity of the wall, it is often necessary to employ
low-Reynolds-number turbulence models, which can adequately resolve the
near-wall region, when computing applications involving wall heating or cool-
ing. However, the turbulence mechanisms near reattachment or impingement
zones are signicantly dierent from those found in simple shear ows where
most turbulence models have been developed. In particular, as will be seen,
the popular based models are often found to predict extremely large length-
scales in such ows, leading to the prediction of excessive heat transfer rates
and the necessity of including additional modelling terms to correct for the
defect. Furthermore, the irrotational straining found in impinging ows ex-
poses a number of weaknesses in both eddy-viscosity based models and in
widely-used stress transport closures.
As an example of the problems encountered in computing impinging ows,
Figure 1 shows the predicted and measured Nusselt number, plotted against
radial distance from the stagnation point, in an axisymmetric impinging jet
ow studied experimentally by Baughn and Shimizu (1989) and Cooper et al.
(1992). The jet issues from a long length of pipe, at a Reynolds number of
23000, and impinges perpendicularly onto a at plate at a distance of 2 jet
diameters from the pipe exit. The ow has been computed using a zonal mod-
elling approach, with a high-Reynolds-number stress transport scheme in the
fully turbulent region, and the LaunderSharma k- model in the near-wall
viscosity-aected region. Without additional modications, it can be seen that,
whilst the predictions are in agreement with the data at large radial distances
341
342 Craft
Figure 1: Nusselt number predictions in an impinging jet.
Without Yap term; - - - With Yap term; Symbols: experiments of
Baughn and Shimizu (1989).
(where the ow becomes a radial wall jet), the model fails, fairly spectacularly,
in the stagnation zone, overpredicting the Nusselt number by a factor of four
or more.
The following sections consider the problems encountered in the modelling of
impinging and reattaching ows, including a discussion of some of the dierent
solutions proposed, and example applications.
2 Turbulence Lengthscales in Impingement and
Reattachment Regions
In an equilibrium boundary layer, the turbulence lengthscale grows linearly
with distance from the wall, and the widely-used k- model has been tuned to
reproduce this behaviour in such a ow. However, in non-equilibrium situations
(for example, where there is separation and reattachment, or impingement) the
model is known to return signicantly larger lengthscales (Yap 1987). Figure 2,
taken from Craft (1991), shows the predicted normalized lengthscale l/(c
l
x)
(where l = k
3/2
/), plotted against distance from the wall x, at two radial
positions in the impinging jet for which heat transfer results were presented
in Figure 1. As can be seen, at a radial distance of 3 jet diameters, where the
ow resembles a simple shear ow, the predicted near-wall lengthscale is only
slightly larger than its equilibrium value. However, along the stagnation line
r/D = 0, the k- model in the near-wall region returns a lengthscale more
than six times greater than that found in an equilibrium boundary layer.
[11] Large modelling of separating and impinging ows 343
2.1 Lengthscale Correction Terms
Yap (1987) studied the case of ow and heat transfer through an abrupt pipe
expansion. He noted that the LaunderSharma k- model overpredicted heat
transfer rates around the reattachment point, and traced this to the fact
that it also returned very large turbulence lengthscales in this region. Fig-
ure 3 (taken from Yap 1987) shows the lengthscale predicted by the Launder
Sharma model, plotted against distance from the wall, in the reattachment
region, together with the linear equilibrium lengthscale that would be pre-
scribed in a 1-equation model. As can be seen, the standard equation returns
lengthscales signicantly larger than those found in an equilibrium situation
and, as a result, the heat-transfer (Figure 4) is predicted considerably too high
around the reattachment point.
To remedy this excessive lengthscale prediction Yap proposed adding an
extra source term in the equation:
Y
= max
_
_
0.83
2
k
_
k
3/2
/
2.5y
1
__
k
3/2
/
2.5y
_
2
, 0
_
_
(2.1)
where y is the distance to the wall.
If the predicted lengthscale l = k
3/2
/ is larger than the equilibrium value
of 2.5y, Y
R
t
)] +B
c
l
R
t
exp(B
R
t
) (2.2)
with c
l
= 2.55 and B
2
k
F(F + 1)
2
, 0
_
(2.5)
They successfully applied this term to the prediction of heat transfer in a num-
ber of rib-roughened ducts and channels, and Figure 5 shows an example of the
predicted Nusselt number in a ribbed pipe, using the LaunderSharma scheme
with both the original Yap correction and their own proposal, equation (2.5).
2.2 Alternative Lengthscale Equations
To avoid the lengthscale prediction problems associated with the equation,
one alternative is to consider the use of a variable other than as the subject for
the lengthscale-determining transport equation. Wilcox (1988, 1991) proposed
a k- model (where /k) which appeared attractive in that it had fewer
additional near-wall terms than those required in the equation, and was
claimed not to need such ne near-wall grids (although does have the rather
unappealing property that it goes to innity at a wall). Soalides (1993) tested
346 Craft
Figure 5: Predicted heat transfer in a ribbed pipe, using the LaunderSharma
k- model. with Yap term; with equation (2.5).
this model in an impinging jet ow and, although the results suered from the
failures associated with the linear EVM stress-strain relation (see later section
on impingement problems), it did return qualitatively the correct shape for the
heat transfer prole in the stagnation region without any additional lengthscale
correction terms. He then retuned the modelled equation, employing it in
conjunction with a non-linear stress-strain relation, and Figures 6 and 7 (taken
from his dissertation) show fully-developed pipe ow results and heat transfer
predictions in the impinging jet. For comparison, it should be noted that if
the -based non-linear eddy-viscosity model of Suga (1995) (for which results
will be presented later) is used without the Yap lengthscale correction in the
equation, the predicted heat transfer shows a signicantly high peak value at
the stagnation point, in contrast to the very at prole returned by the model
of Soalides. Although the model was not widely tested, and was subsequently
found to perform rather poorly in transitional ows, the results reproduced
here do show some promise, suggesting that this may be an area deserving of
further attention.
3 Turbulence Energy Production in Impingement
Regions
Although the addition of a lengthscale correction to the equation can be seen
in Figure 1 to bring a signicant improvement to the prediction of stagnation
heat transfer, the predicted Nusselt number is still too high by almost a factor
of 2, indicating that further modelling renements are needed for this type of
ow situation.
The modelling strategy employed in the calculations of Figure 1 was a zonal
approach, with a high-Reynolds-number RSM coupled to a near-wall low-
[11] Large modelling of separating and impinging ows 347
Reynolds-number k- model. Although one might expect the high predicted
levels of heat transfer to be largely due to weaknesses in the k- model, since
the heat transfer will be strongly aected by the predicted near-wall turbu-
lence, it will be seen later that there are, in fact, signicant weaknesses in both
the k- model and the stress transport model employed in these calculations.
3.1 Weaknesses of the Eddy-Viscosity Formulation
Figure 8 shows corresponding heat transfer results in the impinging jet ow us-
ing four dierent turbulence models: the LaunderSharma k- model through-
out the ow domain; the Basic linear RSM, with the LaunderSharma model
in the near-wall region, and two further RSMs, again employing the Launder
Sharma model in the near-wall layer. The Yap correction is included in all these
calculations and, from the k- results, it is clear that there is still a weakness
in this model, resulting in overprediction of heat transfer in the impingement
region. A reason for this overprediction can be seen from examining the wall-
normal and radial rms velocities v
and u
S
2
where S is the non-dimensional strain rate
S =
k
[2S
ij
S
ij
]
1/2
and S
ij
=
1
2
_
U
i
x
j
+
U
j
x
i
_
(3.4)
To improve predictions they investigated replacing this form of generation by
P
k
= c
S (3.5)
where
=
k
[2
ij
ij
]
1/2
and
ij
=
1
2
_
U
i
x
j
U
j
x
i
_
(3.6)
In a simple shear the strain and vorticity parameters, S and , are identical.
However, in an irrotational strain, such as that found in an impinging ow,
vanishes which will lead to very low production of k, thus reducing the
predicted levels of k in this region.
Figure 11 (taken from Kato and Launder 1993) shows predictions of k along
the centerline of the cylinder using the k- model both with wall-functions, and
with the k- model as a near-wall layer model. When the standard production
term is employed, a substantial increase of k is seen upstream of the cylinder,
which is not present with the modied production of equation (3.5). This
replacement, coupled with a low-Reynolds-number treatment of the near-wall
region, also returns improved predictions downstream of the cylinder.
Although the modication does appear to dramatically improve impinging
ows it is, unfortunately, not universally helpful: in further testing, Suga (1995)
discovered that the replacement of S by in the generation rate worsened the
prediction of ow through a curved channel.
[11] Large modelling of separating and impinging ows 351
3.3 Durbins Modication
Durbin (1996) suggested an alternative method of remedying the production
rate of k in a stagnation region. As noted above, if the eddy-viscosity formu-
lation is employed, then the generation rate can be written as P
k
= c
S
2
.
Durbin argued that towards the stagnation point the timescale k/ becomes
large, leading to high values of S, and the quadratic dependence of P
k
on S
leads to the observed high values of k. By considering what constraint it would
be necessary to impose on the turbulent viscosity in order to prevent one of the
predicted normal stresses from becoming negative in an axisymmetric strain
eld, he suggested that the turbulent viscosity should be modelled as
t
= c
kT (3.7)
where the timescale T is taken as
T = min
_
k
,
2
3c
3
4[(S)
2
[
_
(3.8)
for some tunable constant . In regions where the strain rate S is large, such as
in the stagnation region, the second term in equation (3.8) prevents excessive
growth of T, thus limiting the generation rate of k, which can now be written
as
P
k
= c
k
T
_
S
2
(3.9)
Figure 11: Turbulent kinetic energy along the centerline of a square cylin-
der. k- /k- with modied production; - k- /k- with standard
production; k- with wall-functions; - - - 2nd moment closure with wall-
functions (Franke and Rodi 1991); Symbols: experiments.
352 Craft
When the second term of equation (3.8) is active in T, the generation rate
now depends only linearly on strain rate S, and Durbin demonstrated that
such a treatment can avoid the excessive generation rates otherwise found in
stagnation ows.
4 Application of Higher-Order Models to Impinge-
ment Flows
4.1 Non-Linear Eddy-Viscosity Models
The underlying weakness with the linear EVM in a stagnation ow was seen
above to be associated with the predicted normal stress anisotropy. A non-
linear EVM, which includes higher order terms in the stress-strain relation,
does, of course, have the potential to return improved predictions of the nor-
mal stresses and thus to improve the prediction of k in the stagnation region,
via a better representation of its generation rate. In his development of cu-
bic non-linear eddy-viscosity models, Suga (1995) considered the case of ow
impingement, including the axisymmetric impinging jet amongst the ows
he employed for model tuning. The examples here show results both for the
2-equation NLEVM developed by Suga (Craft et al. 1996), and also his 3-
equation version, where a transport equation was solved for the anisotropy
invariant A
2
in addition to those for k and (Craft et al. 1997). Figure 12
shows the predicted normal stresses along the stagnation line, showing that the
NLEVMs are capable of returning the correct levels of turbulence energy in
this region, leading to improved heat transfer predictions shown in Figure 13.
It is worth noting that even with the improved normal stress predictions,
both of these models also include a term similar to the Yap lengthscale correc-
tion in the equation to improve the heat-transfer predictions. In further work
on the use of the 2-equation NLEVM referred to above, Craft et al. (1999)
demonstrated that a form of lengthscale correction similar to that outlined
in equation (2.5) could be employed with the model, yielding a completely
geometry-independent model that still returned predictions as good as those
shown in Figure 13.
4.2 Reynolds Stress Transport Models
Since stress transport models do not employ the eddy-viscosity formulation for
u
i
u
j
, they do not share quite the same weaknesses in impinging ows as those
encountered above with EVMs. However, equation (3.1) does indicate that if
the turbulence energy is to be correctly predicted, it is necessary for the model
to return the correct normal stress anisotropy. In a stress transport scheme
this suggests that the accurate modelling of the redistribution process
ij
can
be expected to be crucial in capturing such ows. As will be seen below, some
[11] Large modelling of separating and impinging ows 353
widely-used stress models fail to represent this process reliably in stagnation
ows.
In addition to the k- model results, Figure 9 also shows the predicted
normal stresses employing the Basic linear RSM, with the Gibson and Launder
Figure 12: Wall-normal rms velocity component predictions along the stagna-
tion line of an impinging jet using NLEVMs (From Suga 1995). 3-equation
NLEVM; - 2-equation NLEVM; LaunderSharma k- ; Symbols:
measurements of Cooper et al. (1992).
Figure 13: Impinging jet heat transfer predictions using NLEVMs. (From
Suga 1995) 3-equation NLEVM; - 2-equation NLEVM; Launder
Sharma k- ; Symbols: measurements of Baughn and Shimizu (1989).
354 Craft
(1978) wall-reection terms:
ij1
= c
1
(u
i
u
j
/k
2
/
3
ij
) (4.1)
ij2
= c
2
(P
ij
1
/
3
P
kk
ij
) (4.2)
w
ij1
= c
1w
k
(u
l
u
k
n
l
n
k
ij
3
/
2
u
i
u
k
n
j
n
k
3
/
2
u
j
u
k
n
i
n
k
) f
y
(4.3)
w
ij2
= c
2w
(
lk2
n
l
n
k
ij
3
/
2
ik2
n
j
n
k
3
/
2
jk2
n
i
n
k
) f
y
(4.4)
where n
k
is the unit vector normal to the wall, f
y
= l/(2.5y), l is the turbulence
lengthscale k
3/2
/ and y the distance to the wall.
Since the above stress model is only valid at high-Reynolds-numbers, it is
used with the LaunderSharma k- model in the near-wall region, and conse-
quently the very near-wall normal stresses are not shown in the gure. How-
ever, along the stagnation line it is clear that v
2
, the stress normal to the wall,
is signicantly overpredicted by the stress model in fact returning results
only slightly better than those predicted by the linear EVM and leads to
the overprediction of heat transfer seen in Figure 8.
In this case the failure can be traced to the wall-reection model employed
in
w
ij2
. In a simple shear ow (with mean strain U/y), energy is generated
in u
2
by P
11
. A proportion is transferred into v
2
by
222
=
1
/
3
c
2
P
11
, whilst
the wall-reection term
w
222
opposes this transfer, thus leading to the de-
sired damping of v
2
. In an impinging ow, however, v
2
is itself generated by
P
22
= 2v
2
V/y. The pressure-strain element
ij2
then removes some of
this generation and redistributes it into the other stress components. The
wall-reection contribution
w
222
= 2c
w
222
f
y
, in opposing this redistribu-
tion, thus acts to increase v
2
as the wall is approached, and it is to a large
extent this which leads to the overprediction of v
2
seen in Figure 9.
To overcome the above problem, Craft and Launder (1992) proposed an
alternative wall-reection model to be used in place of the above
w
ij2
. They
considered all possible combinations of terms linear in the Reynolds stresses
and mean strains, and included wall-normal vectors to give directional sensi-
tivity to the model. By ensuring that the resultant model behaved correctly in
both wall-parallel shear ows and impinging ows, they arrived at the form:
w
ij2
= 0.08
U
l
x
k
u
l
u
k
(n
t
n
t
ij
3n
i
n
j
) (l/2.5y)
0.1ka
lm
_
U
k
x
m
n
l
n
k
ij
3
/
2
U
i
x
m
n
l
n
j
3
/
2
U
j
x
m
n
l
n
i
_
(l/2.5y)
+0.4k
U
l
x
m
n
l
n
m
(n
i
n
j
1
/
3
n
k
n
k
ij
) (l/2.5y) (4.5)
Predictions of stresses and heat transfer using this model can be seen in
Figures 8 and 9. Clearly the modied
w
ij2
has a signicant eect in the stagna-
tion region, reducing turbulence energy levels, and thus predicted heat transfer
rates, to values close to those measured experimentally.
[11] Large modelling of separating and impinging ows 355
The nal model combination shown in Figures 8 and 9 is the TCL model,
described in [3]. In these calculations it has been implemented in its high-
Reynolds-number form, with the LaunderSharma k- model in the near-
wall region, and includes an additional wall-correction term similar to equa-
tion (4.5), but with slightly modied coecients. It also is seen to return
results in reasonable agreement with the experimental data.
As noted earlier, heat-transfer predictions in impinging ows are strongly
inuenced by the details of the near-wall turbulence modelling. If a full second-
moment closure is employed in the outer, fully turbulent, region, there may
be strong arguments for adopting a similar level of modelling in the near-wall
region (or, at any rate, using something more complex than a linear EVM). Al-
though the discussion of a TCL (two-component limit) model in [3] addressed
only the issues related to devising such a scheme for high-Reynolds-number
ows, the closure described has, more recently, been extended to account for
low-Reynolds-number and inhomogeneity eects found in near-wall regions.
The details are given in Craft and Launder (1996) and Craft (1998a), but
essentially involve the inclusion of a number of damping functions and in-
homogeneity corrections, which employ the turbulent Reynolds number and
the stress anisotropy invariants A and A
2
to account for viscous eects and
those due to high levels of turbulence anisotropy. Instead of employing the
wall-normal vector and distance, lengthscale gradients are used to sensitize
the model to regions and directions where there is strong inhomogeneity (such
as near a wall), without introducing any geometry-related quantities. The dis-
sipation rate equation employed is, again, an extension of that reported in
[3], and includes a lengthscale correction term of the form described in equa-
tion (2.5), but with a smaller coecient than that employed by Iacovides and
Raisee (1997) in their eddy-viscosity model.
The above low-Reynolds-number TCL closure has been applied by Craft
(1998b) to the impinging jet problem, and Figure 14 shows the predicted Nus-
selt number, which is seen to be in reasonable agreement with the experimental
data. The gure also shows the eect of neglecting the lengthscale correction
term in the equation, again highlighting the importance of such a term in
near-wall stagnation regions.
Two more complex applications, employing essentially the same low-Rey-
nolds-number TCL model are shown in Figures 15 and 16. The rst of these
concerns transonic ow (with a freestream Mach number of 0.875) over an
axisymmetric bump, studied experimentally by Bachalo and Johnson (1986).
Besides predicting the correct location of the shock at x/c 0.65, the com-
putation needs to resolve the separation bubble at the end of the bump which
causes the pressure plateau at around x/c = 1. The wall-pressure distribution
in Figure 15 shows the TCL results to be in better agreement with the data
than those of either the linear k- EVM or the Launder and Shima (1989)
stress transport model. The second case presented here is one of the afterbody
356 Craft
ows studied by Carson and Lee (1981). The geometry, shown in Figure 16,
resulted in signicant shock-induced ow separation from the boattail region
Figure 14: Nusselt number predictions in an impinging jet using the low-
Reynolds-number TCL stress model (Craft 1998b). TCL model with
lengthscale correction; - - - TCL model without lengthscale correction; Sym-
bols: measurements of Baughn and Shimizu (1989).
Figure 15: Wall-pressure distributions for the axisymmetric bump ow of
Bachalo and Johnson (1986) (From Batten et al. 1999b). SST: model of Menter
(1994); MLS: model of Launder and Shima (1989); MCL: Low-Re TCL closure.
[11] Large modelling of separating and impinging ows 357
Figure 16: IsoMach contours, boattail C
p
and internal nozzle p/p
T
jet
for the
Carson and Lee (1981) conguration 1 afterbody (From Batten et al. 1999a).
SST: model of Menter (1994); JH: model of Jakirlic and Hanjalic (1995); MCL:
Low-Re TCL closure.
of the afterbody. The gure shows that although the models tested by Batten
et al. (1999a) returned almost identical results in the interior throat section
(where the ow is governed by inviscid processes), the TCL model again re-
turned generally better C
p
values over the boattail than did the k- EVM and
358 Craft
Jakirlic and Hanjalic (1995) models. In both of the above cases it might be
noted that the SST model of Menter (1994) appears to give results very similar
to those returned by the TCL closure. The former model, however, employs
a linear eddy-viscosity relation for the stresses and so cannot be expected to
perform well in other complex ows, such as those where the eects of normal
straining are important. This can be seen in Section 3.4 of [5], which presents
the application of several models to the highly complex ow at Mach 2 around
a n/plate junction. It appears that only second-moment closures are able to
capture the separation and reattachment pattern ahead of the n, and the
above TCL model is again seen to return predictions in reasonable agreement
with the data.
References
Bachalo, W.D., Johnson, D.A. (1986), Transonic, turbulent boundary-layer separa-
tion generated on an axisymmetric ow model, AIAA J., 24, 437443.
Batten, P., Craft, T.J., Leschziner, M.A. (1999a), Reynolds-stress modeling of after-
body ows, in Turbulence and Shear Flow Phenomena 1 (S. Banerjee, J. Eaton,
eds.), Begell House, New York.
Batten, P., Craft, T.J., Leschziner, M.A., Loyau, H. (1999b), Reynolds-stress-
transport modeling for compressible aerodynamics applications, AIAA J., 37, 785
796.
Baughn, J.W., Shimizu, S. (1989), Heat transfer measurements from a surface with
uniform heat ux and an impinging jet, ASME J. Heat Transfer, 111, 10961098.
Carson, G.T., Lee, E.E. (1981), Tech. Rep. Technical report, NASA TP 1953.
Cooper, D., Jackson, D.C., Launder, B.E., Liao, G.X. (1992), Impinging jet studies
for turbulence model assessment: Part 1: Flow eld experiments, Int. J. Heat Mass
Transfer, 36, 26752684.
Craft, T.J. (1991), Second-moment modelling of turbulent scalar transport, Ph.D.
thesis, Faculty of Technology, University of Manchester.
Craft, T.J. (1998a), Developments in a low-Reynolds-number second-moment closure
and its application to separating and reattaching ows, Int. J. Heat Fluid Flow,
19, 541548.
Craft, T.J. (1998b), Prediction of heat transfer in turbulent stagnation ow with a
new second moment closure, in Proc. 2nd Engineering Foundation Conference in
Turbulent Heat Transfer, Manchester, UK.
Craft, T.J., Graham, L.J.W., Launder, B.E. (1993), Impinging jet studies for tur-
bulence model assessment. part ii: An examination of the performance of four
turbulence models, Int. J. Heat Mass Transfer, 36, 2685.
Craft, T.J., Iacovides, H., Yoon, J. (1999), Progress in the use of non-linear two-
equation models in the computation of convective heat-transfer in impinging and
separated ows, Flow, Turbulence and Combustion, 63, 5980.
[11] Large modelling of separating and impinging ows 359
Craft, T.J., Launder, B.E. (1992), New wall-reection model applied to the turbulent
impinging jet, AIAA J., 30, 29702972.
Craft, T.J., Launder, B.E. (1996), A Reynolds stress closure designed for complex
geometries, Int. J. Heat Fluid Flow, 17, 245254.
Craft, T.J., Launder, B.E., Suga, K. (1996), Development and application of a cubic
eddy-viscosity model of turbulence, Int. J. Heat and Fluid Flow, 17, 108115.
Craft, T.J., Launder, B.E., Suga, K. (1997), Prediction of turbulent transitional phe-
nomena with a nonlinear eddy-viscosity model, Int. J. Heat Fluid Flow, 18, 15.
Durbin, P.A. (1996), On the k-3 stagnation point anomaly, Int. J. Heat Fluid Flow,
17, 8990.
Franke, R., Rodi, W. (1991), Calculation of vortex shedding past a square cylinder
with various turbulence models, Paper 20.1, Proc. 8th Turbulent Shear Flows
Symposium, Munich.
Gibson, M.M., Launder, B.E. (1978), Ground eects on pressure uctuations in the
atmospheric boundary layer, J. Fluid Mech., 86, 491.
Hanjalic, K. (1996), Some resolved and unresolved issues in modelling non-
equilibrium and unsteady turbulent ows, in Engineering Turbulence Modelling
and Experiments, 3, W. Rodi, G. Bergeles (eds.), 318.
Iacovides, H., Raisee, M. (1997), Computation of ow and heat transfer in 2D rib
roughened passages, in Proceedings of the Second International Symposium on
Turbulence, Heat and Mass Transfer (K. Hanjalic, T. Peeters, eds.), Delft.
Jakirlic, S., Hanjalic, K. (1995), A second-moment closure for non-equilibrium and
separating high- and low-Re-number ows, 23.2323.30, Proc. 10th Turbulent
Shear Flows Symposium, Pennsylvania State University.
Kato, M., Launder, B.E. (1993), The modelling of turbulent ow around stationary
and vibrating cylinders, Paper 104, Proc. 9th Turbulent Shear Flows.
Launder, B.E., Shima, N. (1989), Second-moment closure for the near-wall sublayer:
development and application, AIAA J., 27, 13191325.
Menter, F.R. (1994), Two-equation eddy-viscosity turbulence models for engineering
applications, AIAA J., 32, 15981605.
Soalides, D. (1993), The strain-dependent non-linear k- model of turbulence and its
application, M.Sc. dissertation, Faculty of Technology, University of Manchester.
Suga, K. (1995), Development and application of a non-linear eddy viscosity model
sensitized to stress and strain invariants, Ph.D. thesis, Faculty of Technology, Uni-
versity of Manchester.
Wilcox, D.C. (1988), Reassessment of the scale determining equation for advanced
turbulence models, AIAA J., 26, 12991310.
Wilcox, D.C. (1991), Progress in hypersonic turbulence modelling, in Proc. AIAA
22nd Fluid Dynamics, Plasmadynamics and Laser Conference, Honolulu.
Wolfshtein, M. (1969), The velocity and temperature distribution in one-dimensional
360 Craft
ow with turbulence augmentation and pressure gradient, Int. J. Heat Mass Trans-
fer, 12, 301318.
Yap, C.R. (1987), Turbulent heat and momentum transfer in recirculating and im-
pinging ows, Ph.D. thesis, Faculty of Technology, University of Manchester.
12
Large-Eddy Simulation of the Flow past
Blu Bodies
W. Rodi
Abstract
The ow past blu bodies, which occurs in many engineering situations, is very
complex, involving often unsteady behaviour and dominant large-scale struc-
tures; it is therefore not very amenable to simulation by the RANS method
using statistical turbulence models. The large-eddy simulation technique is
more suitable for these ows. In this section work in the area of large-eddy
simulations of blu body ows is summarised, with emphasis on work by the
authors research group as well as on experiences gained from two LES work-
shops. Results are presented and compared for the vortex-shedding ow past
square and circular cylinders and for the ow around surface-mounted cubes.
The performance, the cost and the potential of the LES method for simulating
blu body ows, also vis-` a-vis RANS methods, is assessed.
1 Introduction
In many engineering situations, blu bodies are exposed to ow, generating
complex phenomena such as ow separation and often even multiple sepa-
ration with partial reattachment, vortex shedding, bi-modal ow behaviour,
high turbulence level and large-scale turbulent structures which contribute
considerably to the momentum, heat and mass transport. For solving practi-
cal problems, there is a great demand for methods for predicting such ows
and associated heat and mass-transfer processes, in particular the loading, in-
cluding dynamic loading on the bodies and the scalar transport in the vicinity
of structures. Usually the Reynolds number is high in practical problems so
that turbulent transport processes are important and must be accounted for
in a prediction method in one way or another. Until recently, mainly RANS
based methods were used in which the entire spectrum of the turbulent motion
is simulated by a statistical turbulence model. In vortex-shedding situations,
unsteady RANS equations are solved to determine the periodic shedding mo-
tion and only the superimposed stochastic turbulent uctuations are simulated
with the turbulence model. So far, mainly variants of the k- eddy-viscosity
turbulence model have been used for calculating the ow around blu bodies,
but some results have been reported that were obtained with Reynolds-stress
models.
361
362 Rodi
The RANS calculations have shown that statistical turbulence models have
diculties with the complex phenomena mentioned above, especially when
large-scale eddy structures dominate the turbulent transport and when un-
steady processes like vortex shedding and bistable behaviour prevail and dy-
namic loading is of importance. The large-eddy simulation (LES) approach
is conceptually more suitable in such situations as it resolves the large-scale
unsteady motions and requires modelling only of the small-scale turbulent mo-
tion which is less inuenced by the boundary conditions. An introduction to
the LES approach is provided in the companion chapter by Fr ohlich and Rodi
(2001). This approach is computationally more expensive than the RANS ap-
proach, but recent advances in computer performance and numerical methods
have made LES calculations feasible for ows around blu bodies (see discus-
sion on relative computer requirements in the concluding section). It should
be added here that proper direct numerical simulations (DNS), in which all
scales of the turbulent motion are resolved and no model is introduced, are
feasible only for relatively low Reynolds numbers as the number of grid points
required for resolution of all scales increases approximately as Re
3
. Calcula-
tions at higher Reynolds numbers have been reported which were called direct
simulations (e.g. Tamura et al. 1990, Verstappen and Veldman 1997) but some
of these were really quasi-LES calculations in which the task of the subgrid-
scale model to withdraw energy was taken over by a dissipative numerical
scheme (e.g. in the case of Tamura et al.).
The chapter presents and discusses LES calculations for the ow around
various blu bodies, in particular for ows around square and circular cylin-
ders and past surface- mounted cubes. The chapter cannot give an exhaustive
review of such calculations and is based mainly on the results presented at
several workshops and produced in the authors group. For the blu bodies
mentioned, the results obtained with various LES methods will be compared
with each other as well as with experiments and in some cases also with RANS
calculations. The capabilities, problem areas and the potential of LES calcu-
lations also vis-`a-vis RANS calculations will be discussed.
2 LES Methods Used
The LES approach is introduced in the companion chapter of Fr ohlich and
Rodi (2001) which summarises briey the most commonly used methods for
ltering, subgrid-scale modelling, near-wall treatment and numerical solutions.
Here a brief summary is given of the methods used for obtaining the blu-body
results presented in this chapter. In all cases, the resolved scales were dened
by the mesh size, i.e. no explicit ltering was used. Basically, the resolved
quantity is the average over a cell of the numerical mesh and this corresponds
to applying a top-hat lter with lter width equal to the mesh size.
The subgrid-scale stresses representing interactions between resolved larger
scales and unresolved smaller scales were mostly modelled explicitly by a
[12] Large-eddy simulation of the ow past blu bodies 363
subgrid-scale model; in some cases no model was used and the dissipative
eect of the stresses was achieved by numerical damping introduced by a
third-order upwind scheme for the convection terms. Many calculations were
carried out with the Smagorinsky (1963) eddy-viscosity model, which can be
considered the LES version of the Prandtl mixing-length model. Values of the
Smagorinsky constant in the range 0.1 to 0.2 have been used for the blu-
body calculations presented below. Near the wall, the length scale is modied
by a van Driest damping function. The popular dynamic approach of Ger-
mano et al. (1991) has also been used in a number of cases, mostly with
the Smagorinsky model as a base so that the approach then determines the
spatial and temporal variation of the Smagorinsky constant C
s
by making
use of the information available from the smallest resolved scales. Some cal-
culations were also carried out with a one-equation model as base in which
the velocity scale of the subgrid-scale stresses is calculated from a transport
equation for the turbulent subgrid-scale energy (Davidson 1997, Menon and
Kim 1996). Finally, mixed models combining a scale-similarity model based
on a double ltering approach with the Smagorinsky model have also been
applied.
Concerning near-wall treatment, often no-slip conditions have been used at
the wall, but at higher Reynolds numbers the resolution in this region was
then not good enough for a proper LES. In a number of higher-Reynolds-
number calculations also wall functions were applied, mainly the Werner
Wengle (1989) approach which assumes a distribution of the instantaneous
velocity inside the rst cell, namely a linear distribution for y
+
< 11.81 and a
1/7 power law for y
+
> 11.8.
The three-dimensional, time-dependent equations governing the resolved
quantities were mostly solved numerically by nite-volume methods; so far,
nite-element methods are used rarely. The nite-volume methods employ
either a staggered or non-staggered variable arrangement; some can accomo-
date general curvilinear grids, but for the ows around rectangular bodies
mostly Cartesian grids were used. Usually the grids were stretched in order to
achieve a better resolution in the near-wall region with high gradients in one
case embedded grids were used to better resolve this region. The schemes are
generally explicit with small time steps to resolve the turbulent uctuations
and various time discretization methods were employed mostly the second-
order AdamsBashforth scheme but also the Runge-Kutta, Euler and leap-frog
methods. For the discretization of the convection terms, mostly second-order
central dierencing was employed, but also the QUICK scheme and third-order
upwind dierencing (especially when no subgrid-scale model was used) and in
some cases also a fth-order upwind dierencing method. At the outlet the
calculations were all done with convective conditions.
364 Rodi
3 Flow Past Long Cylinders
The ow past long cylinders exposed to uniform approach ow is an interesting
test case because the geometry is simple, but the ow is complex with unsteady
separation. Alternating vortices are shed from the cylinder and transported
downstream, where they retain their identity in a K arman vortex street for a
considerable distance. These vortices are predominantly two-dimensional and
so is the time-mean ow, but large-scale three-dimensional structures exist
which lead to a modulation of the shedding frequency. The approach stagna-
tion ow is basically inviscid and thin laminar boundary layers form on the
forward surfaces of the cylinder. Square and circular cylinders are considered
here. In the case of the square cylinder, the ow separates at the front edges
and a apping shear layer develops on the sides of the cylinder, which is ini-
tially laminar but becomes turbulent fairly quickly. In the case of the circular
cylinder, the separation point is not xed but depends on the boundary layer
development before separation. For the subcritical cases considered here, the
boundary layer remains laminar up to separation, and again the separated
shear layer becomes turbulent fairly quickly.
3.1 Square Cylinder
For the square cylinder, the only detailed experiments with phase-resolved
measurements are those reported for Re = 22000 by Lyn et al. (1995) and Lyn
and Rodi (1994). The situation examined by them has become the standard
test case for this ow. RANS calculations obtained with various turbulence
models ranging from the algebraic BaldwinLomax model to a Reynolds-stress
model (RSM) and also a single LES have been reviewed already in Rodi (1993);
this LES suered from a small calculation domain of only 2 cylinder widths D
in the spanwise direction. The ow was then posed as a test case for an LES
workshop held in 1995 in Rottach-Egern, Germany, and 9 groups submitted
16 dierent results that are reported in Rodi et al. (1995) and partly also in
a summary paper of Rodi et al. (1997). These calculations were all performed
on a computation domain of 4D in the spanwise direction, extending 4.5D
upstream of the cylinder, 6.5D on either side of the cylinder (where the tunnel
walls were located) and at least 14.5D downstream of the cylinder. As will
be shown shortly, there was a great variance in the results, and hence the
same test case was posed again for a workshop held at Grenoble, France, in
1996. The same calculation domain was prescribed as in the earlier workshop.
7 groups presented 20 results to the Grenoble workshop (there is an overlap
of 2 sets of results from the authors group University of Karlsruhe) and
these are summarised in Voke (1997). In the following, some sample results
from both workshops are presented and compared with each other and with
the experiments. The samples are the same as those in the summary reports
of Rodi et al. (1997) and Voke (1997) and were chosen to represent the variety
[12] Large-eddy simulation of the ow past blu bodies 365
of methods and to concentrate on those results in which the statistical values
were determined by averaging over a sucient number of shedding cycles.
Later results of Sohankar (1998) (see also Sohankar et al. 1999) obtained with
3 dierent subgrid-scale models are added.
Some details on the various methods like subgrid-scale model, near-wall
treatment and grids used are given in Table 1. The labelling was taken over
from Rodi et al. (1995) for the Rottach-Egern workshop and from Voke (1997)
for the Grenoble workshop (see key at end of text). UKA2, UK1 and UK3
are from the authors group using dierent grids and near-wall treatment and
hence show the inuence of these.
Figure 1 shows samples of streamlines at phase 1 of the shedding cycle in
comparison with the experimental streamlines. The shedding motion is quali-
tatively well reproduced, with a vortex that has just shed from the lower side
of the cylinder at the phase considered. However, there are considerable dier-
ences in the details such as the size and strength of the shed vortex and also in
the recirculation motion near the side surfaces. Unfortunately, in this region
near the walls the streamlines could not be obtained from the measurements
so that the complex behaviour there with a number of vortices appearing can-
not be assessed. The shedding in the experiments and also generally in the
LES calculations is not very regular, as can be seen from excerpts of pressure
and lift signals shown in Figure 2. Clearly, lower frequency amplitude varia-
tions are present, but a clear shedding frequency could still be determined.
The low frequency variations are believed to be due to three-dimensional ow
structures. In the two-dimensional RANS calculations these eects cannot be
simulated and hence the shedding behaviour is generally regular.
Table 1 summarises various global parameters such as the dimensionless
shedding frequency (Strouhal number St = fD/U
0
), the time-mean drag co-
ecient c
D
, the RMS values of the uctuations of drag and lift coecients c
D
and c
L
, respectively, and the reattachment length l
R
indicating the length of
the time-mean separation region behind the cylinder. Most LES calculations
yielded the correct value of St 0.13, and it appears that St is not very sensi-
tive to the parameters of the simulation; there are, however, a few deviations
from this value, notably the calculations without subgrid-scale model yielded
a higher value. Concerning the mean drag coecient, it seems that the LES
calculations using wall functions are generally close to the experimental range
while those using no-slip conditions tend to produce too high values of c
D
.
There is also considerable variation in the recirculation length which will be
discussed in connection with the centre-line velocity distribution in Figure 3.
The uctuations of the force coecients also show fairly large variation here
no experimental results are available for a comparison.
Table 1 also includes results obtained with RANS models, namley by Bosch
(1995) with the standard k- model and with a modication due to Kato
and Launder (1993), and by Franke and Rodi (1993) with the Reynolds-stress
366 Rodi
Table 1: Global parameters for ow past square cylinder (UW = upwind di.,
CD = central dierencing, WF = wall function, NS = no slip;
(1)
= adjusted
for dierent blockage;
(2)
= outermost mesh with embedded meshes).
Calculation method St C
D
C
D
C
L
L
R
/D grid
N
x
N
y
N
z
LES KAWAMU, 3rd UW 0.15 2.58 0.27 1.33 1.68 125 78 20
No SGS, NS
Rottach- UMIST2, CD 0.09 2.02 1.21 140 81 13
Egern Dyn, WF
Rodi et al. UKAHY2, CD 0.13 2.30 0.14 1.15 1.46 146 146 20
(1997) Smag, WF
TAMU2, Dyn, 3rd UW 0.14 2.77 0.19 1.79 0.94 165 113 17
NS/WF
UK1, Smag, CD 0.13 2.20 0.14 1.01 1.32 109 105 20
WF
UK3, Smag, CD 0.13 2.23 0.13 1.02 1.44 146 146 20
NS
LES NT7, LDM, CD 0.131 2.05 0.12 1.39 1.39 140 103 32
WF
Grenoble UOI, Dyn., 5th UW 0.13 2.03
(1)
0.18 1.29 1.20 192 160 48
NS
Voke IS3, Dyn. mix, 5th UW 0.133 2.79 0.36 1.68 1.36 112 104 32
NS
(1997) TIT, Dyn, 3rd UW 0.131 2.62 0.23 1.39 1.23 121 113 127
(2)
NS
ST5, Smag, variable 0.161 2.78 0.28 1.38 1.02 107 103 20
NS 3rd UW
LES Smag, NS 0.126 2.21 0.16 1.47
Sohankar Dyn, NS CD 0.125 2.04 0.20 1.22 1.0 185 105 25
(1998) OEDSM, NS 0.129 2.25 0.20 1.50
RANS Std. k-, WF 0.134 1.64 0 0.305 2.8 100 76
Bosch KatoLaunder 0.142 1.79 0.012 0.614 2.04 100 76
(1995) k-, WF
Two-layer k- 0.137 1.72 0 0.426 2.4 170 170
Two-layer 0.143 2.0 0.07 1.17 1.25 170 170
KatoLaunder
k-
RANS RMS, WF 0.136 2.15 0.27 1.49 0.98 70 64
Franke and Two-layer RSM 0.159 2.43 0.06 1.3 1.0 186 156
Rodi (1993)
Experiments 0.132 1.9 to 1.38
2.2
[12] Large-eddy simulation of the ow past blu bodies 367
Figure 1: Phase-averaged streamlines at phase 1 from some LES simulations
submitted to the Rottach-Egern Workshop (Rodi et al. 1995).
model of Launder et al. (1975). In each case, either wall functions (WF) or a
two-layer approach applying a one-equation model in the near-wall region was
used. The Strouhal number is predicted well also by most of the RANS models
tested, but the KatoLaunder modication (KL) tends to produce somewhat
too large values and the two-layer RSM an excessive value. The mean drag co-
ecient is signicantly underpredicted by the standard k- model, but roughly
the correct value was obtained when the KL modication was used with the
368 Rodi
Figure 2: Time variation of pressure and lift in ow past square cylinder.
two-layer approach. The RSM gives the correct value of c
D
with wall functions
but overpredicts it in the two-layer approach. This result is consistent with
the signicant overprediction of the length of the recirculation region by the
standard k- model and its underprediction by the two-layer RSM.
Figure 3 shows the distribution of the time-mean velocity along the cylinder
centre-line. The experimental velocity recovers very slowly in the downstream
region or even seems to level o at a value of about 0.6 of the upstream free-
stream level. The LES calculations exhibit a great variance in this region.
Most of them predict the recovery to the free-stream level considerably faster
than the experiments, but two (KAWAMU, TAMU2) level o at roughly the
correct value while UOI and ST5 show a decline of the centre-line velocity
beyond x/D 0.5 which seems physically not plausible. Looking now at the
near-cylinder region with recirculation zone, it can be seen that the calcu-
[12] Large-eddy simulation of the ow past blu bodies 369
lations with the lowest velocity values in the far-eld generally exhibit too
short a recirculation length. On the other hand, those calculations with too
fast recovery in the far-eld are in fairly good accord with the experimental
distribution in the near-eld up to x/D 2. The more recent calculations of
Sohankar (1998) exhibit a recirculation that is too short (l
R
1.0) and too
weak and the centre-line velocity levels o at too high a value of U ( 0.8).
There is no great dierence between the 3 subgrid-scale models used. The
calculations of Sohankar et al. (1999) with a dynamic one-equation model ob-
tained on a ner grid (265 161 25) are also not very dierent, but a little
closer to the experiment.
The RANS calculations all show too fast a recovery in the far-eld. The
standard k- model predicts a signicantly too long recirculation; in the near-
eld a fairly good prediction is achieved with the KL modication combined
with the two-layer approach while the recirculation length is underpredicted
with the RSM.
Figure 4 presents the distribution of the total resolved (periodic plus tur-
bulent) uctuating kinetic energy along the centre-line for the Rottach-Egern
workshop calculations plus the UOI results from the Grenoble workshop. Here
the various LES results show an even wider variation with an almost fourfold
dierence in the peak level of k
tot
, but the picture is not entirely consistent.
TAMU2 yields excessive uctuations which cause underprediction of the sepa-
ration length while KAWAMU produces too small uctuations which explain
the excessive separation length. It is dicult to understand why the UMIST2
calculations with an even lower uctuation level lead to an underprediction
in the separation length. It is generally to be observed that the total uctua-
tions are underpredicted while the drag coecient and separation length are
reasonable. The variance of the total uctuation results is somewhat smaller
for the Grenoble calculations but there is still a factor of about 2, as can
be seen from Figure 5 which shows the distributions of the (resolved) total
uctuation components and along the centre-line in comparison with Lyn et
al.s (1995) measurements. None of the calculations is entirely satisfactory. For
these quantities, the results of Sohankar (1998) obtained with the Smagorin-
sky model and with a one-equation dynamic model agree quite well with the
experiments. The RANS calculations are not very satisfactory concerning the
total uctuations. The KL k- model version with the two-layer approach,
which predicted best the velocity distribution, yields the correct shape of the
uctuating energy but overpredicts the level by about 40%.
Considerably worse are the RANS predictions of the turbulent kinetic en-
ergy component of the uctuations (periodic uctuations subtracted) as shown
in Figure 6. All RANS calculations strongly underpredict the k-level, while the
only LES result available for the turbulent component (UKAHY2) is roughly in
accord with the measurements. This means that in RANS calculations where
the total uctuations are realistic or too high, the periodic uctuations are
370 Rodi
Figure 3: Time-mean velocity u along centre-line of square cylinder. (a) Results
from Rottach-Egern Workshop (Rodi et al 1995). (b) Results from Grenoble
Workshop - complete wake; experiments, simulations: UK1 , UK3 O, UOI
*, IS3 +, NT7 , TIT , ST5 (from Voke 1997). (c) Results from Grenoble
Workshop - near wake; key as for (b) (from Voke 1997).
[12] Large-eddy simulation of the ow past blu bodies 371
Figure 4: Total kinetic energy of uctuations (periodic and turbulent) along
centre-line of square cylinder. Results from Rottach-Egern Workshop and UOI
from Grenoble Workshop.
overpredicted. The very dierent behaviour of RANS and LES calculations of
k is most likely due to the fact that the fairly high turbulent kinetic energy
stems from contributions of low frequency uctuations as indicated in Fig-
ure 2. In the experiments these originate from three-dimensional large-scale
structures. LES calculations can capture these structures and count any low-
frequency uctuations originating from them as turbulence, while of course
the two-dimensional RANS calculations cannot and determine k from solving
the turbulence model equations.
The overall behaviour of the vortex-shedding ow is determined largely by
the prediction of the evolution of the separated shear layers on the sides of
the cylinder. It was found from the workshop calculations (e.g. Rodi et al.
1997) that the very thin ( 0.1D) reverse-ow region near the wall is clearly
not well resolved. The calculations with the best resolution near the wall had
a mesh dimension of the rst cell adjacent to the wall of 0.01D, but many
used a coarser mesh so that sometimes the rst mesh point comes to lie at the
maximum of the negative velocity.
One of the main conclusions at both workshops was that higher resolution
is required at the walls. It is surprising that near the side walls of the cylinder,
the resolution was not improved from the rst workshop to the second (in both
cases the best resolution was (/D = 0.01). None of the methods produced
good agreement with experiments in the near wake with recirculation zone
and at the same time for the recovery of the ow in the far wake as well as
372 Rodi
Figure 5: Total u
and v
further
downstream as obtained by the Stanford group in their various calculations
using the dynamic SGS model and compares them with the measurements of
Ong and Wallace (1996). Even the simulations without subgrid-scale model
gave reasonable agreement for the mean quantities. Introducing a subgrid-
scale model has relatively little eect on the mean velocity and the global
parameters, but improves the agreement with experiments, especially for the
stresses and spectra, with the best improvement when the dynamic model is
employed. Breuer found that the best results are obtained with a ne grid
using the dynamic model, but these calculations produced a separation length
L
R
larger than measured by LS. This is conrmed by the Stanford group cal-
culations, but there are dierences in the velocity results which point to the
inuence of details of the numerical treatment. KM studied quite extensively
the grid eects and found that underresolved LES produced shorter separation
lengths which are in agreement with the LS experiment, and this is conrmed
by Breuers coarser-grid calculations. KM argue that any agreement of such
calculations with the LS experiment is fortuitous. They found that inadequate
grid resolution can cause early transition of the separated shear layers and that
in the LS experiment such early transition was probably caused by external
disturbances. They also found that the more accurate numerical treatment
achieved with the B-spline method improved the calculations further down-
stream and that the power spectra obtained with this method are in good
agreement with the measurements while this is not the case with other dis-
cretization methods.
The main message of these calculations is that the inuence of the subgrid-
scale model is not very strong at this relatively low Reynolds number (although
the spectra decay too slowly without a model) and that the best results are
obtained with a ne-resolution simulation using a dynamic SGS model, which
altogether are in good agreement with experiments. The DNS calculations of
Xia et al. (1998) indicate similarly good agreement, but for a fuller comparison
and assessment, more results of these calculations need to be reported.
As was indicated already, the cylinder ow in the upper subcritical regime
(Re = 140, 000) studied in detail experimentally by Cantwell and Coles (1983,
hereafter referred to as CC) is much more demanding. The boundary-layer
thickness behaves like 1/
at x/D = 1, calculations
of Breuer (1999) on 165 165 64 grid: A1 dynamic model; A2 Smagorinsky
model with C
S
= 0.1; A3 Smagorinsky model with C
S
= 0.065.
[12] Large-eddy simulation of the ow past blu bodies 379
Figure 11: Flow around surface-mounted cube according to Martinuzzi (1992).
well with the experiments. The predicted Strouhal number is higher than that
measured by CC, but is in reasonable agreement with other measurements at
a similar Reynolds number. In the far wake the velocity development is not
predicted so well which is most likely due to the O-grid being too coarse in this
region. Further, the v
S
= C
S
, instead
of C
S
.
3 Application to a Tube Bundle and Cylinder
The ow in a portion of a large tube bundle, where periodic boundary condi-
tions can be applied in all 3 directions, is an ideal case for LES (Rollet-Miet
et al. 1999). Industrial interest lies in vibrations caused by uid-structure
coupling or large temperature uctuations that eventually lead to thermal
fatigue. The geometry is relatively simple, yet the ow experiences complex
strains, making this an attractive test case. The experiment provides data on
mean velocities and Reynolds stresses for the ow through an staggered ar-
ray of tubes (see Figure 1). The Reynolds number based on the bulk velocity
in the sub-channel between the tubes and the tube diameter is 40000. This
ow was considered during two ERCOFTAC/IAHR workshops on Turbulence
Modelling and in both cases RANS models performed poorly.
A rst run was computed with the Smagorinsky model and, after initialisa-
tion, statistics were accumulated over three domain-through-ow times; this
was found to be sucient since span-wise averaging and symmetries are used in
gathering statistics. For the second run, with the localised dynamic procedure
of Piomelli and Liu (1994), statistics were computed over ve domain-through-
ow times.
[13] Large eddy simulation of industrial ows? 397
L
V
0
D
x
y
Figure 1: Computation sub-domain.
The value of the Smagorinsky constant was set to C
S
= 0.065, while the
dynamic model yielded a somewhat higher average value of this coecient,
with great spatial variability. Large values occurred in the separated shear
layers, and slightly negative values appear near the wall on the upstream half
of the tube.
For the mean velocity, the agreement of the LES with all available data is
excellent, and the results obtained from the two subgrid-scale models cannot
be distinguished. For the Reynolds stresses, the overall agreement is fairly
good. Figure 2 shows the results in the wake-to-impingement axis and on the
cross section just behind the tube, a region where RANS models performed
poorly. The striking and unusual feature observed on the wake axis is that
the transverse velocity uctuations are far larger than the axial ones. Thus
as the ow approaches the stagnation point, u
2
< v
2
, and since
U
x
< 0, the
production of the stresses and kinetic energy are
P
uu
= 2u
2
U
x
> 0; P
vv
= +2v
2
U
x
P
k
=
_
u
2
v
2
_
U
x
< 0,
where U is the mean velocity.
We notice from the experiment and LES that u
2
is fairly constant (production
balances dissipation) in the impingement region while v
2
decreases rapidly
due to dissipation and negative production. Since u
2
< v
2
, the kinetic energy
is also decreasing, in contrast to what is found at a stagnation point on a
single blu body, or an impinging jet. Very near the stagnation point, the
kinematic blockage by the wall forces u
2
uctuations to be converted into
v
2
(wall echo eect), and the dynamic model seems to capture the peak in v
2
better, although there is only one experimental point to suggest the existence
of such a peak.
On Figure 3 is shown the LES of a single cylinder at Re = 3900. The un-
structured N3SLES mesh of Kessler et al. uses only about one tenth of the
nodes of the structured mesh adopted by Fr ohlich et al., but the extra cost of
using unstructured grids in this case just about balances the saving in nodes.
398 Laurence
Figure 2: LES of the ow across a tube bundle. LES with Smagorinsky model
(solid lines), LES with dynamic model (dashed lines), experiment (symbols).
As commented earlier, in an industrial environment, only a few tens of shed-
ding cycles could be aorded, while Fr ohlich, with virtually unlimited access
to academic supercomputers, was able to conduct the averaging processes over
thousands of cycles. Still the mean ow features of the structured and unstruc-
tured LES are similar and in good agreement with the results of Beaudan and
Moin (1994).
4 Fluid-Structure Coupling
If industry decides to invest in LES, it is probably not for the higher accuracy
of LES predictions in simple ows, but rather for specic problems where it
seems to be the only approach. For instance, in a nuclear power plant the
control rods, which act as brakes, are very long and thin and they could be
[13] Large eddy simulation of industrial ows? 399
Figure 3: LES of the ow around a single Cylinder at Re = 3900 (Fr olich
et al. 1998). LESOCC nite volume code with 1,3 million nodes (left), and
N3SLES unstructured code with 176000 nodes (right). Streamlines for time
averaged ow showing secondary recirculations.
subject to ow induced vibrations; this is typically an area where the extra
cost of LES is negligible in view of the massive investment in safety studies.
The conguration described in Figure 4, is a simplied mock-up of the
actual geometry. A tube is submitted to a turbulent three-dimensional axial
ow crossing a perforated plate located just above the rod and generating high
turbulent uctuations near the tube. Fluid-structure eects are restricted to
added mass eects due to the uid (in the LES the tube is assumed to be a solid
non-xed tube). The calculation by Longatte et al. (2001) yields reasonably
good predictions of the spectrum of the uctuating uid-force on the control
rod (Figure 5).
5 Acoustic Source Terms
The right-hand side of the acoustic equation for pressure contains derivatives of
the instantaneous acceleration which are extremely dicult to model in terms
of RANS, but can be extracted from an LES and inserted into an acoustic
propagation code such as EOLE (the LES cannot directly predict the noise at
a distance from the source with sucient accuracy since the acoustic pressure
is a very small fraction of the calculated pressure). Figure 6 presents an LES of
the CLARINETTE experiment at EDF (Lafon 1997). A diaphragm is inserted
400 Laurence
400
948
260
150
150
254
Tube
15
Figure 4: Mock up of a control rod. Geometry on the left and instantaneous
velocity from LES on the right.
Figure 5: Turbulent force spectra ((N/m
)
2/Hz) computed numerically (lines)
and measured experimentally (points) at three dierent locations along the
tube at 0.3m, 0.6m and 0.9m from the plate.
in a rectangular channel and strong noise sources appear on the edges of the
jet. The noise predicted by N3SLES and subsequent acoustic propagation
calculation agree well with the experimentally established behaviour, namely
proportional to the fourth power of the bulk velocity. Moreover, for the chosen
ratio of diaphragm-to-channel width, a Coanda eect appears, i.e. an early
reattachment of the jet to one of the walls although the geometry is totally
[13] Large eddy simulation of industrial ows? 401
symmetrical. From the LES, an animation explains this phenomenon: once
the jet attaches to one side, an intense recirculation appears, and feeds back
perturbations into the lower edge of the jet (Figure 6). The mixing on this side
is then stronger than on the free jet side. In other words the friction of the
jet on the recirculation side is larger than on the free side thus deecting the
jet towards the lower wall. This research is part of the PREDIT programme, a
collaboration with the rail and car industry, and sponsored by the French gov-
ernment. Another geometry that has been studied consists of a wall-mounted
rectangular box (Figure 7). Here the acoustic sources are mainly located in the
shear layer separating from the leading edge, whereas contributions from the
trailing corner and recirculation are signicantly weaker. This is valuable in-
formation since, while the overall noise is easily measured, identifying precisely
the noise sources experimentally is much more challenging.
A similar geometry also simulated using N3SLES is the rib-roughened
channel encountered in turbine blade cooling. RANS models tended to show
some instability in this case, with the ow structure oscillating between two
and three recirculation bubbles. The time-averaged streamlines from the LES
exhibit three recirculations, but this structure could never be found on the
instantaneous elds (Figure 8). Owing to the high turbulence intensity devel-
oping on this innite series of ribs (periodicity is assumed), the ow between
ribs is extremely complex, with little relation to the time-averaged solution.
Time-averaged plots sometimes only represent the artefact of taking a statis-
tical average, signicantly dierent from the actual ow when the turbulence
intensity is high.
6 Toward Industrial LES
The previous applications are very simplied mock-ups of actual problems. As
explained in the introduction, for most engineering applications, it is usually
not possible to isolate a sub-domain, where LES might be attempted, from
the complex three-dimensional ow in which it is embedded (a car mirror,
a control rod in a nuclear vessel, an aircraft engine housing etc.). An LES
used as a zoom eect within a complex industrial apparatus would need to be
embedded and coupled with a global RANS calculation. A promising solution
is the Detached Eddy Simulation (DES) of Spalart (1999, 2000) whereby the
RANS eddy viscosity is forced to reduce to a subgrid scale viscosity locally
where a transition to LES is desired. The method is ecient for detaching
shear layers around blu bodies which exhibit a natural instability. A more
general method would consist of generating instantaneous structures using all
the information available from the RANS simulation, since, in general, it is
not sucient simply to reduce the viscosity and expect these structures to
appear, without a considerable space- or time-lag to enable small instabilities
to develop into actual turbulence. Such synthetic turbulence boundary condi-
tions are currently being developed, for example by Cabot and Moin (1999) for
402 Laurence
Figure 6: LES of the ow through a diaphragm (CLARINETTE mock-up),
instantaneous velocity levels, and lower half of the unstructured mesh (before
[13] Large eddy simulation of industrial ows? 403
Figure 7: LES of a wall mounted rectangular body.
Figure 8: LES of the ow across over a ribbed wall. Time averaged streamlines
(upper gure), and perspective view of the intensity of the instantaneous ve-
locity eld; the iso-surface in white/light grey corresponds to a zero value of
the streamwise velocity. On the near side, the ow reattaches early, showing a
hole in the iso-surface, while on the far side, the two recirculation bubbles are
connected into a single one lling the cavity between the ribs.
404 Laurence
Figure 9: The 2D velocity eld in the inow plane of the channel.
wall parallel boundaries, or by Sergent et al. (2000) for inow-like conditions
(Figure 9). In the latter, a set of random vortices is generated that match
the statistics given by the RANS calculation. Coherence of these synthetic
structures in space and time is essential for them to be sustained in the LES
calculation.
Such a RANS-embedded LES calculation could then be used in a manner
similar to the coastal engineering situation where it is customary to overlay a
local calculation of, for example, a harbour, into a larger scale simulation of
coastal currents and tides. Before it can be generalised and used on an indus-
trial basis, the method would need to be made automatic via optimal control,
that is, self-adjustment of the synthetic turbulence to match the statistics of
the RANS calculation in an overlap region.
Conclusions
It is now possible to apply LES to moderately complex geometry, at least
beyond homogeneous and channel ows, but, at the same time, industry has
progressed into applying RANS to real-life problems and it is not certain that
it will resort to LES in the near future, apart from specic problems such
as acoustics or uid-structure interaction. Progress concerning conservative
[13] Large eddy simulation of industrial ows? 405
properties of numerical schemes for unstructured grids is however promising,
as well as work on synthetic turbulence for boundary conditions or embed-
ded calculations that would allow LES to be used as a zoom eect in a region
of specic interest. CPU costs are decreasing, but sta time to conduct and
process the data of an LES is still a limiting factor, as well as turn-over time.
The applications presented above still remain academic test cases by indus-
try standards, and will require some years before a panorama of industrial
applications can be presented (without a question mark).
References
Archambeau, F., Laurence, D., Martin, A., Maupu, V., Pot, G. (1997). Rened tur-
bulence modelling for power generetion industry J. Hydraulic Research 35 (6),
749772.
Benhamadouche, S., Laurence, D. (2001). An Energy conserving scheme for LES on
unstructured grids. Int. Conf. on Num. Methods in Fluid Dynamics, Oxford.
Breuer, M. (1998). LES of the subcritical ow past a circular cylinder : numerical
and modeling aspects, Int. J. Numerical Methods in Fluids 28 1281-1302.
Beaudan, P., Moin, P. (1994). Numerical experiments on the ow past a circular
cylinder at sub critical Re number. Reports TF-62 Stanford University, Mechanical
Engineering Department.
Bonnin, J.C., Buchal, T., Rodi, W. (1996). Databases for testing of calculation meth-
ods for turbulent ows, ERCOFTAC Bulletin 28 4854.
Cabot, W., Moin, P. (1999). Approximate wall boundary conditions in the LES of
high Re number ow, Flow Turbulence and Combustion 63 269-291.
Dormy, E. (1999). An accurate compact treatment of pressure for colocated variables,
J. Computational Physics 151 676683.
Fr olich, J., Rodi, W., Kessler, P., Parappis, S., Bertoglio, J.-P., Laurence, D. (1998).
Large eddy simulation of ow around circular cylinders on structured and un-
structured grids. In Notes on Numerical Fluid Mechanics 66, E.H. Hirschel (ed.),
Vieweg 319338.
Germano, M., Piomelli, U., Moin, P., Cabot, W. (1991). A dynamic subgrid scale
eddy viscosity model, Phys. Fluids A. 3 17601765.
Kobayashi, M.H., Pereira, J.M.C, Pereira, J.C.F. (1999). A conservative nite-volume
second-order accurate projection methods on hybrid unstructured grids, J. Com-
putational Physics 150 4075.
Lafon, P. (1997). Noise of conned ows, EDF R&D report HP 53/97/032.
Longatte, E. et al. (2001). Application of large eddy simulation to a ow induced
vibration problem. Submitted to ASME PVP (Pressure Vessel & Piping Confer-
ence), Atlanta.
Mahesh, K., Ruetsch, G. R. and Moin, P. (1999). Towards LES in complex geome-
tries, Center for Turbulence Research Annual Briefs, 379-387.
406 Laurence
Martin, A., Alvarez, D., Cases, F. (1997). 3D calculations of 900Mw PWR Vessel,
Accurate RCP start-up Flow, 5th Int. Conference on Nuclear Engineering, Nice.
Rodi, W., Ferziger, J., Breuer, M., Pourquie M. (1997). Status of large eddy simula-
tion; results of a workshop, J. Fluid Eng. 119 248262.
Rollet-Miet, P., Laurence, D. (1995). Large eddy simulation with unstructured grids
and nite elements, Turbulent Shear Flow 10, Pennsylvania State University.
Rollet-Miet, P., Laurence, D., Ferziger, J. (1999). LES and RANS of turbulent ow
in tube bundles, Int. J. Heat & Fluid Flow 20 (3), 241254.
Rodi W., Bonnin, J.C., Buchal, T., Laurence, D. (1998). Testing of calculation meth-
ods for turbulent ows: workshop results for 5 test cases, Coll. Notes Internes
DER EDF 98NB00004, ISSN 1161-0611.
Sergent, E., Bertoglio, J.-P., Laurence, D. (2000). Coupling between LES and RANS,
Euromech Coll. 412, Munich.
Spalart, P.R. (1999). Strategies for turbulence modelling and simulations. In Eng.
Turbulence Modelling and Experiments 4, W. Rodi, D. Laurence (eds.), Elsevier,
317.
Spalart, P., Travin, A., Shur, M., Strelets, M. (2000). Physical and numerical upgrades
in the detached-eddy simulation of complex turbulent ows. Euromech Coll. 412,
Munich.
14
Application of TCL Modelling to Stratied
Flows
T.J. Craft and B.E. Launder
1 Introduction
Chapter [3] has provided the rationale and the associated analysis for replacing
the Basic Model of the pressure-strain process in second-moment closure by a
more widely applicable approach namely one satisfying the two-component
limit (TCL). The present chapter will show that extending that model to
include ow problems where gravitational eects are important also brings
clear benets in terms of accuracy and, sometimes, reductions in computing
times, too.
However, not all closure problems in stably-stratied ows can be resolved
merely by ensuring that the non-dispersive pressure-containing correlations
satisfy the two-component limit. If generation by mean shear is eectively
obliterated by the sink associated with the stable stratication, the computed
behaviour of the second moments becomes particularly sensitive to second-
moment transport processes. While convective transport is, of course, handled
exactly at this closure level, the same is not true of diusive transport. Usu-
ally a very simple gradient-diusion model is adopted for these processes.
Such a practice, however, is motivated by the desire to adopt a cheap and
stable approximation for a process that is frequently of little importance.
However, transport can be of great importance in strongly inhomogeneous,
stably-stratied ow so something better must be provided.
Section 2 focuses on modelling the extra pressure-containing correlations
arising from buoyancy and provides a comparison of computational results
from implementing this model both with experimental data and, where avail-
able, with the Basic Model. Then Section 3 goes on to consider a case of a
stably-stratied mixing layer where a partial third-moment treatment is re-
quired. The analysis is rst developed before comparisons with experiments
are discussed.
2 Closure Modelling for Stratied Flow
2.1 The Second-Moment Equations
The exact second-moment transport equations, now containing buoyant terms,
407
408 Craft and Launder
may be written:
Du
i
u
j
Dt
=
_
u
i
u
k
U
j
x
k
+u
j
u
k
U
i
x
k
_
. .
P
ij
+
1
u
j
g
i
+
u
i
g
j
_
. .
G
ij
+
1
p
_
u
i
x
j
+
u
j
x
i
_
. .
ij
2v
u
i
x
k
u
j
x
k
. .
ij
x
k
_
u
i
u
j
u
k
+
pu
j
ki
+
pu
i
kj
u
i
u
j
x
k
_
. .
d
ij
.
(2.1)
Usually the density uctuations,
, (2.2)
where is the dimensionless volumetric expansion coecient:
p
.
(For an ideal gas, where denotes temperature, is of course equal to unity).
The buoyant term may thus be rewritten:
G
ij
_
u
j
g
i
+u
i
g
j
_
.
The corresponding transport equation for u
i
then takes the form:
Du
i
Dt
=
_
u
k
U
i
x
k
+u
k
u
i
x
k
_
. .
P
i
2
g
i
. .
G
i
+
p
x
i
. .
i
+d
i
,
(2.3)
while the mean square scalar variance is determined from the following trans-
port equation:
D
2
Dt
= 2u
k
x
k
2
+d
, (2.4)
where, as in (2.1), the terms in equations (2.3) and (2.4) denoted by and d
signify dissipative and diusive processes.
The equation set (2.1)(2.4) is, of course, unclosed, for the processes de-
noted , d and (representing the non-diusive action of uctuating pressure,
[14] Application of TCL modelling to Stratied ows 409
diusion and dissipation) all require approximation. In this section, we focus
especially on the modelling of
ij
and
i
for, in the ows to be considered,
these are the vital processes to be approximated. As a preliminary, however,
the following paragraphs indicate briey the practices adopted for the dissi-
pation and diusion processes.
2.2 Modelling the Dissipation and Diusion Processes
For the dissipation terms, it is convenient, following Lumley (1978), to assume
local isotropy, which implies:
i
= 0;
ij
=
2
3
ij
, (2.5)
where is the viscous dissipation rate of turbulent kinetic energy, k. Any
shortcomings of this assumption (since, for example, such an approximation
for
ij
clearly does not satisfy the TCL) may be considered to be absorbed in
modelling the turbulence or slow parts of
ij
and
i
.
Diusional transport, in the inhomogeneous examples to follow, is approxi-
mated by the rudimentary generalized gradient-diusion hypothesis (GGDH)
of Daly and Harlow (1970).
d
= c
x
m
_
u
m
u
n
k
x
n
_
, (2.6)
where and may denote either a velocity or a scalar uctuation. As in
unstratied ows, the diusive coecient c
u
k
u
l
x
l
_
+
c
1
2
k
(P
kk
+G
kk
) c
2
k
. (2.7)
As signalled in [3], however, for some years the UMIST group has adopted
a dierent strategy in choosing the coecients c
1
and c
2
from that usually
adopted with either the k- EVM or with simpler (older) forms of second-
moment closure. In most of the calculations reported below we have taken:
c
1
= 1.0; c
2
=
1.92
_
1 + 0.7A
1/2
2
A
25
_
, (2.8)
where A
2
is the second invariant of the dimensionless anisotropic Reynolds
stress, a
ij
a
ji
and A
25
max(A, 0.25).
1
1
The quantity A denotes the atness invariant: 19(A
2
A
3
)/8 where the third invariant,
A
3
, is just a
ij
a
jk
a
ki
.
410 Craft and Launder
The change in the coecients from their erstwhile standard values (c
1
=
1.44; c
2
= 1.92) has brought a number of benets in uniform-density ows
which [3] has already mentioned. In stratied ows a major additional benet
is that we have been able to retain the same coecient (c
1
) multiplying both
P
kk
and G
kk
irrespective of whether the ow is orientated vertically or hori-
zontally and whether the stratication is stable or unstable. This is far from
the case when the Basic Model is adopted (see, for example, Hossain and Rodi
1982, p. 142).
The remaining quantity for whose determination a route must be prescribed,
is the scalar dissipation rate,
from:
=
r
2
2k
, (2.9)
where the time-scale ratio
2
r is related to the heat-ux correlation coecient
as
r =
3
2
(1 +A
2
) ; A
2
u
i
u
i
k
2
. (2.10)
2.3 TCL Modelling of Pressure Interaction in Buoyancy Af-
fected Flows
The treatment here parallels that reported in [3]. Note rst that in buoyant
ows the pressure containing correlations may be written:
ij
p
_
u
i
x
j
+
u
j
x
i
_
=
ij1
+
ij2
+
1
4
_
g
k
k
_
u
i
x
j
+
u
j
x
i
_
dVol
[r[
. .
ij3
,
(2.11)
i
p
x
i
=
i1
+
i2
+
1
4
_
g
k
x
i
dVol
[r[
. .
i3
. (2.12)
The nal terms in these equations provide the additional eects due to the
gravitational eld. The primes applied to and merely indicate that the
quantities are evaluated at x
r
/
/
/
/
/
/
/
/
/
/`
`
`
`
`
`
`
Figure 1:
Table 1: TCL models for
ij3
and
i3
.
ij3
=
_
4
10
+
3A
2
80
_
(G
ij
1
/
3
ij
G
kk
) +
1
4
a
ij
G
kk
+
3
20
_
i
u
m
u
j
k
+
j
u
m
u
i
k
_
u
m
1
10
ij
k
u
m
u
k
k
u
m
1
4
k
_
u
k
u
i
k
u
j
+
u
k
u
j
k
u
i
_
+
1
20
ij
k
u
m
u
n
u
m
u
k
k
2
u
n
1
8
_
u
m
u
j
k
u
i
+
u
m
u
i
k
u
j
_
u
m
u
k
k
k
+
1
8
_
u
k
u
i
u
m
u
j
k
2
+
u
k
u
j
u
m
u
i
k
2
_
k
u
m
3
40
_
i
u
m
u
j
k
+
j
u
m
u
i
k
_
u
m
u
n
k
u
n
+
1
4
k
u
m
u
k
u
i
u
j
k
2
u
m
i3
=
1
/
3
2
a
ik
where the ow structure is undergoing very rapid change (as across the buer
layer), it seems reasonable to replace the mean ow variable to be evaluated
at x
and x is mutually
independent.
Consequently:
ij3
=
k
4
_
_
u
i
r
k
r
j
+
2
u
j
r
k
r
i
_
dVol
[r[
k
_
b
i
kj
+b
j
ki
_
. (2.13)
In the above
k
is shorthand for g
k
/. Notice that the tensor b
i
kj
is precisely
the same tensor as appeared earlier in
i2
, (see [3], equation (4.5)). Thus, by
substituting the expressions for this tensor we obtain the model for
ij3
given
in Table 1, Craft (1991), Craft et al. (1996). Likewise:
i3
=
k
b
ki
. (2.14)
412 Craft and Launder
The quantity b
ki
is simply:
1
4
_
x
i
dVol
[r[
=
1
4
_
2
r
k
r
i
dVol
[r[
. (2.15)
To devise an approximation for b
ki
consistent with that obtained for the other
two-point correlations, we suppose:
b
ki
=
1
ki
+
2
2
a
ki
+
3
2
a
mn
a
nm
ik
+
4
2
a
im
a
km
. (2.16)
By requiring that on contracting and performing the above integral, b
kk
=
2
,
one concludes that:
1
=
1
3
; 3
3
+
4
= 0.
Then, on imposing the requirement that if u
2
2
should fall to zero:
G
2
+
23
= 0,
one concludes that
2
= 1,
3
=
4
= 0. Thus nally:
i3
=
1
3
2
a
ik
. (2.17)
It is to be noted that neither the expression for
ij3
(Table 1) nor
i3
contain
any empirical coecients. Thus, to an extent that is rarely found in closure
modelling, comparisons with experimental or DNS data can be called pre-
dictions. In the case of isotropic turbulence, these formulae reduce to the
quasi-isotropic forms (Launder 1975; Lumley 1975):
ij3
=
3
10
_
G
kk
ij
3
G
kk
_
i3
=
1
3
G
i
.
2.4 Some Applications of the Model to Buoyant Flows
In this section comparisons are made between the TCL and the Basic Mod-
els for a number of ows. The TCL model was rst applied by Craft (1991)
to consider homogeneous, horizontal, stably-stratied ow that had been cre-
ated by passing a uniform ow past a screen of dierentially heated horizontal
rods, Webster (1964), Young (1975). The computations shown in Figure 2 have
assumed local equilibrium (i.e.
1
2
(P
kk
+ G
kk
) = ) which is what has tradi-
tionally been adopted for this test case (though it may have been some way
from the truth). Evidently, as the gradient Richardson number, R
i
increases,
the shear-stress correlation coecient decays, the turbulent Prandtl number
increases and the horizontal heat-ux correlation decreases moderately. The
TCL model reproduces these measured responses at least as accurately as
[14] Application of TCL modelling to Stratied ows 413
Figure 2: Second-moment sensitivity to buoyant stratication in nominally
horizontal, equilibrium, homogeneous ow. (a) Shear-stress correlation coe-
cient; (b) normalized turbulent Prandtl number; (c) Correlation coecient of
horizontal heat ux. TCL model; Basic model; symbols: experiments
of Webster (1964), Young (1975).
the Basic Model
3
even though the two empirical coecients in the buoyant
terms of the latter model were optimized by reference to these same data. As
noted above, the TCL scheme, in contrast, has no adjustable coecients in
the buoyant parts of
ij
and
i
.
A further application of the model, which focuses especially on the buoyant
terms, has been reported by Van Haren (1992). The model is believed to be the
same as that presented in Section 2.2 save that c
1
and c
2
took the constant
3
An extensive review of the capabilities of the Basic Model in buoyant ow has been given
by Launder (1989).
414 Craft and Launder
Figure 3: Development of oscillatory waves in stably-stratied grid turbulence,
from Van Haren (1992). In the right-hand graph, realisable R
ij
- refers to
the TCL model; in the left-hand graph R
ij
- refers either to the Basic Model
or a similar linear model.
values 1.44 and 1.76 and a transport equation was solved for
. Van Haren
considered the decay, in the absence of mean strain, of stably-stratied turbu-
lence and, in particular, the oscillatory pattern that is known to be established
during the decay due to reversals in sign of the vertical heat ux. Van Haren
generated 2-point EDQNM results of such a ow and then tested how well var-
ious single-point closures did in reproducing the behaviour. Figure 3 compares
the time history of the normalized vertical heat ux w/
_
w
2
2
_
1/2
versus
normalized time where N is the BruntV aisal a frequency. Quite clearly, the
extended k- model and a simple second-moment closure
4
shown in the left-
hand gure exhibit a signicantly too long period and a too rapid decay of the
heat ux compared with the EDQNM data. In contrast the new formulation
shown in Figure 3b is rather successful at mimicking the EDQNM results.
We turn now to inhomogeneous cases of self-preserving free shear ows:
the plane and axisymmetric vertical, buoyantly-driven plumes. Their spread-
ing behaviour, obtained by Cresswell et al. (1989), is summarized in Table
2; for comparison the spreading rates of the plane and axisymmetric jets in
stagnant surroundings (quoted in [3]) are also given. Comparisons are drawn
both with experiments and with predictions obtained with the Basic Model.
What is evident is that a far better overall agreement is achieved with the
TCL invariant-dependent closure than with the Basic Model. Even where dis-
4
No details were provided of this scheme but it was presumably the Basic Model or the
similar quasi-isotropic model (Launder et al. 1975) frequently used by the group at the
ECL.
[14] Application of TCL modelling to Stratied ows 415
Table 2: Rate of growth of half-width for some self-preserving free shear ows
Flow Basic Model Recommended TCL Model
experimental
values
Plane plume 0.078 0.120 0.118
Axisymmetric plume 0.088 0.112 0.122
Plane jet 0.100 0.110 0.110
Axisymmetric jet 0.105 0.093 0.101
crepancies remain with experimental data, as in the case of the axisymmetric
ows, these are both smaller in magnitude than with the Basic Model and
show a consistent behaviour across both the plume and the jet.
The same workers extended their study to the case of a hot jet discharged
vertically downwards into a cold water environment moving upwards at less
than 2% of the jet velocity. Because of the net vertical force due to buoyancy on
the less dense uid, the penetration length, of course, is crucially dependent on
mixing. Figure 4a is a vector velocity plot in the vicinity of discharge showing
the reversal in the jet direction while Figures 4b and 4c present respectively
the shear stress and resultant mean velocity proles. Evidently, the new model
does signicantly better in capturing the eects of buoyancy and shear in this
quite complex recirculating ow. Cresswell et al. (1989) were, moreover, agree-
ably surprised to report that, because the TCL model respected realizability, it
led to a faster rate of numerical convergence and to a reduction of some 30%
in computing time per run relative to the Basic Model, despite the greater
algebraic complexity of the model itself.
3 Third-Moment Modelling and Application
3.1 Modelling
As signalled in the Introduction, the presence of strongly stable stratication
may lead to a situation where the local state of turbulence is predominantly
determined by diusional transport from neighbouring, less stratied layers.
In these circumstances the generalized gradient-diusion hypothesis, usually
adopted to represent diusional transport of the second moments, is inade-
quate.
The GGDH approximation can be regarded as a severe truncation of the
third-moment equations for cases unaected by buoyancy. Buoyant terms ap-
pear in the third-moment equations just as they do in the second-moment
equations. Thus, it is possible to envisage a modest elaboration of the diu-
sion model merely by including buoyant terms; such a strategy was adopted
416 Craft and Launder
Figure 4: Axisymmetric downward directed bouyancy jet from Cresswell et al.
(1989). (a) Computed velocity vector plot; (b) turbulent shear stress one di-
ameter downstream of the jet discharge; (c) mean velocity prole one diameter
downstream of the jet discharge. TCL model; Basic model; symbols:
experiments.
[14] Application of TCL modelling to Stratied ows 417
in Craft et al. (1996). However, as stable stratication becomes increasingly
dominant, it becomes more and more dicult to make the system of equa-
tions, with their multiple intercouplings, converge. For that reason, it is now
advocated that one should actually solve transport equations for some of the
third-moments. But which ones? Even in a two-dimensional stratied shear
ow, there are twelve third-moments to be determined which represents a se-
vere enlargement of the model. Accordingly, our current strategy has been
to nd from transport equations simply those triple moments which contain
density uctuations since these are the source of the gravitational eects on
the turbulence structure. Thus the elements of u
i
u
j
u
k
are obtained from the
GGDH as in Section 2 while all other third moments are obtained from trans-
port equations. Following Craft et al. (1997) these equations may be expressed
in compact notation as:
Du
k
u
j
Dt
= P
1
kj
+P
2
kj
+G
kj
+
kj
kj
+d
kj
(3.1)
Du
k
2
Dt
= P
1
k
+P
2
k
+G
k
+
k
k
+d
k
(3.2)
D
3
Dt
= P
1
+P
2
+G
+d
, (3.3)
where, in all equations, the P
1
s denote production by second-moment gra-
dients, P
2
s arise from mean velocity or mean scalar gradients while the Gs
are direct buoyant inuences. The detailed form of these processes is given
in Chapter [15], equation (4) but it is remarked that, within third-moment
closure they can all be handled without modelling approximations. As before,
the terms denoted , and d represent non-dispersive pressure interaction
processes, dissipation and diusion and all require closure approximations.
Regarding the non-dispersive pressure interactions, for consistency with the
modelling of the second-moments, one should logically apply the TCL strat-
egy. However, the appropriate analysis does not appear to have been made.
Accordingly, Craft et al. (1997, 2001) adopted precisely the same modelling
strategy as is applied to the pressure-strain correlation when adopting the Ba-
sic Model; that is, in physical terms the pressure uctuations are regarded as
an agent for promoting the return to isotropy of the turbulence eld. Thus:
k
= c
1
k
u
k
c
2
_
P
1
k
+P
2
k
+G
k
_
, (3.4)
where c
1
= 1/0.075, c
2
= 0.5, and on the right-hand side denotes either
u
j
or .
5
5
Concerning the choice of coecients, c
2
is taken as 0.5 to accord with the corresponding
second-moment coecient while c
1
is selected following Dekeyser and Launder (1985).
418 Craft and Launder
Concerning dissipative processes, the proposals advocated by Dekeyser and
Launder (1985), hereafter referred to as DL, are retained for
kj
and
k
:
jk
= c
3
k
u
l
x
l
jk
(3.5)
k
= c
3
k
u
k
u
l
x
l
, (3.6)
where c
3
= 0.1, following DL. For
= 3
_
2r
_
, (3.7)
is adopted with the coecient r chosen as in equation (2.10).
In earlier closure proposals for the diusion of third moments, the usual
route has been to express the fourth rank products in terms of products of the
constituent quantities taken two at a time (Millionshtchikov 1941):
= + + . (3.8)
An argument against using such a form is that the Gaussian distribution of
the uctuations on which the approximation, equation (3.8), rests, will be
least accurate in regions where quadruple products are most inuential, i.e.
where the turbulence is strongly inhomogeneous. Possibly with such thoughts
in mind, Kawamura et al. (1995) proposed instead that, to model u
i
u
j
u
k
u
m
,
departures from equation (3.8) should be accounted for by a gradient transport
model. We apply this idea to all the quadruple products as follows:
d
kj
x
m
_
u
m
u
j
u
k
_
x
m
_
u
m
u
j
u
k
_
u
m
u
j
u
k
+u
m
u
k
u
j
+u
m
u
j
u
k
__
x
m
_
u
m
u
j
u
k
+u
m
u
k
u
j
+u
m
u
j
u
k
_
. (3.9a)
The rst line of equation (3.9a) is then approximated by the GGDH, i.e,
d
jk
= c
3d1
x
m
_
u
m
u
n
k
x
n
u
j
u
k
x
m
_
u
m
u
j
u
k
+u
m
u
k
u
j
+u
m
u
j
u
k
_
.
(3.9b)
The other quadruple moments are modelled analogously as:
d
k
= c
3d2
x
m
_
u
m
u
n
k
x
n
u
k
2
_
x
m
_
u
m
u
k
2
+ 2u
m
u
k
_
(3.10)
d
= c
3d3
x
m
_
u
m
u
n
k
x
n
3
_
x
m
_
3u
m
2
_
, (3.11)
[14] Application of TCL modelling to Stratied ows 419
Fluid 2
Fluid 1
x=0m x=5m x=10m x=40m
g
0.560m
0.323m
Splitter Plate
Free Surface
y
x
z
Bottom Wall
Figure 5: Geometry of the stably-stratied mixing layer of Uittenbogaard
(1988).
where c
3d1
= c
3d2
= c
3d3
= 0.1. Introduction of the Kawamura modication in
modelling the above processes signicantly improved the stability behaviour
and the quality of the computations.
3.2 Application
The test ow chosen for examination was the salinity-stratied mixing layer
measured at the Delft Hydraulics Laboratory by Uittenbogaard (1988), see
Figure 5. The ow is nominally two-dimensional, a light, fresh-water stream
being injected above a more dense saline stream. The two streams are then
allowed to mix together. A previous exploration of this ow (Craft et al. 1996)
had persuaded us that a more rened treatment of the second-moment diu-
sion processes was needed though, on that occasion, as remarked above, just
a fairly rudimentary extension of the GGDH concept was adopted.
Figures 68 from Craft et al. (2001) (see also Kidger 1999) compare proles
of second-moment and partial third-moment closures for a range of measured
quantities. Figure 6 shows the vertical prole of relative density at a distance
40m downstream. We note from the experimental data (crosses) that the con-
centration has by no means equalized even at this most downstream position,
with the mixing region being only about 0.2m in width. By contrast the linear
k- EVM predicts that the mixture was approaching a fully mixed state, Craft
et al. (1996). At second-moment level the TCL model clearly shows too much
mixing while, even more serious, near the bottom and top, the concentration
levels are respectively above the maximum and below the minimum levels in
the entering streams. Such a development is impossible, of course. Figure 6b
shows, however, that by including the 3rd-moment transport equations, satis-
factory agreement with the experimental data is achieved. The reason for this
dramatic improvement is the very dierent proles of the predicted salinity
420 Craft and Launder
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
0.0
0.1
0.2
0.3
0.4
0.5
0.6
y (m)
max min
max
TCL 2
Basic 2
k-
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
0.0
0.1
0.2
0.3
0.4
0.5
0.6
y (m)
max min
max
TCL 3
TCL 2
Figure 6: Relative density at x = 40m in the stably-stratied mixing layer. (a)
Predictions using TCL and Basic second-moment closures and the k- model;
(b) predictions using the TCL second-moment (TCL 2) and partial triple-
moment (TCL 3) closures. Symbols: experiments of Uittenbogaard (1988).
From Kidger (1999).
ux, shown in Figure 7. The second-moment closure exhibits strong reversals
in scalar ux near the bottom of the channel and near the free surface. It is
that feature which causes the under- and over-shoots in the mean concentra-
tion. The 3rd moment closure, however, gives a similar variation of salinity
ux to the measured levels.
We come now to the reason for this striking dierence in behaviour between
the two models. The most important feature is that the buoyant term in the
salinity ux equation (which is proportional to
2
) is conned within a narrow
inner band with the partial 3rd moment closure whereas it is dispersed over
most of the channel depth when this is modelled by the GGDH, see Figure 7.
6
It is the presence of this substantial buoyant sink in the v equation in regions
where gradients of mean salinity are weak which produces the anomalous sign
of v and the consequent over- and under-shoots in mean salinity. It may be
inferred that the improved representation of the diusive uxes, particularly
of v
2
which appears in the
2
transport equation is thus crucial to getting
the correct behaviour of the resultant mean concentration prole. Finally,
it is noted that while the Basic Model also benets from the 3rd-moment
treatment, Kidger (1999), the improvement is by no means as great with the
mean concentration prole still exhibiting over- and under-shoots.
6
The experimental data conrm the narrowness of spread but the magnitude of the uc-
tuations is much lower possibly because a nite probe will smear the variations.
[14] Application of TCL modelling to Stratied ows 421
y (m)
v x 10
3
(kg/m
2
s)
TCL 3
TCL 2
-5.0 -2.5 0.0 2.5 5.0 7.5 10.0
0.0
0.1
0.2
0.3
0.4
0.5
0.6
Figure 7: Vertical density ux at x = 10m in the stably-stratied mixing layer
with TCL closures. partial triple-moment closure; second-moment
closure; symbols: experiments of Uittenbogaard (1988). From Kidger (1999).
0.000 1.500 3.000 4.500 6.000 7.500
0.000
0.080
0.160
0.240
0.320
0.400
0.480
0.560 0.560
0.480
0.400
0.320
0.240
0.160
0.080
0.000
y
(
m
)
Scalar Variance (kg^2/m^6)
Figure 8: Scalar variance at x = 10m in the stably-stratied mixing layer with
TCL closures. partial triple-moment closure; second-moment closure;
symbols: experiments of Uittenbogaard (1988). From Kidger (1999).
422 Craft and Launder
Acknowledgements
The authors express their appreciation to Dr J.W. Kidger in preparing Figures
58. Mrs C. King wordprocessed the text both for this published version and
the INI Conference.
References
Craft, T.J. and Launder, B.E., (1990). The decay of stably-stratied grid turbulence,
4th Colloquium on CFD, UMIST, Manchester, 2425 April 1990, Department of
Mechanical Engineering, UMIST, Manchester, Paper 1.5.
Craft, T.J., (1991). Second-Moment Modelling of Turbulent Scalar Transport, PhD
Thesis, Faculty of Technology, University of Manchester.
Craft, T.J., Ince, N.Z. and Launder, B.E., (1996). Recent developments in second-
moment closure for buoyancy-aected ows, Dyn. Atmos. Oceans 23, 99114.
Craft, T.J., Kidger, J.W. and Launder, B.E., (1997). Importance of 3rd-moment
modelling in horizontal, stably-stratied Flow. Proc. 11th Symp. on Turbulent
Shear Flows, Grenoble, 1997, 20.1320.18.
Craft, T.J., Kidger, J.W. and Launder, B.E., (2001). Selective 3rd-moment closure
for stably-stratied ows, to be published.
Cresswell, R., Haroutunian, V., Ince, N.Z., Launder, B.E. and Szczepura, R.T., (1989).
Measurements and modelling of buoyancy-modied elliptic turbulent ows, Proc.
7th Symp. on Turbulent Shear Flows, Stanford University, 12.4.112.4.6.
Daly, B.J. and Harlow, F.H., (1970). Transport equations in turbulence, Phys. Fluids
13, 2634.
Dekeyser, I. and Launder, B.E., (1985). A comparison of triple-moment temperature-
velocity correlations in the asymmetric heated jet with alternative closure models.
In Turb. Shear Flows 4, L.J.S. Bradbury et al. (eds.), Springer, 102117.
Hossain, M.S. and Rodi, W., (1982). A turbulence model for buoyant ows and its
application to vertical buoyant jets. In Turbulent Buoyant Jets and Plumes, W.
Rodi (ed.) Pergamon.
Itsweire, E.C., Holland, K.N. and Van Atta, C.W., (1986). The evolution of grid-
generated turbulence in a stably-stratied uid, J. Fluid Mech. 162, 299338.
Kawamura, H., Sasaki, J. and Kobayashi, K., (1995). Budget and modelling of triple-
moment velocity correlations in a turbulent channel ow based on DNS, Proc. 10th
Symp. on Turbulent Shear Flows, Pennsylvania State University, 26.1326.18.
Kidger, J.W., (1999). Turbulence Modelling for Stably-Stratied Shear Flows and Free
Surface Jets, PhD Thesis, UMIST, Manchester.
Launder, B.E., (1975). Prediction methods for turbulent ows, Von Karman Institute
for Fluid Dynamics, Rhode St Gen`ese, Belgium.
Launder, B.E., (1989). The prediction of force-eld eects on turbulent shear ow via
second-moment closure. In Adv. in Turbulence 2, H.H. Fernholz and H.E. Fiedler
(eds.), 338358, Springer.
[14] Application of TCL modelling to Stratied ows 423
Launder, B.E., Reece, G. J. and Rodi, W., (1975). Progress in the development of a
Reynolds stress turbulence closure, J. Fluid Mech. 68, 537.
Lumley, J.L, (1975). Prediction methods for turbulent ows, Lecture Series 76, Von
Karman Institute for Fluid Dynamics, Rhode-St-Gen`ese, Belgium.
Lumley, J.L, (1978). Computational modelling of turbulent ows, Adv. Appl. Mech.
18, 123.
Millionshtchikov M.D., (1941). On the role of the third noments in isotropic turbu-
lence, C.R. Acad. Sci. SSSR 32, 619.
Uittenbogaard R.E., (1988). Measurement of turbulence uxes in a steady, stratied,
mixing layer, 3rd Int. Sym. on Rened Flow Modelling and Turb. Meas., Tokyo,
725732.
Van Haren L., (1992). Comparison between one- and two-point closure models for
freely decaying stratied ow. In Pole Europeen Pilote pour la Turbulence (PEPIT),
Conf. on Stratied Turbulence, Paris, France. See also Etude-Theorique et Modelis-
ation de la Turbulence en Presence dOndes Internes, Th`ese de Docteur-Ingenieur,
Ecole Centrale de Lyon, 1992.
Webster C.A.G., (1964). An experimental study of turbulence in a density stratied
shear ow, J. Fluid Mech. 19, 221245.
Young S.T.B., (1975). Turbulence measurements in a stably-stratied shear ow,
Rep. QMCEP6018, Queen Mary College, London.
15
Higher Moment Diusion in Stable
Stratication
B.B. Ilyushin
1 Introduction
In recent years there has been intensive development of turbulence models both
at the 2-equation level of modelling as well as complete second-moment clo-
sure. These models combine (relative) ease of numerical solution with sucient
accuracy for many applications. However, the use of these models to predict
turbulent transport in stratied ows has given results which in some cases are
even qualitatively incorrect, Lamb (1982). Both experimental and theoretical
studies have identied the existence of large-scale eddy structures, LSES, in
stratied ows taking the form of turbulent spots in stable stratication and
coherent structures in unstable stratication. The long lifetime of LSES results
in a uid pollutant particle being entrained by such an eddy and being trans-
ferred a considerable distance by the (LSES) without any appreciable change in
its direction, see Figure 1. Thus feature does not accord with Eulers concepts
of turbulence diusion in which turbulent transport is considered analogous to
Brownian motion; that is, a random-walk process. In this case a particle can
travel a large distance only from multiple motions of random directions. Thus,
to model turbulent transport in stratied ows where turbulent uctuations
are strongly non-isotropic, one needs to take into account the contribution of
LSES corresponding to the long-wave-length part of the spectrum.
(a) (b)
Figure 1: Transfer of a uid (pollutant) particle in the eld of eddies by (a)
the inertial spectrum range (b) under the action of LSES.
424
[15] Higher moment diusion in stable stratication 425
lg E() Ri < 0
Ri = 0
Ri > 0
~
5/3
lg
Figure 2: Qualitative dependence of the vertical velocity uctuations spectrum
on a character of stratication of the atmospheric boundary layer.
As the ow stability increases the turbulent uctuations decrease in intensity
due to destruction of the LSES. The maximum of the energy spectrum moves
to smaller scales (higher wave number), see Figure 2, and the upper limit of
the interval range (
n+1
, (1)
where the function F depends on space and time coordinates, cumulants of
4
We dene the order of a turbulence model as the order of the highest moment determined
from a dierential transport equation.
428 Ilyushin
Figure 4: Closure strategy for turbulence models.
order no higher than (n+1) (higher-order cumulants are assumed to be negli-
gible) and a relaxation time scale,
n+1
, of the (n+1)th-order cumulant (which
is assumed proportional to the overall turbulence time scale k/, where
in this context may be thought of as the spectral ux of kinetic energy):
n+1
= /
C
n+1
where
C
n+1
is a coecient of proportionality). On expand-
ing the cumulant C
n+1
in a Taylor series about time t + and coordinate
r +U
n+1
, where t
n+1
, and retaining only the rst three terms, equation
(1) can be expressed as:
C
n+1
(t +
n+1
, r +U
n+1
) = C
n+1
(t, r) +
C
n+1
t
n+1
+
C
n+1
x
j
r
U
j
n+1
=
C
n+1
F
_
r, t, C
1
, . . . , C
n+1
_
. (2)
It should be noted that the right hand side of equation (2) involves the time
moment t and coordinate r, whereas the cumulant is calculated at the time
moment t +
n+1
at a point r +U
n+1
. To dene the cumulant C
n+1
at time
[15] Higher moment diusion in stable stratication 429
t and at a point with coordinate r we assume that changes in its value over
the interval
n+1
in the region with size r+U
n+1
are linear and can be taken
into account by a change in the value of the relaxation coecient:
C
n+1
(t) =
C
n+1
F(r, t, C
1
, . . . , C
n+1
). (3)
Evidently, the assumptions made to obtain the algebraic model (3) limit its
use. Allowing for the fact that the typical time scale of relaxation of turbulent
uctuations is larger if the corresponding wave vector is smaller (Monin and
Yaglom 1967), we may conclude that the algebraic parameterization (3) of
the cumulant will result in a larger error if the contribution of the long-wave
uctuations (large-scale structures) to turbulent transfer mechanism is greater.
Using the above approach for creating the nth-order models for n = 2 (see
Figure 6) leads to a standard second-order closure. In fact, using dierential
transport equations to calculate some of the second-order moments and alge-
braic relations for others is often sucient. Of course, to close the second-order
model it is not sucient simply to express the triple correlation by means of
known functions. Other correlations (pressure-containing and dissipative ones)
also need to be approximated. There has recently been signicant progress in
this eld and a large number of papers have been devoted to solving these prob-
lems. Our attention will be mainly directed at algebraic models for the triple
correlations used in second-order models and at third-order closure models.
3 Algebraic models for triple correlations
The transport equations for the triple correlations in stratied ows have the
following form:
Du
i
u
j
u
k
)
Dt
=
C
ijkl
x
l
+
ijk
_
u
i
u
j
u
l
)
U
k
x
l
u
i
u
l
)
u
j
u
k
)
x
l
+g u
i
u
j
)
k3
_
c
3
u
i
u
j
u
k
)
,
Du
i
u
j
)
Dt
=
C
ij
l
x
l
u
i
u
j
u
l
)
x
l
u
i
u
l
)
U
j
x
l
u
l
)
u
i
u
j
)
x
l
+
ij
_
u
i
u
l
)
u
j
)
x
l
+g
u
i
2
_
j3
_
c
3
u
i
u
j
)
,
D
u
i
2
_
Dt
=
C
i
l
x
l
u
i
u
l
)
x
l
u
l
2
_
U
i
x
l
u
l
)
u
i
)
x
l
+g
3
_
i3
c
3
u
i
2
_
,
D
3
_
Dt
=
C
l
x
l
u
l
2
_
x
l
u
l
)
2
_
x
l
c
3
u
i
2
_
,
(4)
where U
i
and u
i
are the mean and uctuating components of the instantaneous
velocity, = 1/ is the coecient of volumetric expansion, and are the
430 Ilyushin
mean and uctuating components of the potential temperature, g is the gravity
force acceleration, = k/ is the time scale of turbulence, k =
1
2
u
i
u
i
) is the
turbulent kinetic energy, its dissipation rate, c
ijk
_
u
i
u
j
u
l
)
U
k
x
l
u
i
u
l
)
u
j
u
k
)
x
l
+g u
i
u
j
)
k3
_
,
u
i
u
j
)=
c
3
_
u
i
u
j
u
l
)
x
l
u
i
)
u
i
u
j
)
x
l
+
ij
_
u
i
u
l
)
U
j
x
l
u
i
u
l
)
u
j
)
x
l
+g
u
i
2
_
j3
__
,
u
i
2
_
=
c
3
_
u
i
u
l
)
x
l
u
l
2
_
U
i
x
l
u
i
u
l
)
2
_
x
l
u
l
)
u
i
)
x
l
+g
3
_
i3
_
,
3
_
=
c
3
_
u
l
2
_
x
l
u
l
)
2
_
x
l
_
.
(5)
The complete set (5) of algebraic equations was applied by Canuto et al.
(1994) to describe the turbulence structure in the case of unstable stratica-
tion. It is evident that use of (5) in a second-order model (to model diusion
terms) requires the time-consuming inversion of twenty equations. Canuto et
al. (1994) inverted the matrix by means of symbolic calculations. However, the
triple-correlation algebraic models can be simplied signicantly for many of
stratied ows. In particular, Hanjalic and Launder (1972) showed that the
contribution of terms with mean velocity gradient to equations (5) was small
and could be neglected. One can also, as adopted for example in Ilyushin and
Kurbatskii (1997) single out the vertical turbulent transport (the direction of
the buoyancy force). In this way, more complete models are used for correla-
tions containing the vertical velocity uctuations than those with a horizontal
velocity uctuations. The algebraic models for the triple correlations
w
3
_
,
w
2
_
and
w
2
_
in the case of a horizontally homogeneous ow take the
[15] Higher moment diusion in stable stratication 431
form:
w
3
_
=
3
3
c
3
_
_
w
2
_
+
g
c
3
w)
_
w
2
_
z
+
2g
c
3
_
w
2
_
+
g
c
3
w)
_
w)
z
+
2(g)
2
c
3
c
3
_
w
2
_
+
3g
c
3
w)
_
2
_
z
_
,
(6a)
w
2
_
=
3
c
3
_
_
w)
3
c
3
w
2
_
z
_
w
2
_
z
+2
_
w
2
_
+
2g
c
3
w)
_
w)
z
+
2g
c
3
_
w
2
_
+
3g
c
3
w)
_
2
_
z
_
,
(6b)
w
2
_
=
3
c
3
_
2
c
3
z
_
w)
3
c
3
w
2
_
_
w
2
_
z
+2
_
w)
3
c
3
w
2
_
_
w)
z
+
_
w
2
_
+
3g
c
3
w)
_
2
_
z
_
,
(6c)
where
3
=
1 +
3
c
3
c
3
2
N
2
+
9
c
3
c
3
c
3
c
3
4
N
4
, =
1 +
3
c
3
c
3
2
N
2
,
3
=
1 +
3c
3
+4c
3
c
3
c
3
c
3
2
N
2
+
9
c
3
c
3
c
3
c
3
4
N
4
, =
1 +
3
c
3
c
3
2
N
2
,
3
=
1 +
3c
3
+4c
3
c
3
c
3
c
3
2
N
2
+
9
c
3
c
3
c
3
c
3
4
N
4
, =
1 +
3c
3
+4c
3
c
3
c
3
c
3
2
N
2
,
where N
2
= g
z
is the BruntVaisala frequency (the correlation
3
_
is
not required for closing the second-order model). One can see that equation
(6) incorporates the mechanism of triple-correlation damping in the case of
a stably stratied ow (
2
N
2
> 0) in contrast to the usually used tensor-
432 Ilyushin
invariant models of the form (see, for example, Hanjalic and Launder 1972):
w
3
_
0
=
3
c
3
w
2
_
w
2
_
z
,
w
2
_
0
=
c
3
_
w)
w
2
_
z
+ 2
w
2
_
w)
z
_
,
w
2
_
0
=
c
3
_
2 w)
w)
z
+
w
2
_
2
_
z
_
,
3
_
0
=
3
c
3
w)
2
_
z
.
(7)
The triple-correlation algebraic models taking into account buoyancy eects
to a rst approximation (only terms linear in are retained) may be derived
from (6) taking into account equation (7):
w
3
_
=
w
3
_
0
+ 3
g
c
3
w
2
_
0
1 +
3
c
3
c
3
2
N
2
,
w
2
_
=
w
2
_
0
+
c
3
_
N
2
g
w
3
_
0
+ 2g
w
2
_
0
_
1 +
3c
3
+4c
3
c
3
c
3
c
3
2
N
2
,
w
2
_
=
w
2
_
0
+
c
3
_
2
N
2
g
w
2
_
0
+g
3
_
0
_
1 +
3c
3
+4c
3
c
3
c
3
c
3
2
N
2
,
(8)
or, for the case of a three-dimensional ow:
u
i
u
j
u
k
) =
u
i
u
j
u
k
)
0
+
g
c
3
u
i
u
j
)
0
1 +
c
3
c
3
2
N
2
,
u
i
u
j
) =
u
i
u
j
)
0
+
c
3
N
2
g
u
i
u
j
u
3
)
0
+g
_
u
i
2
_
0
j3
+
u
j
2
_
0
i3
__
1 +
_
+1
c
3
c
3
+
2
c
3
c
3
_
2
N
2
,
u
i
2
_
=
u
i
2
_
0
+
c
3
_
N
2
g
u
i
u
3
)
0
+g
3
_
0
i3
_
1 +
_
c
3
c
3
+
3
i3
c
3
c
3
_
2
N
2
,
(9)
=
i3
+
j3
+
k3
; =
i3
+
j3
; = 2(1 +
i3
).
The lack of experimental data for the distribution of the triple correlations
in stably stratied ows prevents any direct check on the adequacy of models
(6), (9). However, it is known that the mechanism of damping the vertical
velocity uctuations by buoyancy is similar to that of suppressing radial ve-
locity uctuations in a swirling ow in a straight circular-sectioned pipe. The
[15] Higher moment diusion in stable stratication 433
third-order moments distributions measured in Zaetz et al. (1985) for this ow
enable us to evaluate the adequacy of accounting for the mechanism of swirl
damping in the above approach. The models for the triple correlations
v
3
_
and
u
2
v
_
for swirling ow take the form
5
:
v
3
_
=
v
3
_
0
+ 6
c
3
W
r
v
2
w
_
0
1 + 6
_
c
3
_
2
W
r
_
W
r
+
W
r
_
,
u
2
v
_
=
u
2
v
_
0
+ 2
c
3
W
r
u
2
w
_
0
1 + 2
_
c
3
_
2
W
r
_
W
r
+
W
r
_
,
(10)
where W is the circumferential velocity and the coecient c
3
is taken as 4.75
(1/c
3
0.21) to give best agreement with measurements for the skewness
factor S
v
in non-rotating pipe ow (see Figure 7a) with results of calculation
by the model of Hanjalic and Launder (1972):
v
3
_
0
=
c
3
_
3
v
2
_
v
2
_
r
6
vw)
2
r
_
,
u
2
v
_
0
=
c
3
_
v
2
_
u
2
_
r
+ 2 uv)
uv)
r
2
uw)
2
r
_
,
(11a)
v
2
w
_
0
=
c
3
_
2
v
2
_
_
vw)
r
+
vw)
r
_
+vw)
_
v
2
_
r
4
w
2
_
r
__
,
u
2
w
_
0
=
c
3
_
2 uv)
_
uw)
r
+
uw)
r
_
+vw)
u
2
_
r
_
.
(11b)
One may note the similarity of accounting for swirl inuence on the triple
correlations in the model (10) with that of buoyancy inuence in (8).
Figure 5 shows proles of the radial skewness factor S
v
and triple correlation
u
2
v
_
computed by including (and omitting) eects of damping the third-order
moments by swirl for a ow in a cylindrical pipe by means of the algebraic
models (10)
6
. The results of calculations using the models of Daly and Harlow
(1970) and Hanjalic and Launder (1972), (11), as well as the measured data,
are plotted in Figure 5. It can be seen that, taking into account the mechanism
of swirl damping in (10), the calculated third-order moment proles come
closer to the measured ones.
5
In deriving equation (10), the contribution of the algebraic convective term including the
swirl velocity (that is the term proportional to W/r) in the dierential transport equation
(4) for triple correlations has been also taken into account.
6
The work was carried out with Dr. S.N. Yakovenko.
434 Ilyushin
0.0 0.5 1.0
0.4
0.2
0.0
S
v
r/R
1
2
3
2,3
a)
0.0 0.5 1.0
0.2
0.0
<u v>/u
2
*
3
r/R
1
2,3
1
2
3
b)
Figure 5: Radial (a) skewness factor S
v
=
w
3
_
/
w
2
_
3/2
and (b) correlation
u
2
v
_
distributions in the rotated pipe ow. Lines correspond to calculations
without (dashed lines) and with (solid lines) rotation: 1 by the model (Daly
and Harlow 1970), 2 by the model (14) (Hanjalic and Launder 1972), 3
by the model (13); symbols ( without rotation and with it) are the
experimental data (Zaetz et al. 1985).
The ow that forms in the turbulent wake behind a self-propelled body
in a stably stratied ow is a well-known example of a turbulent stratied
shear ow. For comparatively weak stratication, the wake develops at rst in
practically the same way as in a homogeneous uid and expands symmetrically.
However, buoyancy forces inhibit vertical turbulent diusion. Consequently, at
large distances from the body, the wake becomes attened out and then nally
entirely ceases to grow in the vertical direction. Turbulent mixing causes a
more uniform density distribution within the wake than outside it. Gravity
forces tend to restore the earlier undisturbed state of stable stratication. As
a result, convective streams leading to the formation of internal waves in the
surrounding uid appear in the plane normal to the wake axis. The vertical
extent of the wake is an important characteristic illustrating the process of
turbulent diusion in a stably stratied uid.
The results of numerical simulations of the turbulent wake dynamics in a
linear stratied medium show that taking into account the buoyancy eect in
the triple-correlation algebraic models allows us to describe the behaviour of
the wake height downstream of the body (this height is dened by prole of the
vertical component
w
2
_
of turbulence kinetic energy) in complete agreement
[15] Higher moment diusion in stable stratication 435
H
1
0.1
1
2
F
D
565
314
103
65
Lin
Pao
0.001 0.01 0.1 1 5
t / T
65
Figure 6: Evolution of wake height: lines are calculation results: 1 model with-
out buoyancy terms; 2 model taking into account buoyancy eects; symbols
are experimental data (Lin and Pao 1979).
with the measurements data (Lin and Pao 1979)
7
in contrast to models which
neglect this eect (see Figure 6)
8
.
4 Third-order closure models
There is a considerable number of references on the use of third-order tur-
bulence models in the literature. However, as a rule, these models apply the
Millionshikov quasi-normality hypothesis to model the diusion processes in
the equations for the triple correlations, that is, for n = 3 all cumulants of
fourth and higher order are assumed to be zero (whereas, in the proposed
approach, for n = 3 only cumulants of fth and higher order are regarded as
zero). As a consequence, in the former case the triple-correlation equations
are of rst order without the necessary mechanism of damping the triple cor-
relations. Instead, in Andre et al. (1978) the physically incorrect procedure
of clipping the triple correlations values in accordance with the generalized
Schwartz inequalities has been used. The approach proposed below allows us
to overcome this diculty. It has been corroborated by results of the paper
by Hazen (1963) showing that to predict the early stages of turbulence the
use of the quasi-normal hypothesis (i.e. the assumption of zero fourth-order
cumulants) is limited to small uctuation magnitudes in which third moments
are small. On the other hand, using the hypothesis assuming zero fth-order
cumulants (for non-zero fourth-order cumulants) allows the width of applica-
bility of the model to be markedly widened. The equations for second- and
7
The lack of measurement data for the proper triple correlation distributions prevents us
making a direct comparison of computed third-order moments.
8
This work has been undertaken with Professor G.G. Chernykh and Dr. O.F. Voropaeva.
436 Ilyushin
fourth-order moments of the velocity eld are as follows:
u
i
u
j
)
t
+U
k
u
i
u
j
)
x
k
=
u
i
u
j
u
k
)
x
k
ijk
_
u
i
u
k
)
U
j
x
k
+g
k
u
i
)
jk
+
1
_
u
i
p
x
j
_
_
u
i
2
u
j
x
k
x
k
__
,
(12)
u
i
u
j
u
k
u
l
)
t
+U
m
u
i
u
j
u
k
u
l
)
x
m
=
u
i
u
j
u
k
u
l
u
m
)
x
m
ijkl
_
u
i
u
j
u
k
u
m
)
U
l
x
m
+u
i
u
j
u
k
)
u
l
u
m
)
x
m
+g
m
u
i
u
j
u
k
)
ml
+
1
_
u
i
u
j
u
k
p
x
l
_
_
u
i
u
j
u
k
2
u
l
x
m
x
m
__
,
(13)
where is the density and is the viscosity of uid. To approximate fourth-
order moments in equation (13), the fth-order cumulants are assumed to be
zero:
C
ijklm
= u
i
u
j
u
k
u
l
u
m
) u
i
u
j
) u
k
u
l
u
m
) u
i
u
k
) u
j
u
l
u
m
)
u
i
u
l
) u
k
u
j
u
m
) u
i
u
m
) u
k
u
l
u
j
) u
j
u
k
) u
i
u
l
u
m
)
u
j
u
l
) u
i
u
k
u
m
) u
j
u
m
) u
i
u
k
u
l
) u
k
u
l
) u
i
u
j
u
m
)
u
k
u
m
) u
i
u
j
u
l
) u
l
u
m
) u
i
u
j
u
k
)
= 0.
(14)
Taking into account (12), (13) and (14), one can derive the equation for the
fourth-order cumulant C
ijkl
:
C
ijkl
t
=
ijkl
_
C
ijkm
U
l
x
m
C
ijk
g
m
ml
__
u
i
u
j
u
k
p
x
l
_
u
i
u
j
)
_
u
k
p
x
l
_
u
i
u
k
)
_
u
j
p
x
l
_
u
i
u
l
)
_
u
j
p
x
k
__
+
__
u
i
u
j
u
k
2
u
l
x
m
x
m
_
u
i
u
j
)
_
u
k
2
u
l
x
m
x
m
_
(15)
u
i
u
k
)
_
u
j
2
u
l
x
m
x
m
_
u
i
u
l
)
_
u
j
2
u
k
x
m
x
m
__
u
i
u
k
u
m
)
u
j
u
l
)
x
m
u
i
u
m
)
u
j
u
k
u
l
)
x
m
_
u
i
u
k
u
m
)
u
j
u
l
)
x
m
u
j
u
l
u
m
)
u
i
u
k
)
x
m
,
[15] Higher moment diusion in stable stratication 437
where
C
ijk
= u
i
u
j
u
k
) u
i
u
j
) u
k
) u
i
u
k
) u
j
) u
j
u
k
) u
i
)
is a mixed cumulant of velocity and potential temperature uctuations (see
below). Equation (15) contains the unknown cumulant of velocity uctuations
and uctuating pressure derivative as well as the cumulant including the vis-
cosity (third group of terms on the right-hand side of the equation). To model
these two cumulants we assume that their sum is represented by the relax-
ation term C
ijkl
/
4
(where
4
= /
C
4
with the coecient of proportionality
C
4
between the typical time scale of turbulence and that of the cumulants
relaxation) and equation (15) transforms to:
C
ijkl
t
=
ijkl
_
C
ijkm
U
l
x
m
C
ijk
g
m
ml
u
i
u
j
u
m
)
u
k
u
l
)
x
m
u
i
u
m
)
u
j
u
k
u
l
)
x
m
_
u
i
u
k
u
m
)
u
j
u
l
)
x
m
u
j
u
l
u
m
)
u
i
u
k
)
x
m
C
4
C
ijkl
.
(16)
Using (3), the cumulant algebraic model is derived from (16):
C
ijkl
=
C
4
_
_
_
ijkl
_
C
ijkm
U
l
x
m
+C
ijk
g
m
ml
+u
i
u
j
u
m
)
u
k
u
l
)
x
m
+u
i
u
m
)
u
j
u
k
u
l
)
x
m
_
+ u
i
u
k
u
m
)
u
j
u
l
)
x
m
+u
j
u
l
u
m
)
u
i
u
k
)
x
m
_
_
_
.
(17)
The use of (17) for the cumulant C
ijkl
in determining the evolution of third-
order moments from the dierential transport equations implies a quicker re-
laxation of C
ijkl
than that of the third-order moments. This condition restricts
the value of the coecient C
4
; C
4
> C
3
, where C
3
is the coecient of pro-
portionality between the typical time scale of turbulence and the typical time
scale of the triple-correlation relaxation (the coecient of the model for the
pressure-containing correlation in the third moment equations). The Schwarz
inequality for the triple correlation
u
3
i
_
u
2
i
_
u
4
i
_
u
3
i
_
2
0 (18)
is considered to x C
4
. Equation (18) may be written more strongly as:
C
iiii
u
3
i
_
2
u
2
i
_
. (19)
438 Ilyushin
Next we consider the decay of homogeneous turbulence according to the laws
w
2
_
=
w
2
_
0
exp
_
w
3
_
=
w
3
_
0
exp
_
C
3
t
_
C
iiii
= C
iiii0
exp
_
C
4
t
_
.
_
_
(20)
The condition C
4
2C
3
1 for the upper limit of the coecient C
4
follows
from (19) taking into account (20). In using equation (17) for the cumulant
C
ijkl
in the triple-correlation transport equation, the value of this coecient
(C
3
< C
4
< 2C
3
1) is taken to be equal to its upper limit:
C
4
= 2C
3
1. (21)
In the case of a stratied ow the algebraic model (17) for the cumulant
of the velocity uctuations contains the mixed cumulants of velocity and po-
tential temperature uctuations. Using analogous arguments, the following
equivalent algebraic models can be obtained for these cumulants:
C
ijk
=
C
4
_
_
_
ijkl
_
C
ijm
U
k
x
m
+C
ijkm
x
m
+C
ij
g
m
km
+u
i
u
j
u
m
)
u
k
)
x
m
+ u
i
u
m
)
u
j
u
k
)
x
m
+u
i
u
m
)
u
j
u
k
)
x
m
_
+u
m
)
u
i
u
j
u
k
)
x
m
ijkl
_
_
_
C
ij
=
C
4
_
_
_
ij
_
C
im
U
j
x
m
+ 2C
ijm
x
m
+C
i
g
m
jm
+u
m
)
u
i
u
j
)
x
m
+ 2 u
i
u
m
)
u
j
)
x
m
+u
i
u
m
)
u
j
2
_
x
m
_
+
2
u
m
_
u
i
u
j
)
x
m
+u
i
u
j
u
m
)
2
_
x
m
ijkl
_
_
_
C
i
=
C
4
_
C
m
U
i
x
m
+ 3C
im
x
m
+C
g
m
im
+3
2
u
m
_
u
i
)
x
m
+ 3 u
i
u
m
)
2
_
x
m
+ 3 u
m
)
u
i
2
_
x
m
+u
i
u
m
)
3
_
x
m
_
C
=
C
4
_
4C
m
x
m
+ 6
u
m
2
_
2
_
x
m
+ 4 u
m
)
3
_
x
m
_
. (22)
[15] Higher moment diusion in stable stratication 439
This algebraic model (22) of the cumulant contains the relaxation coecients
which are not considered to be equal to each other because the mechanisms of
decay of dierent cumulants are determined by dierent physical processes.
The Schwarz inequalities for the triple correlations including temperature
uctuations are
u
2
i
_
u
2
i
2
_
u
2
i
_
2
0;
2
_
4
_
3
_
2
0. (23)
Equation (25) thus provides the following inequalities for the cumulants
C
ij
u
2
i
_
2
u
2
i
_
, C
3
_
2
2
)
. (24)
These are used to dene upper limits for the coecients C
4
and C
4
of
equation (22). For this, the decay of thermal inhomogeneity is considered for
homogeneous turbulence according to the relations:
u
i
) = u
i
)
0
exp
_
C
1
t
2
_
=
2
_
0
exp
_
r
t
u
2
i
_
=
u
2
i
_
0
exp
_
C
3
t
3
_
=
3
_
0
exp
_
C
3
t
,
C
ij
= C
ij
0
exp
_
C
4
t
, C
= C
0
exp
_
C
4
t
,
(25)
where C
1
and r = /
2
_
, C
3
and C
3
are the corresponding
coecients for the third-order moments
u
2
i
_
and
3
_
. Applying analogous
arguments to that used to x the value of C
4
, and noting equation (25), we
derive from (24) the following relations between the relaxation coecients of
the second-, third- and fourth-order cumulants:
C
4
= 2C
3
C
1
, C
4
= 2C
3
r. (26)
Finally it should be noted in this section that equations (21) and (26) between
the relaxation coecients of the second-, third- and fourth-order cumulants
have been obtained from rather sweeping assumptions. Further renements in
their values may be made from numerical optimization.
5 Assessment of cumulant model
As mentioned above, the available data base for the statistical structure of tur-
bulence in stably stratied ows is limited to the distribution of second-order
moments. It does not permit us to examine the capability of the cumulant
model in these ows by a direct comparison of third-order moments calcu-
lated from the equations using the cumulant models. Applying the cumulant
9
r is the reciprocal of the time-scale ratio R introduced in Chapter 2.
440 Ilyushin
0.0 0.2
0
1
<wk
> /w
z /z
*
3
i
Figure 7: Proles of vertical ux of turbulent kinetic energy: the solid line
is calculation by the model presented in Ilyushin and Kurbatskii (1996), the
dashed line is calculation by the model in Andre J.C. et al. (1978); are
laboratory experimental data (Willis and Deardor 1974), , , , , _, ,
, are observed data (Lenschow et al. 1980).
model in the third-order closure model for the convective atmospheric bound-
ary layer (Ilyushin and Kurbatskii 1996) enables the positive vertical ux wk
)
of turbulence energy to be predicted in complete agreement over the whole of
the layer height including the near-ground layer (see Figure 7). The lack of
detailed experimental data for the fourth-order moments in the ABL does not
allow us to make a direct test of the cumulant models in this ow.
However, the eect of stable stratication on turbulence is seen by the sup-
pression of vertical velocity uctuations. Noting the action of this suppres-
sion, turbulence in such ows is characterized by strong anisotropy. Analo-
gous mechanisms cause the large turbulence anisotropy in swirling pipe ow
and in turbulence near a wall. The distribution of higher-order moments
has been measured in such ows and this allows us to compare the fourth-
order cumulants, measured and calculated, using the proposed model. Rau-
pach (1981) measured distributions of the vertical velocity intensity, skewness
S
w
=
w
3
_
/
w
2
_
3/2
and atness factor K
w
=
w
4
_
/
w
2
_
2
in the bound-
ary layer on a at rough plate. However, data for the vertical distribution of
the typical time scale of turbulence (or equivalently the spectral ux of the
[15] Higher moment diusion in stable stratication 441
(a) (b) (c)
0 1
0
1
1 1 2 4 6 8 10
Figure 8: The distribution of turbulence intensity (
w
=
w
2
_
1/2
(a), skewness
S
w
=
w
3
_
/
3
w
(b), and kurtosis K
w
=
w
4
_
/
4
w
(c) coecients of vertical
variance: lines correspond to interpolated functions for
w
and S
w
; the line on
the plot (c) is the calculated prole of kurtosis K
w
; symbols are the experi-
mental data (Raupach 1981).
turbulence kinetic energy) are lacking. It makes it impossible to dene the
K
w
prole by a direct substitution of measured values into (17). Nevertheless,
taking into account the fact that for this ow the algebraic gradient model
(7) adequately describes the behaviour of the third moment
w
3
_
, the cumu-
lant value can be calculated from the algebraic expression obtained from (17).
Figure 8c shows the prole of the atness factor K
w
calculated by substitut-
ing into (17) analytic interpolating functions for the distribution of
w
2
_
and
w
3
_
(see the dashed lines in Figure 8 a,b) which were measured in Raupach
(1981) (d is the displacement of the ow thickness by the roughness and
is the boundary layer thickness). It is seen that for the boundary layer on a
at plate, the algebraic model (17) for the cumulant C
3333
correctly describes
the behaviour both within the boundary layer and near the plate where the
structure of turbulent uctuations is characterized by substantial anisotropy
caused by the rigid wall damping of the vertical velocity uctuations.
The results of testing the cumulant model for a rotating ow in a straight
circular sectioned pipe are shown in Figure 9. Measurements (Zaetz et al. 1985)
have been conducted in a rotating section of the pipe where the inuence of
transitional eects on the turbulence is substantial. The experimental data
10
in Figure 9 show that the eect of pipe rotation does not inuence the ow
structure for r/R > 0.6. The swirl ow damps radial uctuations of velocity
and decreases the atness factor. The calculations shown in Figure 9 qualita-
10
Near the axis the experimental points have a large error and the dierence in the kur-
tosis values at r/R = 0.9 for both cases (with and without rotation) is apparently due to
measurement errors.
442 Ilyushin
3 4
0.0
0.2
0.4
0.6
0.8
v
r/R
K
Figure 9: Proles of the atness factor K
v
=
v
4
_
/
v
2
_
2
in circular pipe
ow. Lines (solid in swirling ow, dashed without swirl) are calculations;
symbols ( in swirling ow; without swirl) are the experimental data
(Zaetz et al. 1985).
tively conrm the eects noted. However, agreement of the calculations with
the measured data is worse than in the boundary layer on a plate (Figure
8). To ascertain the reasons for this disagreement between the measured at-
ness factor and those calculated, more extensive studies of swirling ow (both
experimental and numerical) are required.
6 Conclusions
1. An increase in stable stratication leads to a decrease in the inertial
range of the turbulent uctuation spectrum. This decrease is caused
mainly by the destruction of large-scale eddy structures. As a conse-
quence, the inuence of the spatially non-local turbulent transfer (i.e.,
the transfer to a large distance by the action of large-scale eddy struc-
tures) on the dynamics of a ow becomes weaker whereas the inuence of
the non-local memory eects conditioned by intermittency of turbulence
increases. Gradient diusion models which are applicable to turbulent
transport on scales corresponding to the inertial range are inadequate
for predicting such ows.
[15] Higher moment diusion in stable stratication 443
2. Large values of the atness coecients of the velocity uctuation PDFs
(fourth-order cumulants) for values of the skewness (third-order mo-
ments) close to zero in stably stratied ows preclude the application of
third-order closure models using the quasi-normality hypothesis of Mil-
lionshikov (the assumption of fourth-order cumulants) if one is to obtain
an adequate description of the turbulence structure.
3. In many cases second-order closure models give all the required informa-
tion about the statistical structure of the investigated ow and it is then
not useful to adopt higher-order closures to describe the turbulence.
4. Algebraic expressions for the triple correlations, taking into account
buoyancy eects, allow one to reproduce the suppression of vertical ve-
locity uctuations by the action of stable stratication. They can be used
to model the diusion terms in the transport equations of second-order
closure models.
5. To calculate the distribution of higher-order moments in stably stratied
ows, one should apply higher-order closure models. The closure strategy
adopted enables third-order turbulent transport models to be formulated
accounting for non-zero values of the fourth-order cumulants.
Appendix
For a complete statistical description of the hydrodynamic characteristic elds
in a turbulent ow, it is necessary to set the whole of the multi-dimensional
joint distributions of probability for the values of these characteristics at ev-
ery possible ensemble of space-time points. However the denition of such
multi-dimensional distributions is a highly complex problem. Moreover these
distributions are often inconvenient for applications because they are bulky.
Thus, generally one is restricted to employing only some simpler statistical
parameters describing particular properties of a stream. The most important
parameters of probability distributions are the moments and some of their
combinations (in this section the central moments of one-point distributions
P(u
i
) will be considered for simplicity):
_
u
1
u
2
u
3
_
=
_
R
u
1
u
2
u
3
P(u) du moments,
i
=
u
2
i
_
1/2
dispersions,
S
i
=
u
3
i
_
3
i
skewness factors,
i
=
u
4
i
_
4
i
3; excess coecients,
444 Ilyushin
where u
i
is the vector of turbulent uctuations of velocity (i, j, k = 1, 2, 3). The
two latter coecients, dependent on values of higher-order moments, charac-
terize the departure of the probability density function (PDF) from the normal
(Gaussian) one (see Figure 10). The distribution with non-zero asymmetry
(S > 0) corresponds to a ow with intense but rare uxes directed positively
alternating with slow but more probable uxes directed negatively (and vice
versa for S < 0). A positive excess value indicates that the distribution is
less at than normal and a quantity is concentrated in regions of very large
and very small values. Flows with such a distribution are characterized by a
mixture of long periods of relative steadiness (when uctuations are small)
with periods of increased activity (with turbulent uctuations). Such a sta-
tistical structure is termed spotted
11
. The coecient of asymmetry S
of
uctuating velocity derivatives along the coordinate has an important phys-
ical sense. Since mainly small-scale uctuations contribute to this quantity,
its value reects the statistical structure of small-scale turbulence. For scales
in the inertial range of the turbulent spectrum, the S
n,m,k=0
C
n+m+k
(iq
1
)
n
)
n!
(iq
2
)
m
m!
(iq
3
)
k
k!
+O(q
r
),
(q) =
_
R
P(u) exp(iqu) du.
Coecients of this expansion are termed cumulants (semi-invariants). They
are more convenient for use as statistical characteristics of distributions than
the moments because of the fact that cumulants of order higher than second
describe how non-Gaussian (non-equilibrium) are distributions of uctuations
and also because of their invariant properties. The rst four cumulants are:
C
1
i
= u
i
) ,
C
2
ij
= u
i
u
j
) ,
C
3
ijk
= u
i
u
j
u
k
) ,
C
4
ijkl
= u
i
u
j
u
k
u
l
) u
i
u
j
) u
k
u
l
) u
i
u
k
) u
j
u
l
) u
i
u
l
) u
j
u
k
) ,
11
Such a statistical structure of uctuations is typical of small-scale turbulence (Monin
and Yaglom 1967) and also for stably stratied turbulent ows (see below).
12
It should be noted that in regions with strong intermittency (for example, near a jet
boundary) such a unique connection of the sign of the velocity derivative asymmetry and
the direction of the energy spectral ux is not found because in this region the approximation
of homogeneous and isotropic turbulence is not satised.
[15] Higher moment diusion in stable stratication 445
4 2 0 2
S
> 0
> 0
0
0.5
Figure 10: Distributions with coecients of skewness and excess being signif-
icant in comparison with those of the equilibrium (Gaussian) distribution.
It follows from the foregoing that to describe the turbulent state of a ow
it is sucient to dene cumulants of this expansion. However, following the
theorem of Marcinkiewicz (1939), the expansion of the characteristic function
can be a polynomial of second degree (normal distribution
13
) or an innite
series. It seems to be impossible to dene a complete (innite) set of coef-
cients. The innite chain of transport equations for cumulants (moments)
should be broken and closed by the application of models. It should be noted
that such a breaking can lead to incorrect results (obtaining negative values
of quantities which in practice, can only be positive: PDF, dispersions, energy,
dissipation). The last circumstance is a consequence that by the denition of a
nite number of cumulants (moments) the corresponding probability may not
even exist. It is connected closely with the inadmissibility of arbitrarily clip-
ping the Taylor series of the characteristic function logarithm.
14
In the present
chapter the assumption of a relaxation character of the PDF evolution to the
equilibrium state is applied (by analogy with the approximation used in a
range of problems in the kinetic theory of gases). This assumption means there
13
The normal distribution of the turbulent uctuations supposes their isotropy and homo-
geneity. Practically the only ow that reasonably approximates these conditions is the tur-
bulent ow behind a grid. The overwhelming majority of applied problems have anisotropic
external eects (stratication, rigid boundary). In such ows experiment xes signicant
values of the coecients of skewness and excess (characteristics reected in the deviation of
the PDF from the normal one). To describe the statistical structure of anisotropic turbulence
the Gaussian PDF approximation appears to be insucient.
14
A signicant number of papers (see, for example Schumann (1977)), is devoted to the
problem of the realizability of turbulence models.
446 Ilyushin
Figure 11: Spectrum of turbulent uctuations.
will be no strong deviations of the PDF from the equilibrium one.
Moreover, the Kolmogorov hypothesis concerning the locally isotropic struc-
ture of small-scale turbulence in the inertial range of the turbulent energy spec-
trum (see Figure 11) will be used as well. At present the correctness of this
hypothesis is not in doubt. The statistical structure of this region of the spec-
trum is described with good accuracy by the equilibrium PDF and a motion
of uid particles (or contaminant particles) in the eld of isotropic turbulence
can be characterized as a Brownian motion (diusion).
The short-wave part of the spectrum corresponds to scales of dissipation of
the turbulent kinetic energy into heat by viscous action. This region contains
a comparatively small part of the turbulence energy. Corresponding to this
region, uctuations have a complex statistical structure and are characterized
by considerable values of the skewness and atness coecients (with a strongly
[15] Higher moment diusion in stable stratication 447
non-Gaussian distribution). However, when modelling turbulent ows in the
frame of the method of statistical moments, their inuence on the processes of
turbulent transfer is believed to be negligible. The primary role of uctuations
in this range providing the sink of turbulent kinetic energy is taken into
account. In the case of developed turbulence, the rate of viscous dissipation is
equal with good accuracy to the spectral ux of turbulence energy and, for its
denition, does not require the consideration of a complex statistical structure
of small-scale turbulence.
In contrast, the energy range corresponding to long-wave (large-scale) uc-
tuations contains the main part of the turbulence energy and mainly deter-
mines the character of turbulent transport. Long-wave uctuations correspond
to large-scale eddy structures (LSES). They are characterized by a compar-
atively large time of relaxation and contain information about the history
and structure of the averaged ow (memory eects). Therefore, this region
of the spectrum is usually anisotropic and the PDF (being considerably non-
Gaussian) is characterized by substantial values of skewness and atness coef-
cients. Thus, uctuations of the energetic and inertial ranges of the spectrum
are the main objects of study in problems of turbulence modelling
15
.
References
Andre J.C., De Moor G., Lacarr`ere P. and Du Vachat R. (1976). Turbulence approxi-
mation for inhomogeneous ow: Part 1. The clipping approximation, J. Atmos. Sci.
33 476; Part 2. The numerical simulation of penetrative convection experiment,
J. Atmos. Sci. 33 482.
Andre J.C., De Moor G., Lacarr`ere P., Therry G., Du Vachat R. (1978). Modelling the
24-hour evolution of the mean and turbulent structures of the planetary boundary
layer, J. Atmos. Sci. 35 1861.
Byzova N.L., Ivanov V.N., Garger E.K. (1989). Turbulence in the boundary layer of
the atmosphere. Gidrometeoizdat, Leningrad. (in Russian).
Canuto V.M., Minotti F., Ronchi C., Ypma R.M., Zeman O. (1994). Second-order
closure PBL model with new third-order moments: comparison with LES data, J.
Atmos. Sci. 51 1605.
Caughey, S.J. (1977). Boundary layer turbulence spectra in stable conditions, Bound-
ary-Layer Meteorol. 11 3.
Caughey S.G. (1982). Observed characteristics of atmospheric boundary layer. In
Atmospheric Turbulence and Air Pollution Modelling, F.T.M. Nieuwstadt and H.
van Dop (eds.), Reidel.
15
The complex statistical structure of large-scale eddy formations on the one hand, and
their governing eect on the dynamics of a turbulent ow on the other, have led to the
development of the method of describing turbulence based on the exact resolution of large
eddies with the use of parameterizations to take into account small-scale uctuations (the
LES method). See Chapters [8], [12], [13], [25].
448 Ilyushin
Craft T.J., Kidger J.W., Launder B.E., (1997). Importance of third-moment mod-
elling in horizontal, stably-stratied ows. Eleventh Symp. on Turbulent Shear
Flows, Grenoble, France.
Finnigan J.J., Einaudi F., Fua D. (1984). The interaction between an internal gravity
wave and turbulence in the stably stratied nocturnal boundary layer, J. Atmos.
Sci. 16 2409.
Hanjalic K., Launder B.E. (1972). A Reynolds stress model of turbulence and its
application to thin shear ows, J. Fluid Mech. 52 609.
Hazen A.M. (1963). Towards a non-linear theory of appearance of turbulence, DAN
USSR. 163 1282 (in Russian).
Ilyushin B.B., Kurbatskii A.F. (1996). New models for calculation of third-order
moments in planetary boundary layer, Izv. RAN. Phys. Atmos. Ocean. 34 772 (in
Russian).
Keller L., Friedmann A. (1924). Die Dierenzialgleichungen f ur die turbulente Bewe-
gung einer Kompressible Flussigkeit. Proc. Int. Congr. Appl. Mech. (Delft).
Lamb R.G. (1982). Diusion in convective boundary layer. In Atmospheric Tur-
bulence and Air Pollution Modelling, F.T.M. Nieuwstadt and H. van Dop (eds.),
Reidel.
Lenschow D.H., Wyngaard J.C., Pennel W.T. (1980). Mean-eld and second-moment
budgets in a baroclinic, convective boundary layer, J. Atmos. Sci. 37 1313.
Lin J.T., Pao Y.H. (1979). Wakes in stratied uids, Ann. Rev. Fluid Mech. 11 317.
Marcinkiewicz, J. (1939). Sur une propriete de la loi de Gauss, Math. Zeitschr. 44
612.
Moeng C.-H., Randall D.A. (1984). Problems in simulating the stratocumulus-topped
boundary layer with a third-order closure model, J. Atmos. Sci. 41 1588.
Monin A.S., Yaglom A.M. (1967). Statistical Fluid Mechanics (Vol 1,2), Nauka (in
Russian) See also English translation (ed. J.L. Lumley) Vol 1 (1971) Vol 2 (1975)
MIT Press.
Pinus N.Z., Cherbakova E.F. (1966). About wind velocity eld in stratied atmo-
sphere, Izv. Acad. Sci. SSSR, Phys. Atmos. Ocean. 2 1126.
Raupach M.R. (1981). Conditional statistics of Reynolds stress in rough-wall and
smooth turbulent boundary layers, J. Fluid Mech. 108 363.
Schumann U. (1977). Realizability of Reynolds-stress turbulence models, Phys. Flu-
ids. 20 721.
Willis G.F., Deardor J.W. (1974). A laboratory model of the unstable planetary
boundary layer, J. Atmos. Sci. 31 1297.
Zaetz, P.G., Safarov N.A., Safarov R.A. (1985). Experimental study of turbulent ow
characteristics behaviour by channel rotation around longitudinal axis. In Modern
Problems of Continuous Medium Mechanics, MFTI, Moscow, 136, (in Russian).
16
DNS of Bypass Transition
P.A. Durbin, R.G. Jacobs and X. Wu
Although direct numerical simulation (DNS) databases have had an impact
on the development of models for fully turbulent ows, computer simulations
of bypass transition have only recently become available and their impact on
modeling has been very much less. The conditions of bypass transition make
direct simulation a natural resource; there is little doubt that its use in this
eld will continue to grow. Direct simulation is restricted to low Reynolds
number and to simple geometries because of its great demand on computer
speed and memory. These restrictions may not be a debilitating handicap for
many transition applications: by their very nature, transitional phenomena
occur at Reynolds numbers that are low compared to fully turbulent ow.
Since focus is on the vicinity of the transition line, geometrical complexity is
not a controlling factor.
Simulation databases increasingly will be used to develop and test models,
and to understand phenomena; for these purposes they provide highly resolved
data, with well dened inlet conditions. Computational data are often of higher
quality than analogous laboratory experiments. The geometry and the nature
of the external disturbances that provoke transition are prescribed analytically
and can be reproduced exactly in Reynolds averaged simulations. Free-stream
turbulence intensity and integral scale can be specied at the inlet. A wealth
of instantaneous ow elds and ensemble averaged data are available for basic
and applied inquiries. Instantaneous ow elds provide views of the develop-
ment of perturbations in the transitional zone and elucidate the initial stages of
coupling between free-stream eddies and boundary layer disturbances. These
are the phenomological underpinnings of the averaged data. The gures con-
tained herein provide a sampling of the phenomological and statical resources
of DNS.
Certain simulations have already begun to ll a niche left open by laboratory
experiments. The periodic, wake induced transition discussed in 3 is a case in
point. In some applications the simulation codes themselves might be a tool
to use for predictive purposes. For instance, bypass transition in low pressure
turbines is critical to some designs. The need for accurate prediction might
warrant direct simulation.
1 The role of bypass transition
Under a quiet free-stream, a laminar boundary layer becomes linearly unstable
beyond a critical Reynolds number at which TollmeinSchlichting waves start
449
450 Durbin, Jacobs and Wu
Figure 1: The ragged edge of the turbulent region is maintained by merging
spots. Three side by side copies of the same image are displayed.
to grow. The instability is via a subtle mechanism whereby viscosity desta-
bilizes the waves (Drazin and Reid 1981). That subtle mechanism produces
very slow growth. Transition to turbulence might occur 20 times farther down-
stream from the position of linear instability. Orderly transition occurs only
after the waves have become nonlinear and inviscid mechanisms have come
into play. Given the delicacy and slowness of the orderly transition route, it
is not surprising that the boundary layer will forgo this process when it is
subjected to relatively small disturbances. It is also not surprising that the
orderly route is beyond the scope of statistical turbulence modeling. Statisti-
cal models are inherently restricted to chaotic processes. What, then, is the
province of turbulence models in the eld of transition?
When a laminar boundary layer is exposed to about 1% or more free-stream
turbulence, the TollmeinSchlichting route to transition is bypassed. A rather
dierent scenario is then encountered: highly localized, irregular motions oc-
cur inside the boundary layer. These develop into small patches of turbulence,
called turbulent spots. Spots spread and intensify as they propagate down-
stream. Eventually they join onto a main turbulent region, thereby maintain-
ing its upstream edge. As the spots join the main region they form interlaced
zones of turbulent and laminar ow, just upstream of a fully turbulent region.
Figure 1 illustrates the ragged edge of the turbulent zone, as is maintained by
merging spots. This gure is three copies of a single image; the periodicity is
an artice.
The streamwise length in which bypass transition occurs is quite short com-
pared to the length of the orderly route; bypass transition is complete in a
distance comparable to the distance to its onset (Simon and Ashpis 1996). In
gure 3 the onset of transition is at about R
250 (Re
x
1.5 10
5
) and it
is complete by R
600 (Re
x
3 10
5
).
Figure 2 shows two turbulent spots in the transition zone, as extracted from
a direct numerical simulation. This gure consists of contours of the uctu-
ating u-component velocity in a plane near the wall. When portrayed in this
[16] DNS of bypass transition 451
x
z
0.875
1.0 0.75 1.125 1.25 1.375 1.5
0.05
0.10
0.15
0.20
0.00
flow
Figure 2: Spots in the transition zone as seen via contours of u near the wall.
manner, streamwise elongated structure within the oval patches of u
contours
becomes apparent. It is a generic property of strongly sheared perturbations
to develop elongated contours of u-component velocity uctuations. Turbulent
spots certainly can be detected in experiments, but high resolution space-time
data are hard to obtain by laboratory measurements. Numerical simulation is
quite the opposite. A plenitude of data are available; the diculties lie in how
to process and archive all that information.
The occurence of spots and precursors to spots is random in space and time.
Hence, bypass transition falls within the province of statistical uid dynamics.
Whether conventional turbulence models are applicable to this phenomenon
is an open question (Saville 1999), but the fact that it is stochastic is not. The
application of turbulence modeling to bypass transition will be discussed in
detail in other chapters of this volume. The present chapter is addressed to
the phenomenon as revealed by direct numerical simulation (DNS) and to the
potential for DNS to generate unique databases that can be used to develop
and assess models.
Under zero pressure gradient, when the level of turbulence is on the order of
3%, transition takes place near the critical Reynolds number of linear theory.
This may well be a reection of a global critical Reynolds number, rather than
a relevance of linear stability per se. Upstream of global instability, even large
perturbations of the boundary layer do not provoke transition. That upstream
region can be described as a buetted laminar boundary layer. The random
nature of this region puts it into the province of Reynolds averaged (RANS)
modeling. In most cases it is more important for the RANS model to capture
transition than for it to predict the buetted laminar region.
The virtues of numerical simulation include the wealth of full-eld instan-
taneous and ensemble averaged data that is provided, the ability to prescribe
precise inow conditions and geometry, and the ability to perform thought
452 Durbin, Jacobs and Wu
200 400 600 800 1000
Re
0
0.002
0.004
0.006
0.008
C
f
Blasius
Turbulent
T3A
DNS
Figure 3: DNS and experimental Cf.
experiments. Thought experiments are simulations of real uid dynamical
behavior that would be virtually impossible to achieve in a physical experi-
ment. Such experiments have already elucidated the dominance of the normal
component of free-stream velocity uctuations in provoking transition (Yang
and Voke 1995). They have also shown that turbulent uctuations in imping-
ing wakes exert far greater control of bypass transition than does the mean
velocity decit (Wu et al. 1999).
Drawbacks to DNS include its restriction to relatively low Reynolds numbers
and to small, simple geometries. It imposes substantial CPU and computer
memory requirements. The stringent requirements on computer resources are
rapidly being alleviated by massively parallel computers. The simulations de-
scribed herein were performed on computers with from 64 to 512 processors.
On the order of 5 10
7
grid points were used. A single run might require
10
2
hours of CPU time and be performed in 4 weeks of wall clock time.
Of course, post-processing the wealth of data takes place over a much larger
time period. This all is analogous to laboratory experimentation: data are
obtained with highly specialized equipment and are archived for subsequent
analysis.
2 Transition induced by free-stream turbulence
The simplest case of bypass transition is that induced by free-stream tur-
bulence passing above a at plate with no imposed pressure gradient. The
RollsRoyce T3a experiment has become the denitive data set for this con-
guration (Roach and Brierly 1990). Although our DNS experiments were not
designed to accurately reproduce the conditions of that experiment, a fairly
[16] DNS of bypass transition 453
u-component
v-component
Figure 4: Streaks, spots and transition in a boundary layer beneath free-stream
turbulence as evidenced in contours of uctuating u and v components of
velocity.
close approximation has been produced. Figure 3 compares ensemble averaged
skin friction versus momentum thickness Reynolds number from the DNS of
Jacobs and Durbin (2000) to the experiments of Roach and Brierly (1990).
The smallest x-location at which dC
f
/dx = 0 can be dened as the onset of
transition. At this location data elds like those in gure 4 show the onset
of turbulent spot precursors. These stand out more starkly in the v-contours
than in those of u. The spots themselves are initially highly intermittent. They
mature through the transition region. Downstream, where the skin friction in
gure 3 has reached the fully turbulent level, the spots merge into the contin-
uously turbulent boundary layer. Statistical elds are obtained by spanwise
and temporal averaging of random elds like gure 2.
The evolution of mean velocity proles with downstream distance is dis-
played in gure 5 by log-linear plots. These show that the trends of experi-
mental ensemble averaged data are closely mirrored in the DNS. The lower
portion of gure 5 is a comparison of the root-mean-square streamwise velocity
uctuation to measurements. The DNS elds are a very comprehensive data
set. Only selected locations are plotted for comparison to experiment; further
data are available in Jacobs and Durbin (2000). The contour plots in gure 6
illustrate the ability of DNS to provide full eld data in this case contours
of turbulent kinetic energy production are plotted; they show the increase in
production as the boundary layer undergoes transition.
In the present example of transition induced by free-stream turbulence, the
comprehensiveness of the DNS elds might be seen as the primary merit of
the statistical data. In section 3 below, on wake induced transition, the DNS
dataset lls a gap left by a lack of experimental data. However, even in the
present case, instantaneous ow elds obtained from the numerical simula-
tion provide views of the coupling between free-stream eddies and boundary
layer disturbances that could not be obtained in the lab. These instantaneous
pictures are the subject of the remainder of this section.
454 Durbin, Jacobs and Wu
10
-1
10
0
10
1
10
2
10
3
y+
0
10
20
30
U
+
Re
= 323
Re
= 385
Re
= 456
Re
= 980
Log Law
U proles
0 50 100 150 200
y+
0
1
2
3
u
+
Re
= 177
Re
= 385
Re
= 539
Re
= 897
u proles
Figure 5: DNS and experimental proles of U, u at a few streamwise location.
Symbols are the data of Roach and Brierly (1990).
2.1 Streaks, spots and shear ltering
An intriguing phenomenon occurs in the laminar boundary layer upstream of
the onset of transition. Long streaks of u-component velocity perturbations are
observed: see the u-component of gure 4. Streaks are a feature of zero pressure
gradient laminar boundary layers bueted by free-stream eddies. The streaks
appear to be identical to what have been called Klebano modes (Kendall
1991). The latter refer to proles of uctuating velocity that peak in the outer
part of the boundary layer. (Ironically, these modes are amplied by the
mean shear through a process of that has recently been renamed non-modal
growth. The term Klebano mode is a misnomer.) Figure 7 shows proles of
uctuating velocity in the streak region. These proles peak at y/
1 and
are in excellent accord with Klebano mode measurements.
[16] DNS of bypass transition 455
Re
y
/
0
1.0E+05 2.0E+05 3.0E+05 4.0E+05 5.0E+05
0
1
2
3
4
5
Figure 6: Contour plot rate of turbulent kinetic energy production.
The role of laminar streaks in the subsequent bypass transition is unclear.
Some insights are currently emerging from examination of DNS elds. Al-
though it has been proposed that the streaks might undergo a sinuous insta-
bility, leading to turbulent spots (Elofsson et al. 1999), no evidence of that was
seen in any of our direct simulations. Indeed the streaks in gure 4 persist up
to the turbulent zone with no indication of sinuous instability. Velocity vector
plots show that the streaky structure consists of narrow, intense regions where
u
tr
= 163 +e
6.91Tu
(2.1)
for the transition Reynolds number of zero pressure gradient boundary layers.
Tu is the turbulence intensity in percentage, 100u
Figure 8: Horizontal sections through the free-stream and the boundary layer.
of space and time; for instance the mean velocity is of the form U(x, y, ))
and the skin friction coecient is a function, C
f
(x, ), of downstream distance
and phase. The phase angle is equivalent to time, modulo the passing period
= 2(t/T )
mod 1
. In a lab experiment it is a challenge to accumulate these
data.
However, in DNS experiments averaging over samples both at constant
phase and in the direction of homogeneity enables statistical elds to be accu-
rately computed. A database of transition induced by period wakes has been
created by Wu et al. (1999). Reynolds averaged computations of these data
are discussed in Wu and Durbin (2000).
The ow conguration is illustrated by gure 9. This shows wakes in the
free-stream sweeping over the boundary layer. As time progresses the wakes
convect from left to right. As they convect downstream, the centerline of a
given wake traverses down a constant x section; indeed, the sloped wake was
created by traversing a wake down the inlet plane. Figure 9 illustrates how the
boundary layer thickens in the vicinity of x = 1.0, where it is highly perturbed
by the impinging wake. Downstream of that it thins before thickening again
beyond x = 1.5. In this furthest downstream region, the boundary layer has
become fully turbulent, but upstream it is a highly buetted, laminar layer.
The upper frame of gure 9 shows instantaneous contours of the uctuating
u-velocity. The central frame is a spanwise average at the same phase, while
the lower frame is a RANS computation. All three show local thickening of
the boundary layer beneath the wake within the laminar region and transition
to a thicker, turbulent boundary layer farther downstream. It is clear from
this gure that RANS models have the potential to represent, to a signicant
degree, the bueted laminar layer and the transition to turbulence. However,
too much should not be read into the level of agreement between RANS and
[16] DNS of bypass transition 459
Figure 9: Contours of streamwise velocity component: (a) instantaneous u over
one xy plane; (b) phase-averaged u); (c) unsteady RANS using v
2
-f model.
From Wu and Durbin (2000).
DNS in gures 9 and 11. For example, it is unlikely that pressure gradient and
curvature eects would be captured correctly at the level of modeling used in
these gures. The RANS computations were done with a model developed and
calibrated for fully turbulent ow (Wu and Durbin 1999), with no modica-
tion specically for transition. There is probably a need to modify turbulence
models if they are to be used in ows in which transition must be computed
with a high degree of accuracy. In other ows, the crude ability of turbulence
models to undergo a form of bypass transition might be sucient.
Prior to the fully turbulent zone, demarcated by boundary layer growth in
gure 9, pus are seen in contour plots of u near the wall (Wu et al. 1999).
These sometimes develop into turbulent spots, depending on how they interact
with the free-stream turbulence. The initial irregular motion that leads to
turbulent spots is created by certain free-stream eddies that cause jetting
motion near to the wall. The jet is in the upstream direction relative to the
mean ow. In gure 10 the phase averaged velocity has been subtracted from
the instantaneous velocity to show the motions that are responsible for spot
precursors. The gure shows a lifted shear layer that undergoes instabilities
on a rapid time-scale. The three sections of this gure are three successive
460 Durbin, Jacobs and Wu
Figure 10: Development of a backward jet in the relative velocity eld and its
breakdown to turbulence. The gures show the development of a spot precursor
and its breakdown via inviscid instability. These are successive instants, with
time increasing from top to bottom.
instants in time, with time increasing from top to bottom. They are spaced
by a time interval of 0.1 of the wake passing period, T . The time-step in
the simulation was 10
3
T . The x range portrayed moves downstream with
the wake. The topmost gure is suggestive of an eddy at x 1.0, y 0.01
triggering subsequent highly irregular motions. The instabilities that grow
on the negative jet develop small-scale eddies that evolve into the turbulent
spots.
[16] DNS of bypass transition 461
0 3.5
0
6
x
0 3.5
0
6
Figure 11: Contours of skin friction vs. x and . (a) phase-averaged u); (b)
unsteady RANS using v
2
f model. From Wu and Durbin (2000).
A plan view of velocity contours would show a nascent spot at the last in-
stant of gure 10. At a somewhat later stage a fully formed spot, as in gure 2,
would emerge. Although the occurence of turbulent spots in transitional ow
has been known from experiments, an understanding of the precursors to de-
veloped spots has only recently begun to evolve from numerical simulations
such at these. By the time turbulent spots are detected, the process has already
become chaotic.
Phase averaged data for periodic wake induced transition can be found in
Wu and Durbin (2000). Figure 11 shows C
f
(x, ) contours for a at plate
boundary layer beneath periodically passing wakes. The upper part of the
gure consists of DNS data; the lower shows a RANS simulation of the same
conguration. These are contour plots. In a slice parallel to the x-axis, at
constant , the contour elevations become a skin friction curve C
f
(x) at that
particular phase. Because the tongues between x 1 and x 2.25 in gure 11
462 Durbin, Jacobs and Wu
extend upstream at an angle to the axes, the constant phase section traverses
from low, to high, to low and back to high C
f
. The rst region of high C
f
occurs where the external wake impinges on the laminar boundary layer. The
periodic tongues in gure 11 are imprints upon the laminar boundary layer of
the passing wake. The boundary layer is highly perturbed but the wakes do
not cause transition until further downstream. Beyond x 2.5 the boundary
layer is fully turbulent. The periodic buetting by the passing wake is still
seen in the turbulent region, but there it is a minor eect.
Single point closure models are meant for use in Reynolds averaged com-
putational uid dynamics codes. They are routinely tested against a range
of experiments. The same should be done for bypass transition models. DNS
provides comprehensive databases to use in testing. In the present example
of wake induced transition, quantitative phase averaged elds of velocity, skin
friction and second order statistics are now available. Eects of pressure gra-
dients and curvature will be included in future simulations. These will greatly
enhance the ability to tackle computational predictions of complex engineering
ows.
References
Abu-Gannam, B.J. and Shaw, R. (1980). Natural transition of boundary layersthe
eect of turbulence, pressure gradient and ow history, J. Mech. Eng. Sci. 22,
213228.
Andersson, P., Berggren, M. and Henningson, D. S. (1999). Optimal disturbances
and bypass transition in boundary layers, Phys. Fluids 11, 134150.
Drazin, P. and Reid, W.H. (1981). Hydrodynamic Stability, Cambridge University
Press.
Hunt, J.C.R. and Durbin, P.A. (1999). Perturbed shear layers, Fluid Dyn. Res. 24,
375404.
Jacobs, R.G. and Durbin, P.A. (2000). Bypass transition phenomena studied by com-
puter simulation, report TF-77, Department of Mechanical Engineering, Stanford
University.
Jacobs, R.G. and Durbin, P.A. (1998). Shear sheltering and the continuous spectrum
of the OrrSommerfeld equation, Phys. Fluids 10, 20062011.
Kendall, J.M. (1991). Studies on laminar boundary layer receptivity to free stream
turbulence near a leading edge. In Boundary Layer Stability and Transition to
Turbulence, D.C. Reda et al. (eds.), ASME-FED, 114, 2330.
Liu, X. and Rodi, W. (1991). Experiments on transitional boundary layers with
wake-induced unsteadiness, J. Fluid Mech. 231, 229256.
Roach, P.E. and Brierly, D.H. (1990). The inuence of a turbulent freestream on zero
pressure gradient transitional boundary layer development, part I: test cases T3A
and T3b. In Numerical Simulation of Unsteady Flows and Transition to Turbu-
lence, O. Pironneau et al. (eds.), Cambridge University Press, 229256.
[16] DNS of bypass transition 463
Savill, A.M. (1999). One point closures applied to transition. In Turbulence and
Transition, M. Hallb ack et al. (eds.), Kluwer, 233268.
Simon, F. and Ashpis, D. (1996). Progress in modeling of laminar to turbulent
transition on turbine vanes and blades, NASA-TM 107180.
Wu, X. and Durbin, P.A. (2000). Boundary layer transition induced by periodically
passing wakes, J. Turbomachinery 122, 442449.
Wu, X., Jacobs, R.G., Hunt. J.C.R. and Durbin, P.A. (1999). Simulation of boundary
layer transition induced by periodically passing wakes, J. Fluid Mech. 398, 109
153.
Yang, Z. and Voke, P. (1995). Numerical study of bypass transition, Phys. Fluids 7,
22562264.
17
By-Pass Transition using Conventional
Closures
A.M. Savill
1 Introduction
In attempting to apply turbulence models to transitional ows most practi-
tioners either assume point transition at an estimated transition location,
and force a switch between laminar and turbulent computations at that point,
or make use of experimentally-derived correlations to estimate the start of
transition and then ramp-up the turbulent eddy viscosity by scaling this
with an empirical transition function until fully turbulent conditions are at-
tained. These two approaches are generally used in conjunction with mixing
length models or even simpler integral methods which remain the most com-
mon methods used today for everyday engineering design purposes.
However Priddin (1975) showed that a low-Re k- model treatment (due to
Jones and Launder (1972)) could be used to predict transitional ow as well.
Subsequently a few other successful low-Re k- model variant applications
were reported by Arnal et al. (1980), Dutoya and Michard (1981) and then
Liu (1989). Other encouraging results were obtained by Arad et al. (1982),
using the alternative two-layer k-kl model scheme of Ng, (1971), by Wilcox
(1992) using his alternative two-equation k- model, and even with a one-
equation k-l model adapted by Manartonakis and Grundmann (1991).
At the same time Rodi and Scheuerer (1984) used the low-Re k- model of
Lam and Bremhorst see Patel, Rodi and Scheuerer (1985) to perform the
rst extensive evaluation of such transition predictions for a range of simple
test cases and a turbine blade. For free-stream turbulence intensities > 1%
they found this model could quite accurately predict the observed by-pass
transition (provided the initial conditions were carefully controlled). But early
attempts to employ higher level Reynolds Stress Transport (RST) closures,
by Donaldson (1969) and Finson (1975) failed to give quantitatively similar
agreement with experiment.
The rst direct inter-comparison of dierent modelling techniques for pre-
dicting by-pass transition ranging from the simplest correlation methods to
direct numerical simulations was reported by Pironneau, Rodi, Ryhming,
Savill and Truong (1992). Two test cases were considered, based on data pro-
vided taken from a series of experimental studies conducted by researchers
at Rolls-Royce. These involved transition on a sharp-leading-edge test plate
464
[17] By-pass transition using conventional closures 465
(carefully controlled to avoid leading-edge separation and to maintain nom-
inally zero pressure gradient) under the inuence of precisely isotropic 3%
(T3A Test Case) and 6% (T3B) free-stream turbulence (fst).
A very wide range of results was obtained from the dierent computational
approaches adopted. Although industrial design methods incorporating the
well-known Abu-Ghannam and Shaw (1980) correlations for the start and end
of transition (which in fact tted the Rolls-Royce data well) provided adequate
predictions, it was clear that the success of more academic, but potentially
more predictive, transport model schemes was highly dependent on the exact
closure approximations adopted. However the best of these produced excellent
results, and the interest this generated led to the setting up of an ERCOFTAC
Transition Modelling Special Interest Group (SIG) Project to evaluate a wider
range of candidate turbulence models against a larger series of transition test
cases see Savill (1992, 1993a,b, 1994, 1996).
The present contribution is based on the work of this Transition SIG and
a number of similar projects to assess turbulence models for predicting tran-
sition. These include studies for natural convection ows, reported by Henkes
(1992); for pipe ow, as reported by Jackson and He (1995); and for relaminar-
ising ow reported by Viala et al. (1995); as well as a similar AFOSR/NASA-
led evaluation exercise on by-pass transition modelling by USA research groups
see Simon (1994) and others, all of which have suggested that the Launder-
Sharma low Re k- model oers generally the best predictions of any standard
eddy viscosity approach.
The ERCOFTAC Project results have proved particularly instructive be-
cause Test Case computations were initially done blind (i.e. without access
to, or indeed in some cases in advance of, Test Case data), using only specied
initial/boundary conditions. In each case experimental data of high quality
and detail, complemented by suciently resolved numerical simulations, were
provided to assess the predictions. The necessary industrial input was also
secured to ensure relevance of the comparisons. In addition participants were
subsequently encouraged to perform tests of the sensitivity of their results
to the imposed conditions, model constants, and mesh resolution. This was
particularly important since the very wide range of models being evaluated
precluded the specication of xed grids. Finally authors were asked to rene
or otherwise modify their model approaches before re-testing these on a wider
range of cases listed in the table below.
2 Applying turbulence models to prediction of tran-
sitional ows
In attempting to use (high-Re) turbulence models to predict transition one
can take two dierent approaches. One can either adopt the premise that the
transition process can be modelled as a superposition of laminar and turbulent
466 Savill
Table 1: ERCOFTAC Transition Modelling SIG Test Cases
T3A-: zero pressure gradient, 1% isotropic free-stream turbulence
(theoretical or experimental initial conditions)
T3A: zero pressure gradient, 3% isotropic free-stream turbulence
(theoretical or experimental initial conditions)
T3A+: zero pressure gradient, 3% isotropic fst, but variable Lue
(theoretical or experimental initial conditions)
T3B: zero pressure gradient, 6% isotropic free-stream turbulence
(theoretical or experimental initial conditions)
T3BLES: zero pressure gradient, 4.5% weakly anisotropic free-stream
turbulence (Simulated initial conditions)
T3B+: zero pressure gradient, 10% weakly anisotropic free-stream
turbulence (experimental initial conditions)
T3C1-5: pressure gradient representative of aft-loaded turbine blade
(expt. initial conditions; various Re): C1 (6%fst)
& C2 (3%fst, same design Re); C3 & C4 (3%fst , lower Re
without & with laminar separation) ; C5 (3%fst, higher Re)
T3D1-3: zero pressure gradient, 0.1% isotropic fst, following laminar
separation (experimental conditions; various Re)
T3E: strong favourable/adverse pressure gradient, 0.1%fst,
relaminarisation/retransition (experimental initial conditions)
T3F : weak & strong convex curvature, 0.7 & 2% free stream
turbulence (experimental conditions)
T3G : weak & strong concave curvature, 0.6% and 8.5% or 2.5%
fst (experimental conditions)
T3H : zero pressure gradient, 5% fst with heated surface
(experimental conditions, variable Re etc.)
T3L1-6: semi-circular leading edge, 0.2% free-stream turbulence
(experimental initial conditions; various Re, fst, etc)
T3K: Low-Speed (LP Rotor) Turbine Cascade, 4% free-stream
turbulence (experimental initial conditions)
NB. The T3A&B Test Cases have also been adopted for the NASA Project
ow in changing proportions of time, , that the ow is turbulent, i.e. taking
times a turbulent solution and (1 ) times laminar (where is an inter-
mittency factor varying from 0 at the start of transition to 1 at the end) or
introduce additional (low-Re) model approximations to handle the turbulence
development through transition.
The rst of these approaches could, in principle, be applied to any type
of transition (natural, TaylorG ortler, by-pass), but requires some empirical
input regarding the location of transition onset and the precise streamwise
variation of . It is often used to improve point transition methods by including
an allowance for the length-of-transition over which varies from 0 to 1.
[17] By-pass transition using conventional closures 467
The second approach requires that there should be some initial source of
turbulence activity within, and/or outside the initial (pseudo-laminar) shear
layer, and makes the inherent assumption that the transition is diusion-
controlled in the sense that it is triggered by diusion of free-stream tur-
bulence into the ow. Such by-pass transition only occurs for free-stream
turbulence intensity levels > 1%. It was given this name by Morkovin because
most of the development stages in natural transition are then by-passed and
transition to turbulence then occurs at a point where breakdown of free-stream
turbulence-induced streaks and secondary instability produces (in 2D) a line
of turbulent spots across the span of the ow. This is why most current in-
dustrial design methods still use a simple point transition approach whereby
an initial laminar boundary layer computation is triggered to turbulent at a
point determined either empirically or via an established correlation of relevant
experimental data.
The present contribution concentrates on the latter approach, because this
appears to oer potentially the greatest generality in terms of predictability
and range of applicability, but some discussion is included concerning the
former, not least because recent work has shown the advantages to be gained
from combining both approaches.
It has to be stressed however that neither approach considers many of
the individual mechanisms and stages of transition discussed for example by
Hallb ack et al. (1996). The low-Re modelling approach takes no specic ac-
count of receptivity, algebraic growth, secondary instability, or turbulent spots.
Instead, the modelling implicitly assumes that diusion of (generally isotropic)
free-stream turbulence leads to a build-up of weakly correlated turbulence ac-
tivity in the initial pseudo-laminar (essentially Blasius prole) boundary layer
and that transition is initiated once the local production of turbulence en-
ergy suciently exceeds the local dissipation rate, . As in the case of fully
turbulent ows, all low-Re models are rather insensitive to the actual turbu-
lent (spot) structure. Intermittency models do attempt to account for spot
formation and growth rates, with varying degrees of sophistication based on
an ever-expanding experimental database; but only a few of these make any
allowance for receptivity and algebraic growth, and even those that do, lump
these together with secondary instability eects in so-called sub-transition
corrections.
2.1 Low-Reynolds Number Transport Models
A very large number of low-Re one and two-equation models (k-l, k-, k-
and k-) have now been evaluated. These include most, if not all, of the best
known variants including the models of Abid, Birch/HassidPoreh; Chien,
Chang, DutoyaMichard, McDonald and Fish, Hanjalic, Jones and Launder,
Lam and Bremhorst, LaunderSharma, LienLeschziner, Michelassi, Myong
Kasagi, NaganoHishida, NaganoTagawa, KasagiSikasano, Biswas and Fuku-
468 Savill
yama, Norris and Reynolds, Shih, To and Humphrey, Yang and Shih, Wilcox
and Wolfshtein (for summary presentations and cross comparisons of formula-
tions see Patel et al (1985), Pironneau et al (1992) and Henkes (1992)). Each
version introduces dierent low-Re extensions to the standard high-Re tur-
bulence model equations for turbulence energy k =
1
2
_
u
2
+v
2
+w
2
_
and
dissipation, , (or length scale) combinations see for example Hallb ack et al.
(1996) within the following basic set of low-Re equations for 2D shear ow:
Dk
Dt
=
t
_
U
y
_
2
+
__
+
t
k
_
k
y
_
ADVECTION PRODUCTION VELOCITY DIFFUSION
LAMINAR & TURBULENT
+
k
PRESSURE DISSIPATION
DIFFUSION
(1)
D
Dt
= E +f
1
C
k
_
U
y
_
2
+
y
__
+
t
_
y
_
+
f
2
C
2
k
(2)
(for numerical convenience often written in terms of which equals 0 at
the wall boundary) where
= f
k
2
/ and = D; D =
wall
,= 0;
E = C
t
_
2
U/y
2
_
2
. The quantities E, f
, f
1
and f
2
are low-Re damping
factors, and C
1
, C
2
and C
3
model constants, which vary for dierent low-Re
formulations.
The associated eddy viscosity and dissipation damping factors (f
and F
)
are always functions of the Reynolds-number formulations adopted by the
various model developers:
R
y
k
1/2
y
; y
+
; R
y
(/)
1/4
; R
t
k
2
; R
L
k
3/2
=
l
.
It must be stressed that not all variants require inclusion of the full set of low-
Re damping factors, and/or additional D and E factors in the representative
velocity and length-scale transport equations, to ensure the modelling can be
extended right through the buer layer where high-Re models normally employ
wall functions.
It should also be noted that all current low-Re, eddy-viscosity models were
originally developed to handle low-Re near-wall regions of turbulent ows and
that no models have so far been proposed primarily to model low-Re transi-
tion regions. This applies equally to low-Re Reynolds-stress transport (RST)
schemes.
2.2 Alternative Correlation and Intermittency-Scaling Models
Intermittency weighting for transitional ows has until recently only been
employed in simple closure models, at least for design purposes. In particular
[17] By-pass transition using conventional closures 469
a number of dierent integral/correlation methods have been developed and
used to some eect. The rst moderately successful model was that of Dhawan
and Narasimha (1958), which determined the onset of transition (at Re
xs
) by
extrapolating Narasimhas own assumed universal intermittency distribution
as a function of x, and proposed a correlation for the length of transition, ,
in terms of Re
xs
itself.
Other well-known examples are the methods of Fraser and Milne (1986),
which uses the Abu-Ghannam and Shaw (1980) correlation for the start of
transition and their own correlation for transition length, and the Narasimha
and Dey (1989) approach, which uses Narasimhas alternative correlations for
Re
xs
coupled with his later spot-formation model calibrated against a slightly
modied version of Dhawan and Narasimhas correlation for transition length.
More recently Costa and Arts (1991) have modied Dhawan and Narasimhas
model to account also for the eects of acoustic excitation at discrete frequen-
cies for low fst (Tu = 0.4%).
The same type of intermittency-weighting has now been utilised in a wide
range of mixing-length models. Boyle (1990) has obtained some reasonable
predictions for heat transfer on a range of turbine-blade test cases by
introducing the standard correlations of Dunham (1972), Seyb (1971) and
Abu-Ghannam and Shaw (1980) for Re
xs
, and Dhawan and Narasimhas
correlation for Re
t
_
U U
o
__
x
and
S
= C
(/k)S
k
,
where C
k
= 60, C
= C
1
C
k
, denotes the local boundary layer thickness and
U U
o
an entrainment velocity.
However, in practice it was found that C
-dependence;
- - converted to R
t
dependence and; with in place of also.
the zero and variable pressure gradient cases as the LaunderSharma model
see Savill (1993a, 1993b).
The value of switching from R
y
to R
t
damping factors has also been in-
dicated by the success of the Dawes model which assumes R
t
= 0.4R
y
(that
simulations show is the case for the very near wall region of turbulent ows), in
order to adapt a Reynolds (f
) hybrid y-dependent
low-Re model for use in an unstructured code.
(Note that Rodi and colleagues have shown that it is also possible to obtain
very good predictions with the original R
y
-dependent Lam and Bremhorst
model, at least in parabolic boundary layer computations, provided specic
empirical adjustments are made to take account of free-stream turbulence
level and pressure gradient variations again see Pironneau et al. 1992).
Regarding the key deciencies noted for the LaunderSharma model (and
its closest equivalents):
(a) Transition length prediction deciencies may be rectied by introducing
a Production Transition Modication (PTM), whereby the production
of turbulence energy is articially limited by an empirical scaling through
transition to ensure a close t of the predicted end of transition with the
correlation of Abu-Ghannam and Shaw, although the same result can be
obtained by introducing an intermittency factor see Simon (1994).
(b) It had been thought that there would be even greater benets associ-
ated with the switch from to ( 1/) as the length-scale depen-
dent variable, since it is known that the k- model does not require a
[17] By-pass transition using conventional closures 479
full low-Re implementation to handle turbulent near-wall ows and is
therefore considerably less demanding of grid resolution for boundary
layers. However there are then problems with the imposition of a free-
stream cut-o boundary condition and low-Re damping factors are still
required to handle transitional ows. Although the Wilcox (1992) low-
Re k- model employs solely R
t
-dependent damping factors, predictions
obtained at UMIST and elsewhere within the SIG, for the start and end
of transition in the T3A & B cases have proved disappointing.
(c) All k- models exhibit signicant discrepancies in their predictions for
the eect of variable pressure-gradient on transition. As indicated above,
this is also a problem encountered with eddy viscosity models in gen-
eral, and correction factors need to be introduced to sensitise them to
individual strain rates in transitional and turbulent ows. Unfortunately
although several empirical connections have been tested, these have so
far been found to lack generality for more complex ows. The same is
true of the strain-dependent RNG k- model which, although derived
from a purely theoretical (Renormalisation Group Theory) analysis of
the NavierStokes equations and dierent simplifying assumptions, only
provides an alternative, equally approximate, viscosity transport scheme,
(albeit with directly implied tunable strain-dependent constants), and
has not shown any advantages for transitional ow predictions.
Some progress has been made in increasing the sensitivity of eddy viscos-
ity models to the irrotational straining encountered in stagnation regions,
but the modications employed are not suciently general to also handle
subsequent shear-layer curvature and pressure-gradient eects. For variable
pressure-gradient ows, greater success has been obtained using a two layer k-
/k-l treatment with the near-wall prescribed length scale providing a match-
ing condition for the k- model employed in the outer boundary layer. How-
ever, this approach necessitates the inclusion of empirical correlations which
are better dened for adverse than favourable pressure gradients and therefore
it is of limited predictive capability, particularly where strain history eects
are important.
3.4 Non-Linear Viscosity Models
To improve predictions with isotropic eddy-viscosity transport models requires
better accounting for the local strain-eld eects encountered even in sim-
ple attached variable-pressure-gradient boundary layers. To handle those sub-
jected to convex and concave streamline curvature (and in addition swirl and
rotational eects on shear layers in general) it is clear that some kind of for-
mal anisotropic extension to the Boussinesq stress-strain relation is necessary).
This requirement has led to the development of a range of explicit non-linear
480 Savill
Figure 7: Nonlinear k- model predictions () versus linear LaunderSharma
model (- -) and experimental results for 3%fst () and 6%fst ().
k- (and even k-) models with terms up to (most commonly) quadratic, but
sometimes cubic, order in both the rotational and shear strain invariants; as
well as some alternative (implicit) algebraic-stress-model formulations derived
from full Reynolds-stress-transport model closures. Careful analysis has shown
that the cubic level treatment is required to capture both curvature and ro-
tational eects, but a suciently general description can only be obtained
by extending to at least quartic level, which has yet to be done. The UMIST
group (Craft et al. 1997) and others have demonstrated that both a 2-equation
(low Re k- with additional strain-dependent f
3
is in fact half the value applied in the
original KLY model for fully turbulent ow and this scaling of f
ensures a
very close t to the transition simulations of Voke and Yang (1993), within
the transitional buer layer to y
+
= 40 (which is approximately equivalent
to
+
in the pre-transitional boundary layer). Outside this region the E term
quickly becomes negligible.
This SLY model can accurately predict the onset of transition under nomi-
nally zero pressure-gradient conditions for levels of free-stream turbulence in
the range 1 to 10% (see Figure 8). Equally good predictions are obtained for
C
f
, H, and , and hence for the mean velocity prole development in all cases.
The growth of the initial u
= 10,000).
Certainly for the 3 and 6%fst test cases, the length of transition and shape
factor distribution, are predicted slightly better than with the LaunderSharma
[17] By-pass transition using conventional closures 483
Figure 9: Comparison of T3B test case experimental () and DNS () steam-
wise growth of turbulence intensity with predictions using LaunderSharma
k- model (- - - -) and SavillLaunderYounis (SLY) RST model with ()
and without ( ) non-local model extensions.
k- model and the same is true of the turbulent prole development although
the maximum value of k is still underpredicted compared to both the experi-
mental and simulation results as illustrated for the 6%fst case in Figure 9.
Similar results have been obtained with the LaunderShima low-Re RST
treatment for this case see Figure 10 probably because the f
damping
introduced in the SLY scheme only has an indirect eect through the E factor.
However, it has been found that modifying the E factor to provide a better t
to low-Re near-wall DNS data for turbulent ows as done in the Launder
Shima model does not improve transition predictions.
In comparison, the SLY model exhibits exactly the correct sensitivity to
pressure gradient for the on-design aft-loaded turbine-blade test case with
3%fst and also captures the correct trend of varying Reynolds number for the
pressure gradient distributions representative of the same aft-loaded turbine
o-design (T3C3-5 Test Cases), greatly improving on results obtained with
both standard and corrected LaunderSharma k- model treatments see
Figure 11.
The SLY model has also been shown capable of handling both relaminarisa-
tion and retransition of a turbulent boundary layer see Figure 12 provided
the same degree of mesh resolution is maintained as the boundary layer thick-
484 Savill
Figure 10: Comparison of LaunderShima RST () and SavillLaunder
Younis RST ( ) model predictions for 6%fst experimental test case T3B ().
ness varies (dramatically in this case). In addition it has been found to predict
the eect of a three-fold variation in free-stream length scale for zero pressure
gradient 3%fst and the separate inuence of both weak and strong streamline
curvature see Savill (1995).
4 Towards practical computations for engineering
ows
Several of the k- models evaluated within the ERCOFTAC Transition SIG
Project have now been applied to additional heat transfer and blading test
cases see Figures 13 and 14. In nearly all cases computations have been
initiated near the leading edge for a 2D or mid-span computations. Again the
LaunderSharma model has been found to provide the best overall predictions
of the un-modied models that have been evaluated, although even better re-
[17] By-pass transition using conventional closures 485
Figure 11: Comparison LaunderSharma k- model, with () and without
( ) pressure gradient correction versus SavillLaunderYounis RST ( )
model predictions for variable fst and pressure-gradient representative of an
aft-loaded turbine blade operating at various experimental conditions ().
sults have been obtained by introducing PTM scaling and similar alternative
R
t
-dependent low-Re treatments. Good results have also been reported with
a multi-scale k- approach being developed by Crawford see Simon (1994).
In some cases superior predictions have been obtained with a PTM version of
the R
y
-dependent Lam and Bremhorst scheme, although only after introduc-
ing some empirical input . This last model has also been successfully applied
in modied form to engine ows by researchers at General Electric see Dailey
et al. (1994). However it should be noted that the very early successful predic-
tions reported by Rodi and Scheuerer (1984) for heat transfer variation with
increasing free stream turbulence were possibly misleading. It now appears
that the reproduction of the experimentally observed shift to a monotonic
decay of Stanton number may have been due to the LamBremhorst model
predicting a similar, but far too early C
f
transition.
1
1
A more recent application of non-linear and linear eddy viscosity models to compute
heat-transfer coecients over a cascade of gas-turbine blades has been reported by Craft et
al. (1997). This shows very clearly the superiority of the non-linear EVM which avoids the
excessive heat-transfer coecients generated by the linear model both in the vicinity of the
stagnation point and on the pressure surface towards the downstream end of the blade. (Ed)
486 Savill
Figure 12: Comparison of SLY model predictions (, - - -) and measured C
f
results () from T3E test case (relaminarisation and retransition).
Some 3D computations have recently been undertaken for the ERCOFTAC
linear turbine cascade (T3K) test case, using a variety of closures including the
Dawes low-Re k- model. However even with up to 1.2 million cells the predic-
tions still appear to be limited by the degree of boundary-layer grid resolution
that can be incorporated in the computations and the model treatment used
to handle the leading edge stagnation region. At least as good results have
been obtained by Gregory-Smith (1995) and colleagues at Durham, using a
simple mixing length model and assuming laminar conditions throughout the
blade passage up to 80% chord their experimentally determined location of
transition on the suction surface.
RST models such as the SLY scheme need to be tested more extensively,
particularly within an elliptic framework for nite leading edge ows such as
the T3L case, prior to 2D and 3D blading applications, but already exhibit
the correct response to both streamline and in-plane curvature.
It can therefore certainly be anticipated that the RST approach will even-
tually provide superior predictions to k- models for the T3L, T3K and other
blading-type test cases, with nite leading edges, because such stress transport
models correctly predict only a small growth in turbulence energy along stag-
nation streamlines. Already acceptable engineering results have been achieved
for many turbomachinery ows using the Dawes k- model combined with the
Durbin treatment for irrotational strain and curved ow regions.
[17] By-pass transition using conventional closures 487
Figure 13: Model predictions for Daniel Turbine Blade test case: Launder
Sharma (LS), LamBremhorst (LB) both with Production Transition
Modication (PTM), Crawford Multiscale (MTS), and standard Launder
Sharma (LS).
Successful prediction of attached-ow by-pass transition can also be ex-
tended to laminar separation-induced transition by the inclusion of appropri-
ate alternative transition correlation criteria or, with progressively less em-
piricism, intermittency presumptions and intermittency transport see Savill
[18].
Acknowledgements
This contribution would not have been possible without the continued support
of Rolls-Royce plc, DRA, the CEC & COST. Several developments owe their
origins to discussions with close colleagues at UMIST (especially Professor
Brian Launder), researchers at the University of Surrey (in particular Profes-
sor Peter Voke) and Dr John Coupland of Rolls-Royce Aerothermal Methods;
with whom regular interaction has been maintained as part of an overall tran-
sition strategy. The assistance of Dr B.A. Younis of City University, who freely
provided the original version of the low-Re RST code and assisted with the
implementation of this, has been vital to the development of the SLY model, as
have contacts with other researchers, too many to mention individually, whose
help is also gratefully acknowledged. The Editors, organisers of the INI Pro-
gramme and other colleagues contributed a number of constructive criticisms
which I hope are reected in this nal version of my notes.
488 Savill
Figure 14: Model predictions for Daniel Nozzle Guide Vane test case ();
computations: LaunderSharma (+), LamBremhorst (), NaganoTagawa
(), KasagiSikasano (), and BiswasFukuyama ().
References
Abid, R. (1993) Evaluation of two-equation turbulence models for predicting transi-
tional ows, Int. J. Engineering Science 11 25.
Abu-Ghannam, B.J. and Shaw, R. (1980) Natural transition of boundary layers: the
eects of turbulence, pressure gradients and ow history, J. Mech. Eng. Sci. 22,
213.
Arad, E., Berger, M. and Wolfshtein, M. (1982) Numerical calculations of transitional
boundary layers, Int. J. Num. Methods Fluids 2 123.
Arnal, D. (1991) Transition description and prediction. Numerical simulation of un-
steady ows and transition to turbulence. In Proc. 1st ERCOFTAC Workshop on
Numerical Simulation of Unsteady Flows Transition to Turbulence and Combus-
tion, Lausanne, O. Pironneau et al. (eds.), Cambridge University Press, 304316.
Arnal, D., Habiballah, M. and Delcourt, V. (1980) Synth`ese sur les methodes de
calcul de la transition developpees `a DERAT. ONERA, Rapp. Techn., OA No.
11/5018.
Baldwin, B. and Lomax, H. (1978) Thin-layer approximation and algebraic model
for separated turbulent ows, AIAA Paper 78257.
Biswas, D. and Fukuyama, Y. (1994) Calculation of transitional boundary layers
with an improved low-Reynolds number version of the k- model, ASME Journal
Turbomachinery 25 11.
[17] By-pass transition using conventional closures 489
Blair, M.F. and Werle, M.J. (1981) Inuence of free-stream turbulence on the zero-
pressure gradient, fully turbulent boundary layer, AFOSR Technical Report 81
0514.
Boyle, R.J. (1990) NavierStokes analysis of turbine heat transfer. NASA-TM-102496.
Cebeci, T. and Smith, A.M.O. (1974) Analysis of Turbulent Boundary Layers Aca-
demic Press.
Chevray, R. and Tutu, N.K. (1978) Intermittency and transport of heat in round jet,
J. Fluid Mech. 88 133.
Costa, J. and Arts, T. (1991) Boundary layer transition under the presence of discrete
frequencies in the free-stream turbulence, Proc. 10th ISABE 2 807.
Craft, T.J., Launder, B.E. and Suga, K. (1997) The prediction of turbulent transi-
tional phenomena with a non-linear eddy-viscosity model, Int. J. Heat Fluid Flow
18 1528
Dailey, L. D., Jennions, I. K. and Orkwis, P. D. (1994) Simulated laminar-turbulent
transition with a low Reynolds number k- turbulence model in a NavierStokes
ow solver, AIAA Paper 94-0189.
Dawes, W.N. (1992) The simulation of three-dimensional viscous ow in turbo-
machinery geometries using a solution-adaptive unstructured mesh methodology,
J. Turbomachinery 114 (3) 528537.
Dhawan, D. and Narasimha, R. 1958) Some properties of boundary layer ow during
transition from laminar to turbulent motion, J. Fluid Mech. 3 418.
Dick, E. (ed.) (1995) Theme section on transition, ERCOFTAC Bulletin 24.
Donaldson, C. DuP. 1969) A computer study of an analytical model of boundary
layer transition, AIAA J. 7 271278.
Dopazo, C. (1977) On conditional averages for intermittent turbulent ow, J. Fluid
Mech. 81 433.
Dunham, J. (1972) Prediction of boundary layer transition on turbomachinery blades,
AGARDO graph 164, 55.
Dutoya, D and Michard, P (1981) A program for calculating boundary layers along
compressor and turbine blades. In Numerical Methods in Heat Transfer (eds. R.W.
Lewis, K. Morgan and O.C. Zienkiewicz), Wiley, 413428.
Finson, M.L. (1975) A Reynolds stress model for boundary layer transition. Phys.
Sc. Inc., Rep.-No. TR-34.
Fraser, C.J. and Milne, J.S. (1986) The eect of pressure gradient and free-stream tur-
bulence intensity on the length of transitional boundary layers, Proc. Inst. Mech.
Eng. 200 (C3), 179.
Ganey, R.L., Salas, M.D. and Hassan, H.A. (1990) An abbreviated Reynolds stress
turbulence model for airfoil ows, AIAA Paper 90-1468.
Gostelow, J.P., Hang, G., Walker, G.J. and Dey, J. (1994) Modelling of boundary
layer transition in turbulent ows by linear combination integral method, ASME-
94-G1-358.
Gregory-Smith, D. G. (1995) 3D ow simulation in turbomachinery, Proc. VKI Tur-
bomachinery Conference, Von Karman Inst. Rhode-St-Gen`ese Belgium.
490 Savill
Hanjalic, K., Jakirlic, S. and Hadzic, I. (1997) Expanding the limits of equilibrium
second-moment closures, Fluid Dynamics Research 20 25-41.
Henkes, R.A.W.M. (1992) Test case on natural convection, Proc. EROCFTAC/IAHR
Seminar on Assessment of Turbulence Models of Engineering Applications, EPF
Lausanne.
Hallb ack, M., Henningson, D. S., Johansson, A. J. and Alfredsson, H. (eds.) (1996)
Turbulence and Transition Modelling, Kluwer.
Jackson, J.D. and He, S. (1995) Simulation of transient turbulent ow using various
2-equation low-Reynolds number turbulence models, Proc. 10th Turbulent Shear
Flows Symposium, Pennsylvania State University.
Jones, W.P. and Launder, B.E. (1972) The prediction of laminarization with a two-
equation model of turbulence, Int. J. Heat Mass Transfer 15 301314.
Kebede, W., Launder, B.E. and Younis, B.A. (1985) Large amplitude periodic pipe
ow: a second-moment closure, Proceedings 5th Symposium Turbulent Shear Flows,
16.2316.29, Cornell Univ.
Launder, B.E., and Cho, J.R. (1994) The modelling of diusion-controlled transi-
tion with the new model, Proc. 6th Biennial Colloquium on Computational Fluid
Dynamics, UMIST.
Launder, B.E., Reece, G.J. and Rodi, W. (1975) Progress in the development of a
Reynolds-stress turbulence closure, J. Fluid Mech. 68 537.
Launder, B.E. and Shima, N. (1989) Second-moment closure for the near-wall sub-
layer: development and application, AIAA J. 27 (10) 13191325.
Libby, P.A. (1975) On the prediction of intermittent turbulent ows, J. Fluid Mech.
68 273.
Libby, P.A. (1976) Prediction of the intermittent turbulent wake of a heated cylinder,
Phys. Fluids 19 494.
Liu, S.L. (1989) The prediction of boundary layer transition using low Reynolds
number k- turbulence model, Proc. 4th Asian Congress of Fluid Mechanics, Hong
Kong, Vol. 1, A9.
McDonald, H. and Fish, R.W. (1973) Practical calculation of transitional boundary
layers, Int. J. Heat Mass Transfer 16 1729.
Manartonakis, G. and Grundmann, R. (1991) Transition in three-dimensional bound-
ary layers with a one-equation model, Proc. Boundary Layer Transition and Con-
trol Conference, Cambridge, UK, RAeS Publishers.
Micheltree, R.A., Salas, M.D. and Hassan, H.A. (1990) One-equation turbulence
model fo transonic airfoil ows, AIAA J. 28 16251632.
Myong, H. K. and Kasagi, N. (1990) A new approach to the improvement of k-
turbulence model for wall-bounded shear ows, JSME J. Series II 33 (1), 63.
Nagano, Y. and Hishida, M. (1987) An improved form of the k- model for wall
turbulent shear ows, Trans. ASME J. Fluids Engineering 109 156.
Nagano, Y. and Tagawa, M. (1990) An improved k- model for boundary layer ows,
Trans ASME J. Fluids Engineering 112 33.
[17] By-pass transition using conventional closures 491
Narasimha, R. and Dey, J. (1989) Transition-zone models for two-dimensional bound-
ary layers, Indian Academy Proceedings in Engineering Sciences: Sadhana 14 (2),
93.
Ng, K.-H. (1971) Predictions of Turbulent Boundary-Layer Developments using a
Two-Equation Model of Turbulence. PhD thesis, Imperial College, London.
Patel, V.C., Rodi, W. and Scheuerer, G. (1985) Turbulence models for near-wall and
low-Reynolds number ows: A review, AIAA J. 23 1308.
Pironneau O., Rodi W., I.L. Ryhming, Savill A.M. and Truong T.V. (eds.) (1992)
Proc. 1st ERCOFTAC Workshop on Numerical Simulation of Unsteady Flows
Transition to Turbulence and Combustion, Lausanne, Cambridge University Press;
see in particular A.M. Savill: A synthesis of T3 test case predictions.
Priddin, C.H. (1975) The Behaviour of the Turbulent Boundary Layer on Curved
Porous Walls. PhD thesis, Imperial College, London.
Prihoda, J., Hlava, T. and Kozel, K. (1999) Modelling of bypass transition including
pseudolaminar part of the boundary layer, ZAMM 79 S3, 5699700.
Rodi, W. and Mansour, N.N. (1993) Low-Reynolds number k- modelling with the
aid of direct simulation data, J. Fluid Mech. 250 509529.
Rodi, W. and Scheuerer, G. (1984) Calculation of laminar-turbulent boundary layer
transition on turbine blades, AGARD CP 390 on heat transfer and cooling in gas
turbines, 18-1.
Savill, A.M. (1992) Evaluating turbulence model predictions of transition an ER-
COFTAC Special Interest Group project, Appl. Scientic Research 51 555.
Savill, A.M. (1993a) Some recent progress in the turbulence modelling of by-pass
transition. In Near-Wall Turbulent Flows (eds. R.M.C. So and B.E. Launder) El-
sevier, 829.
Savill, A.M. (1993b) Further progress in the turbulence modelling of by-pass transi-
tion. In Engineering Turbulence Modelling and Experiments 2 (eds. W. Rodi and
F. Martelli) Elsevier, 583.
Savill, A.M. (1994) Transition modelling for turbomachinery II, Summary Proceed-
ings of the 1st ERCOFTAC Transition SIG Workshop of the BRITE-EURAM
AERO-CT92-0052 Project Workshop on Transition in Turbomachinery, VUB.
Savill, A.M. (1995) Predicting transition with turbulence models: the eect of variable
free-stream turbulence length scale. In Advances in Turbulence VI, S. Gavrilakis,
L. Machiels and P.A. Monkewitz (eds.), Kluwer, 7980.
Savill, A.M. (1996) Evaluation of turbulence models for predicting transition in tur-
bomachinery ows, Minowbrook II Workshop on Boundary Layer Transition in
Turbomachines, Sycracuse University, USA.
Schubaer, G.B. and Klebano, P.S. (1956) Contributions on the mechanics of bound-
ary layer transition, NACA Technical Note 3489.
Seyb, N.J. (1971) Transitional boundary layers, Rolls-Royce Report, Bristol Engine
Division.
Simon, F.F. (1994) Proceedings of the AFOSR Workshop on End-Stage Boundary
Layer Transition, Minowbrook, Syracuse University, USA.
492 Savill
Simon, F.F. and Stephens, C.A. (1991) Modelling of the heat transfer in bypass
transitional boundary layer ows. NASA-TP-3170.
Singer, A., Dinavahi, S.P.G. and Iyer, V. (1991) Testing of transition-region models:
test cases and data. NASA CR 4371 see also Trans. ASME JFE 114 73.
Steelant, J. and Dick, E (1995) Transition modelling by conditionally averaged equa-
tions, ERCOFTAC Bulletin 24 6264.
Theodoridis, G., Prinos, P. and Goulas, A. (1991) Prediction of transition boundary-
layer ows: a comparison of various low-Re turbulence models, Proc. 10th ISABE
2, 16.3.
Viala, S., Deniau, H., Hambruger, J. and Aupoix, B. (1995) Prediction of boundary
layer relaminarization using low-Reynolds number turbulence models, Proc. 10th
Turbulent Shear Flows symposium, Pennsylvania State University.
Voke, P.R. and Yang, Z.Y. (1993) Numerical solutions of the mechanisms of by-pass
transition in the at plate boundary layer, Proc. 9th Symp Turbulent Shear Flows,
Kyoto, 21-2-1.
Westin, K.J.A. and Henkes, R.A.W.M. (1997) Application of turbulence models to
bypass transition, J. Fluids Eng. 119 (4), 859866.
Wilcox, D.C. (1992) The remarkable ability of turbulence model equations to describe
transition, SAE Conference on Numerical and Physical Aspects of Aerodynamic
Flows, Paper SAE-92, Long Beach, CA.
Yang, Z. and Shih, T.H. (1991) Extending turbulence modelling to bypass transition,
NASA CMOTT Research Briefs 83.
18
New Strategies in Modelling By-Pass
Transition
A.M. Savill
1 Introduction
The preceding chapter [17] has discussed the various earlier attempts to predict
by-pass transition using conventional turbulence-model closure schemes across
the full range of complexity from simple mixing-length/correlation approaches
to Large Eddy Simulations (LES). A perspective has thus been provided on the
state-of-the-art as evidenced by research journal publications and actual in-
dustrial applications, which accurately reects developments in both academia
and industry during the twenty-year period from the early 70s to mid 90s.
However, over the last ve years a number of existing established method-
ologies have been signicantly rened in the light of improved physical un-
derstanding emerging from further experimental studies and, more especially,
new numerical database developments. Moreover some additional alternative
approaches have also now been proposed, researched and, in some cases, tested
with greater success. It is these newer strategies for turbulence modelling of
by-pass transition, which have not yet reached either the archival research lit-
erature or full industrial implementation, that form the subject of the present
review.
2 Results from by-pass transition simulations
A number of (fully resolved) simulations have now been carried out for ows
undergoing by-pass transition. Some initial rather coarse mesh zero pressure
gradient Direct Numerical Simulations performed by Yang and Voke for the 1st
ERCOFTAC Workshop see Pironneau et al. (1992) in fact very accurately
reproduced the experimental behaviour observed in experimental test cases
with 3 and 6% free-stream turbulence (fst). Encouraged by this success, a large
amount of the UK National Supercomputer resources (more than 10,000hr
CPU on a Cray YMP) has been devoted to establishing a suciently ne mesh
by-pass transition simulation database for a (T3BLES) Test Case similar to
the Rolls-Royce T3B (6%fst) experiment. In order to avoid the sub-grid-scale
modelling approximations as far as possible, this simulation has been run as a
Direct Numerical Simulation so far as the low-Re transitional boundary layer is
493
494 Savill
concerned, but a Smagorinsky-type SGS model has been employed to simulate
the development of the higher Re free-stream turbulence which was extracted
from a precursor simulation.
In addition Yang and Voke have now also performed a number of coarser
mesh Large-Eddy Simulations for the ERCOFTAC T3A, T3B, T3C1, T3D2
& T3L Test Cases see Table 1 of [17] and Savill (1995). Each simulation
appears to capture the essential features of the development of the mean ow
and turbulence elds, including, in the last case, the detailed features of the
laminar separation region, and a second ne resolution data base is now being
created for this as part of a current European Thematic Network project see
Savill and Pittaluga (2000).
In parallel the same researchers have carried out a series of LES numerical
experiments for variations on the T3BLES case in order to investigate further
the mechanisms of transition and the eect of varying degrees of free-stream
anisotropy.
2.1 Low-Reynolds Number Damping
As discussed in [17], the transition simulations provide some explanation for
the success of particular low-Re k- models in that they show the damping
eect of the wall on the eddy viscosity hardly changes as the boundary layer
develops from its pre-transitional pseudo-laminar state to its post-transitional
fully turbulent form. Of the most widely used low-Re k- model approaches,
the Dawes model formulation for f
and S
t
= f
k
2
= 0.25
(with and without the LaunderSharma f
3
is then automatically reduced from its turbulent value of 0.5
to the transition optimised value of 0.25 in the middle of the transition region
before recovering again to its fully turbulent level. Such a variation in the
LaunderSharma f
and T3A
test cases.
After careful consideration of the original SLY intermittency model devel-
opment and its subsequent renement and optimisation for by-pass transition
ows (as well as an open reassessment of the underlying physical modelling)
by Barton and Savill (unpublished), a revised, more robust, intermittency
transport model equation has now been adopted as follows:
Transport Equation:
D
Dt
= d
+P
1
+P
+G (4.1)
Diusion: d =
x
j
__
(1 )
3/2
c
_
k
_
uv +
_
x
j
_
(4.2)
Production (1): P
1
= c
1
(1 )
1
2
P
kk
k
(4.3)
Production (2): P
2
= c
2
_
k
2
__
x
j
_
2
(4.4)
Dissipation:
= c
3
(1 )
_
k
_
(4.5)
510 Savill
Entrainment G = c
4
(1 )
_
k
_
uv
k
1/2
x
j
(4.6)
Reynolds Stress Source: S
u
i
u
j
= c
5
(1 )
_
k
_
u
i
u
j
(4.7)
Dissipation Source Term: S
= c
4
uv
_
k
1/2
_
x
j
. (4.8)
This set has rst been tested in a 2D boundary layer code on the 1% and
3% test cases. For each case the code was run with or without the intermit-
tency equation coupled to the mean ow and Reynolds stress transport equa-
tions. Normal manual optimisation of the model constants conrmed that good
agreement could be achieved with the experimentally measured intermittency
and skin friction results see Figure 11 in uncoupled mode using a single
set of constants:
= 1.0, c
= 0.15, c
1
= 2.2, c
2
= 0.6,
c
3
= 0.1, c
4
= 0.16, c
5
= 0, c
4
= 0.1,
in line with the previous ndings of Savill (1996). Furthermore, for fully cou-
pled solutions, even better overall agreement could be achieved with the same
set, provided that then c
5
= 0.1 and c
1
becomes a linear function of the
free-stream turbulence intensity: c
1
= 1.0 for T3A
= 1.0, c
= 7 or 11.5, c
1
= 0.13, c
2
= 1.35 or 6.0,
c
3
= 0.1, c
4
= 0.05, c
5
= 0, c
4
= 0.1.
Using the manually optimised set of constants it was found that transition
predictions could successfully be extended down to a free-stream turbulence
level of only 0.3% (below the limit for other turbulence models). However,
switching to the second, automatically optimised set, allowed transition to be
predicted at external turbulence levels an order of magnitude less (0.04%),
[18] New strategies in modelling by-pass transition 511
Figure 11: Comparison of measured maximum intermittency growth rate ()
and C
f
(+) with SLY-( model predictions () for 3%fst test case T3A.
closer to that achieved in the best laminar ow research tunnels (0.008%)
and on external aircraft surfaces under ight conditions.
The same intermittency-transport model equation has also now been cou-
pled to the Dawes low-Re k- model employed within his unstructured Navier
Stokes code. Results presented for the Durham T3K Turbine test case by
Biesinger and Savill (1997) reproduce the experimentally observed transition
behaviour at midspan on the suction surface and both relaminarisation and re-
transition on the pressure surface. Subsequently a similar methodology has also
been successfully applied to a multi-element aerofoil conguration by Vicedo,
Vilmin, Savill and Dawes (2000).
512 Savill
Figure 12: e
n
predicted n factors for the T3A (1%fst) test case.
5 Other possible approaches
The large eddy simulations performed by Voke and Yang (1993) have revealed
evidence of wave-like behaviour, as well as free-stream-turbulence-induced
streaks and isolated spots in the wall region ahead of transition onset (de-
tected as the point of minimum C
f
). The division between natural and by-pass
transition regimes is certainly not as clear as Morkovin and others have sug-
gested, and the simulations certainly support ample evidence in the literature
showing that theT-S mechanism can co-exist with other strongly amplifying
mechanisms over a wide range of fst intensities from well below to well above
1% see also [17].
Interest is thus now developing in possible extensions of e
n
and parabolised
stability equation (PSE) approaches to variable fst by-pass transition cases,
because there is already some evidence from the NASA Transition Project and
elsewhere that both methods can be used successfully to predict the onset of
transition for non-zero free-stream turbulence on both at and curved surfaces,
although the n factor then becomes an empirical function of the free-stream
turbulence intensity. In particular Pereira and colleagues see Sousa et al.
[18] New strategies in modelling by-pass transition 513
Figure 13: PSE predictions (-) of C
f
for the T3A test case versus experimental
data ().
(2000) have demonstrated that an e
n
method can predict transition onset
for the ERCOFTAC Transition Modelling SIG T3A
+
, are better at predicting transition than those which satisfy either
only one or neither of these.
Low-Re models which employ damping factors that are functions of a
general property of low-Re ows (in particular turbulent Reynolds num-
ber R
t
) are more appropriate for the prediction of low-Re transition
regions than ones that only introduce a dependence on wall-proximity
(e.g. via R
y
, R
).
Many of the more recent low-Re turbulence models (which have been
carefully calibrated against turbulence simulations to meet additional
wall-limiting conditions and thus produce superior predictions for a wide
range of fully turbulent ows) contain damping factors that are functions
of y and thus fail to predict transitional ows correctly.
Such models can be improved for transition prediction by introducing
extra x-dependent damping factors and/or replacing R
y
, R
by R
t
; elim-
inating the D and/or E factors; switching to v for the velocity scale; and
introducing a nonlinear stress-strain relationship, a specic allowance for
pressure-diusion, and better sensitivity to straining (history).
Of all the low-Re model treatments examined thus far the use of Launder
Sharma-type damping factors (which are functions of R
t
and satisfy the
wall-limiting conditions for uv
+
and
+
is recommended for most accu-
rate prediction of mean ow quantities.
However this model is demanding of grid resolution and requires specic
sensitising to irrotational straining. It is better implemented in a low-Re
RST model which can also capture some eects of free-stream turbu-
lence stress anisotropy and does not require excessive grid renement or
specic strain-dependent corrections.
Unfortunately all models are sensitive to initial and free-stream bound-
ary conditions for , which are usually not well-dened (estimates based
516 Savill
on turbulent values for the mixing length or structure parameter distri-
butions, and integral length scale, respectively are subject to consider-
able uncertainty). Consequently the eects of variable and anisotropic
free-stream length scales encountered in many practical ows cannot yet
be predicted.
All transport models require greater mesh resolution than can be accom-
modated in current 3D computations for complex engineering ows on
all but the very latest massively parallel computers.
Future research must focus on developing methods less demanding of CPU/
store, but oering greater predictive abilities and applicability. The combina-
tion of PSE with intermittency transport modelling may oer a novel route
towards this goal, and, in view of the simulation ndings, a good candidate
transport model may be that of Durbin which adopts v as velocity scale and
solves an elliptic relaxation equation to take specic account of the non-local
damping eect of walls. A new two-time-scale model of Yang and Shih, which
uses a combination of R
t
and a dimensionless strain rate parameter also merits
investigation, particularly as this model has been shown to provide the best k-
model predictions for ow through rows of tube-bundle rods see Watterson
et al. (1999). Other studies suggest further consideration of models combin-
ing an R
y
dependence for (low-Re) near wall regions and an R
t
or perhaps
R
L
-dependence for (low-Re) transition regions see [17].
In principle even better predictions should be feasible with the hybrid low-
Re/TCL approach, since this allows a clearer distinction to be drawn between
the modelling of low-Re transitional and (low-Re) near-wall eects. However
this approach clearly requires further development along the lines currently
being revised by Hanjalic (1997, 1999), Hadzic (1997, 1999) and others.
Continued work on a hierarchy of models, evaluated step-by-step against
a larger series of progressively more complex test cases, is therefore recom-
mended, but it is clear that input is also required from studies of much more
complex engineering ows. These last are needed both to assist the improve-
ment of current physical modelling and ensure that the current best models
really will be able to improve predictions when many more complicating fac-
tors, including compressibility, rotation and unsteadiness, have to be taken
into account.
The ow through turbomachinery and many other practical ow geome-
tries, such as tube bundles and heat exchangers, is of course subjected to the
sequential eects of convex and concave curvature as well as the combined
eect of streamline curvature and irrotational straining. Rather less attention
has so far been paid in general to the modelling of such sequences and combi-
nations of simple strain rates, although it has been known for some time that
these may produce unexpected nonlinear eects. Fortunately Savill and Tse-
lepidakis (1994) have shown that the basic LRR RST model can qualitatively
[18] New strategies in modelling by-pass transition 517
reproduce the eects of sequences of convex and concave curvature including
their unequal eect on the shear stress (which is suppressed more by the stabil-
ising eect of convex curvature than it is increased by the destabilising eect
of concave curvature), but similar magnitude inuence on the normal stress
anisotropy. The same authors have also shown that the same model can also
capture the manner in which the response to and recovery from each strain
is modied by the opposing action of the other sense of curvature, at least
when diusion is negligible (as in wakes). However, in practice it is clear that
signicant further lag eects occur in such ows.
Alternative TCL modelling can provide better quantitative predictions than
the LRR model for the variation of Reynolds stress anisotropy in homogeneous
shear ow subjected to both weak and strong convex or concave curvature.
However the production term in the equation needs to be scaled by the non-
equilibrium parameter P
k
/ to produce equally good results when applied to a
shear ow subjected to sequential convex-to-concave curvature. It is likely that
similar non-equilibrium extensions will be required to most models to account
correctly for lag and other strain-history eects encountered when applying
them to real engineering ows.
Further attention also needs to be directed to the handling of nite leading
(and trailing) edges and associated problems posed by transition in leading-
edge separation bubbles as well as unsteady interactions, including wake-
passing transition with intermediate boundary-layer calming eects. The T3L
semi-circular and leading-edge test case has now become a new entry-level for
model validation see Savill and Pittaluga (2000) and it has been clearly es-
tablished that stagnation-region corrections (such as those proposed by Kato
and Launder 1993, and Durbin 1983) are essential to maintain good predictions
with even the Dawes and LaunderSharma low-Re k- models. Somewhat un-
expectedly, only poor results (too large a bubble and delayed transition) have
so far been obtained with nonlinear k- schemes perhaps reecting insuf-
cient generality of model calibration. Very much better results have been
reported by Papanicolau and Rodi (1999) using a two-layer approach with al-
ternative Chen and Thyson-type correlations more representative of separated
shear layer transition, and by Hadzic and Hanjalic (2000) using their TCL
RST model approach. Moreover, Vicedo et al. (2000) have recently demon-
strated that the addition of a simple prescribed-intermittency scaling (and
subsequently intermittency transport) to the unmodied Dawes low-Re k-
model can provide excellent predictions both up to and downstream of the
leading edge. They have also now begun applying the same approach(es) to
the prediction of unsteady transition in an LA turbine blade test case with
equal success.
518 Savill
Acknowledgements
This contribution would not have been possible without the continued sup-
port of Rolls-Royce plc, DRA, the CEC and COST. Several developments owe
their origins to discussions with close colleagues at UMIST (especially Pro-
fessor Brian Launder), researchers at the University of Surrey (in particular
Professor Peter Voke) and Dr. John Coupland of RollsRoyce Aerothermal
Methods; with whom regular interaction has been maintained as part of an
overall Transition Strategy. The assistance of Dr. B.A. Younis of City Uni-
versity, who freely provided the original version of the low-Re RST code and
assisted with the implementation of this, has been vital to the development
of the SLY model, as have contacts with other researchers, too many to men-
tion individually, whose help is also gratefully acknowledged. The Editors,
organisers of the INI Programme and other colleagues contributed a number
of constructive criticisms which I hope are reected in this nal version.
References
Abid, R. (1993). Evaluation of two-equation turbulence models for predicting tran-
sitional ows, Int. J. Engineering Science 11 25.
Biesinger, T.E. and Savill, A.M. (1998). A contribution to the Durham Linear Turbine
Cascade Test Case. Proc. ERCOFTAC Turbomachinery SIG Workshop, Aussois.
Byggstoyl, S. and Kollmann, W. (1986). A closure model for conditioned stress equa-
tions and its application to turbulent shear ows, Phys. Fluids 29 1430.
Byggstoyl, S. and Kollmann, W. (1981). Closure model for intermittent turbulent
ows, Int. J. Heat Mass Transfer 24 1811.
Cho, J.R. and Chung, M.K. (1992). A proposal of -- turbulence model, J. Fluid
Mech. 237 301.
Cho, J.R. and Chung, M.K. (1990). Intermittency modeling based on interaction
between intermittency and mean velocity gradients. In Engineering Turbulence
Modelling and Experiments, W. Rodi and E. Ganic (eds.), Elsevier, 101.
Cotton, M.A. (1989). Personal Communication.
Dick, E. (Ed) (1995). Theme section on transition, ERCOFTAC Bulletin 24.
Durbin, P.A. (1993). Application of a near-wall turbulence model to boundary layers
and heat transfer, Int. J. Heat Fluid Flow 14 316.
Hadzic, I. and Hanjalic, K. (2000). Separation-induced transition to turbulence:
second-moment closure modelling, J. Flow Turb. and Comb., 63, 153273.
Hallback, M., Henningson, D. S., Johansson, A. J. and Alfredsson, H. (eds.) (1996).
Turbulence and Transition Modelling, Kluwer Academic Publishers.
Hanjalic, K., Jakirlic, S. and Hadzic, I. (1997). Expanding the limits of equilibrium
second-moment closures, Fluid Dynamics Research 20 2541.
Hui, J. (2000). Turbulence and Transition Predictions of Internal Flows. PhD Thesis,
KTH, Stockholm
[18] New strategies in modelling by-pass transition 519
Janicka, J. and Kollmann, W. (1983). Reynolds stress closure model for conditional
variables. Proc. 4th Symp. Turbulent Shear Flows, 14.1314.18 Karlsruhe, 14.13
Janour, P. (1997). A new k- model for predicting transition. Institute of Thermo-
mechanics Report, Prague.
Johnson, M.W. and Ercan, A.H. (1999). A physical model for bypass transition, Int.
J. Heat Fluid Flow 20, 95104.
Kato, M. and Launder, B.E. (1993). The modelling of turbulent ows around sta-
tionary and vibrating square cylinders. Proc. 9th Symp. Turbulent Shear Flows,
10-4.110-4.6 Kyoto.
Kawamura, H. and Hada, K. (1992). A k- two-equation model for fully developed
channel ow. Submission to Stanford collaborative testing turbulence models, Sci-
ence Univ. Tokyo.
Kebede, W., Launder, B. E. and Younis, B.A. (1985). Large amplitude periodic pipe
ow: A second-moment closure. Proc. 5th Symp. Turbulent Shear Flows, 16.23
16.28.
Lumley, J.L. (1978). Computational modeling of turbulent ow, Adv. Appl. Mech.
18 123176.
Mansour, N.M, Kim, J. and Moin, P. (1988). Reynolds-stress and dissipation-rate
budgets in a turbulent channel ow, J. Fluid Mech. 194 1544
Papanicolau, E.L. and Rodi, W. (1999). Computation of separated-ow transition
using a two-layer model of turbulence, J. Turbomachinery 121 78.
Patel, V.C., Rodi, W. and Scheuerer, G. (1985). Turbulence models for near-wall and
low-Reynolds number ows: a review, AIAA J. 23 1308.
Pereira, J.C.F. (2000). Coupling NavierStokes and stability calculations in the in-
compressible regime. Proc. ECCOMAS Conference, Barcelona.
Pironneau O., Rodi W., I.L. Ryhming, Savill A.M. and Truong T.V. (eds.) (1992).
Proc. 1st ERCOFTAC Workshop on Numerical Simulation of Unsteady Flows
Transition to Turbulence and Combustion, Lausanne, Cambridge University Press;
see in particular A.M. Savill: A synthesis of T3 test case predictions.
Rodi, W. and Mansour, N.N. (1993). Low-Reynolds number k- modelling with the
aid of direct simulation data, J. Fluid Mech. 250 509529.
Rotta, J.C. (1968). Statistical theory of non-homogeneous turbulence. Papers 1 and
2. (translated by W. Rodi) Imperial College Reports TWF/TN/3839.
Savill, A.M. (1993a). Some recent progress in the turbulence modelling of by-pass
transition. In Near-wall Turbulent Flows, R.M.C. So, C. Speziale and B.E. Launder
(eds.), Elsevier, 829848.
Savill, A.M. (1993b). Further progress in the turbulence modelling of by-pass tran-
sition. In Engineering Turbulence Modelling and Experiments 2, W. Rodi and F
Martelli (eds.), Elsevier, 583592.
Savill, A.M. (1994). Transition modelling for turbomachinery II. Summary Proc.
of the 1st ERCOFTAC Transition SIG Workshop of the BRITE-EURAM AERO-
CT92-0052 Project Workshop on Transition in Turbomachinery, VUB, Brussels.
520 Savill
Savill, A.M. (1995) A summary report on the COST ERCOFTAC Transition SIG
Project: Evaluating turbulence models for predicting transition, ERCOFTAC Bul-
letin 24, 5761.
Savill, A.M. (1996). Evaluation of turbulence models for predicting transition in
turbomachinery ows. Minowbrook II Workshop on Boundary Layer Transition in
Turbomachines, Sycracuse University.
Savill, A.M. (1999a). Transition Prediction with turbulence models. Proc. INI Work-
shop on Breakdown to Turbulence and its Control, Isaac Newton Institute, Cam-
bridge.
Savill, A.M. (1999b). Non-local counter-gradient transport modelling. Proc. INI
Workshop on Mathematics of Closure, Isaac Newton Institute, Cambridge.
Savill, A.M. and Pittaluga, F. (2000). Implementation and further application of
rened transition prediction methods for turbomachinery and other aerodynamic
ows, ERCOFTAC Bulletin 44, 45.
Savill, A.M and Tselepidakis, D.P. (1994). Adopting the new TCL Reynolds stress
model for curved ows. Proc. 6th Biennial Colloquium on Computational Fluid
Dynamics, UMIST.
Sieger, K. (1999). Prediction of Transitional Boundary Layers with a New Two-layer
Turbulence Model. PhD Thesis, University of Karlsruhe
Simon, F.F. (1994). Proc. AFOSR Workshop on End-Stage Boundary Layer Transi-
tion, Minowbrook, Syracuse University.
Sousa, J.M.M., Margues, N.P.C. and Pereira, J.C.F. (2000). Coupling NavierStokes
calculations with linear stability analysis in wing design. Proc. ECCOMAS
2000 Special Session on Transition Prediction using Tubulence Models and PSE,
Barcelona.
Steelant, J. and Dick, E. (1995). Transition modelling by conditionally averaged
equations, ERCOFTAC Bulletin 24, 6264.
Vicedo, J. Dawes, W.N. and Savill, A.M. (2000). Coupling PSE to a low Re k- -
intermittency transport RANS approach. Proc. ECCOMAS 2000 Special session
on Transition Prediction using Turbulence Models and PSE, Barcelona.
Vicedo, J., Vilmin, S., Dawes, W.N. and Savill, A.M. (2000). Extension of intermit-
tency transport modelling to natural transition in external aerodynamic applica-
tions. Proc. RAeS Aerodynamics Research Conference 2000, Royal Aeronautical
Society, London.
Vicedo, J., Vilmin, S., Dawes, W.N., Hodson, H.P. and Savill, A.M. (2000). The
extension of CFD-friendly turbulence modelling to include transition. Proc. 8th
European Turbulence Conference, Barcelona.
Voke, P.R. and Yang, Z.Y. (1993). Numerical solutions of the mechanisms of by-pass
transition in the at plate boundary layer. Proc. 9th Symp Turbulent Shear Flows,
Kyoto, 21-2-121.2-6.
Voke, P.R., Yang, Z.Y. and Savill, A.M. (1996). LES and modelling of transition
following a leading-edge separation bubble, In Engineering Turbulence Modelling
and Experiments-3, W. Rodi and G. Bergeles (eds.) 601610, Elsevier Science.
[18] New strategies in modelling by-pass transition 521
Westin, K.J.A. and Henkes, R.A.W.M. (1997). Application of turbulence models to
bypass transition, J. Fluids Eng. 119 859866.
Yang, Z. and Shih, T. H. (1991). Extending turbulence modelling to by-pass transi-
tion, NASA CMOTT Research Briefs 83.
Yao, Y.F., Savill, A.M., Sandham, N.D. and Dawes, W.N. (2000). Simulation of a
turbulent trailing edge ow using unsteady RANS and DNS. Proc. International
Heat and Mass Transfer Conference, Nagoya.
19
Compressible, High Speed Flows
S. Barre, J.-P. Bonnet, T.B. Gatski and
N.D. Sandham
1 Introduction
There is an ever-increasing need to be able to predict and control high subsonic
to high supersonic speed ows for the optimal design of aerospace vehicles. In
this Mach number range, large variations in pressure can occur which can also
lead to large density variations. Such variations in the state variables can have
a signicant impact on both the mean and turbulence dynamics.
Improving the prediction and control of compressible ows requires an accu-
rate description of the large scale (structure) dynamics and of the mean and
other statistical properties of the ow. Thus, a better understanding of the
dynamics and improvements in the modeling of these ow features is essen-
tial. Within the framework of the ensemble-averaged NavierStokes equations,
compressible ows have generally been computed using mass-weighted (Favre-
averaged) variables. Higher-order correlations with incompressible counter-
parts have been modeled by variable density extensions to the incompressible
form; meanwhile higher-order correlations unique to the compressible form
have been isolated. Some of these compressibility terms, such as dilatation
dissipation and the compressible heat ux, have been widely used. Others,
such as mass ux or pressure-dilatation, have seen more limited application,
as have explicit compressibility corrections to standard model terms. Many
calculations have been performed neglecting (with or without justication)
extra compressible terms and using simple variable density extensions of the
equations.
The goal of this chapter is to provide the reader with an overall perspective
of the experimental and numerical study of compressible, turbulent shear lay-
ers including the eect of shock-turbulence interactions. Concepts and relevant
modeling issues associated with turbulent boundary layers, mixing layers and
wakes are discussed. In addition, the eects of shock/turbulence interactions
on ow eld dynamics in both isotropic, homogeneous turbulence, and inho-
mogeneous (boundary-layer and mixing layer) ows are discussed. Coupled
with some fundamental theoretical aspects of compressible turbulent ows,
this should provide a summary of the current state of the art and a basis for
further research.
522
[19] Compressible, high speed ows 523
2 Background: Experimental and Simulation Data
Detailed and accurate data for mean and turbulent quantities in supersonic
turbulent ows are very important in order to increase our knowledge about
compressible turbulent ow dynamics and develop turbulence closures. Such
data can conceivably be provided by both laboratory experiments and numer-
ical simulations (DNS, and to a more limited extent LES). However for both
experiments and simulation there are additional diculties beyond those faced
for low speed ows. In this section we briey review the current status.
The rst issue with experiments is to obtain suciently high ow quality.
Great care must be taken concerning the initial turbulence levels. Parasitic
acoustic perturbations can alter the external conditions of the ow and make
the results dicult to generalize and not very easy to use as test cases for
turbulence model validation. One example of this is in a supersonic mixing
layer where it is dicult to obtain, in a given wind tunnel, a wide range of
convective Mach numbers M
c
. Even in a classical air-air wind tunnel with
good quality external conditions such as low external pressure gradients, low
external noise and turbulence, and a suciently large test section to avoid
connement eects, a variation in M
c
from 0.4 to 1 can be considered large,
but even this is not easy to obtain in a lot of experimental facilities. Another
example is the case of a normal shock wave interacting with homogeneous and
isotropic turbulence. In this case, it is very dicult to generate an incoming
turbulent eld free from parasitic eects like pressure waves and low frequency
pulsation.
Another major diculty with experiments is concerned with measurement
techniques. The most popular techniques used to measure turbulent elds
in supersonic ows are hot-wire anemometry and Laser Doppler Velocimetry
(LDV). While these techniques allow one to obtain a lot of turbulence data,
they are not free of diculty in these types of ows. For the hot wire, the main
problem lies with calibration, particularly when transonic Mach numbers and
low Reynolds numbers are involved. It has been shown that neglecting tran-
sonic eects may lead to about a 50% underestimation in root mean square
velocity uctuations (Barre et al. 1992). LDV does not need calibration, but
the ow must be seeded with particles in order to generate Mie scattering light.
In highly sheared supersonic compressible ows, such as mixing layers at high
M
c
, seeding problems may corrupt the experimental results. Another problem
with LDV seeding arises in the case of shock/turbulence interaction. The in-
ertia associated with the seeding particles makes them unable to follow the
local air velocity in the immediate downstream vicinity of the shock. In these
cases, a particle drag correction must be applied to the data. This law must
be computed accurately in order to keep experimental results representative
of reality. Even more measuring diculties arise when 3D measurements are
necessary. Three-dimensional LDV in supersonic ows is not an easy task, as
discussed by Bonnet et al. (1996) and Gruber et al. (1993), so that it is some-
524 Barre et al.
times necessary to develop mixed measurements techniques including hot wire
and LDV to deduce three-dimensional Reynolds stress behavior (Chambres
1997).
Some new optical techniques, such as Filtered Rayleigh Scattering, Par-
ticle Image Velocimetry (PIV) and Collective Light Scattering have begun
to be used (see Bonnet et al. (1998) for a complete description). These new
techniques are very promising, but at this time no complete results for the
Reynolds stress tensor in supersonic compressible ows have been obtained.
Hopefully, in the near future, they will be able to provide new information for
variables like the thermodynamic turbulent uctuations which would be very
useful in the eld of compressible turbulence modeling.
Direct numerical simulations (DNS) of compressible ows have mainly fo-
cused on simple homogeneous ows, with and without shocks, to study the
eects of compressibility on turbulence. Studies of inhomogeneous turbulent
shear ows, such as boundary-layer and mixing-layer ows, are less common
but have been performed. Such ows suer from the same constraints as their
incompressible counterparts in that grid resolution restrictions prevent high-
Reynolds number computations. In the compressible, turbulent boundary-
layer case, such constraints are somewhat mitigated by the fact that the vis-
cosity increases near the wall, thereby decreasing the local eective Reynolds
number in a region where grid resolution is most critical.
Large-eddy simulation for high-speed ow is in its infancy. Extra sub-grid
terms requiring modeling appear in the energy equation. Thus far attempts to
model these terms have been based on extensions of standard techniques such
as dynamic modeling from incompressible ow, with some limited guidance
from DNS. The interaction of sub-grid model and numerical method can be
even more important for compressible ow with shock waves, which require
special schemes, often relying on extra dissipation. Since this is introduced
near the grid scale it will have a strong eect on the energy transfer from
resolved to unresolved scales. This may well have implications of the choice of
sub-grid model.
3 Compressible NavierStokes Equations
It is a useful prelude to both the experimental and numerical results to be
presented in this chapter to detail the equations governing the motion in the
compressible regime. In addition to the numerical simulations or the calcu-
lation of the mean statistics through a Reynolds-averaged approach, these
equations provide a useful basis of analysis for the experimental results. It
will be shown how an analysis of the energy-budgets within the context of the
ensemble-averaged equations leads to a better understanding of the dynamics
of compressible turbulent ow.
[19] Compressible, high speed ows 525
The starting point is the mass, momentum, and (total) energy conservation
equations. The conservation of mass equation is
t
+
(u
j
)
x
j
= 0, (2.1)
and the conservation of momentum equation is
(u
i
)
t
+
(u
i
u
j
)
x
j
=
p
x
i
+
ij
x
j
, (2.2)
with the viscous stress tensor
ij
dened as
ij
= 2
_
S
ij
1
3
S
kk
ij
_
, (2.3)
where and u
i
are respectively the density and velocity components, is the
molecular viscosity, and S
ij
is the strain rate
S
ij
=
1
2
_
u
i
x
j
+
u
j
x
i
_
. (2.4)
The total energy equation is given by
(E)
t
+
(u
j
H)
x
j
=
(u
i
ij
)
x
j
q
j
x
j
, (2.5)
where the total energy and total enthalpy, E and H, respectively, are
E =
_
e +
u
i
u
i
2
_
(2.6)
H =
_
E +
p
_
(2.7)
and e = c
v
T is the internal energy (c
v
is the specic heat at constant volume
and is assumed constant), q
j
= k
T
T/x
j
is the heat ux (k
T
is the thermal
conductivity), and the equation of state is the perfect gas law p = RT.
These equations, or reformulations thereof, can be used in a direct simula-
tion approach. In a large-eddy simulation approach, the governing equations
are ltered to yield a set of equations for the large-scale motions of the ow,
and in a RANS approach the equations are Reynolds averaged to yield a set
of equations for the statistical mean motion of the ow.
2.1 Reynolds-averaged equations
While direct or large-eddy simulations of compressible turbulent ows have
previously focused on homogeneous and temporally developing ows, with only
526 Barre et al.
recent applications to inhomogeneous ows, numerical solutions of the com-
pressible, Reynolds averaged NavierStokes (RANS) equations with suitable
correlation closure models, have been available for more than three decades.
In the compressible formulation, a Reynolds averaging approach is once
again applied to the conservation equations (2.1), (2.2), and (2.5). It has been
found that a rewriting of the equations using mass-weighted, or Favre (Favre
1965), variables is advantageous, since the equations take a more compact form
relative to the Reynolds averaged variables and the terms in the equations can
be shown to have analogous counterparts in the incompressible formulation.
For a dependent variable f, the Favre average is dened as
f =
f
. (2.8)
The instantaneous value f can then be decomposed into either the usual
Reynolds averaged variables or the Favre-averaged variables
f = f +f
=
f +f
. (2.9)
As might be expected, an extensive list of relations exist between the Reynolds-
averaged variables and the Favre-averaged variables. Some of these relations
are provided in Appendix A.
The equations for the mean density and mean momentum u
i
are given
by
t
+
x
j
( u
j
) = 0, (2.10)
D u
i
Dt
=
( u
i
)
t
+
x
j
( u
j
u
i
) =
p
x
i
+
ij
x
j
(
ij
)
x
j
, (2.11)
and the mean viscous stress tensor
ij
is
ij
= 2
_
S
ij
1
3
S
kk
ij
_
2
_
S
ij
1
3
S
kk
ij
_
, (2.12)
where is the mean molecular viscosity, and
ij
=
u
i
u
j
is the Favre-averaged
correlation tensor. Equation (2.12) neglects contributions from
, and assumes
that U
i
u
i
. This assumed equality between the average velocities implies that
the average uctuating velocity u
i
is small since u
i
u
i
= u
i
. Results from
a DNS of a spatially evolving supersonic ow (M
E is
(
E)
t
+
x
j
_
u
j
H
_
=
x
j
_
q
j
+E
j
_
, (2.13)
[19] Compressible, high speed ows 527
1 10 100 1000
y
+
0.0
0.2
0.4
0.6
0.8
1.0
u
,
u
_
~
u
_
u
~
1 10 100 1000
y
+
0.030
0.025
0.020
0.015
0.010
0.005
0.000
u
u
_
~
(a) (b)
Figure 1: Mean streamwise velocity prole across supersonic boundary layer:
(a) Favre and Reynolds mean, (b) mean velocity dierence. (Velocities scaled
with freestream velocity.)
where
E = c
v
T +
u
i
u
i
2
+
i
u
i
2
, (2.14)
H =
E +
p
, (2.15)
q
j
= k
T
T
x
j
k
T
T
x
j
, (2.16)
and
E
j
= c
p
j
T
+ u
i
(
ij
ij
) +
u
i
u
i
u
j
2
ij
u
ij
u
i
. (2.17)
In the approximation of (2.16), uctuations in the thermal conductivity are
neglected, and the Favre-averaged and Reynolds-averaged mean temperatures
are taken as approximately equal. The equality of the two averages may once
again be checked against DNS results. As Figure 2 shows from the results of Rai
et al. (1995), the maximum variation within the boundary layer between the
two temperature averages is approximately 1 percent occurring at y
+
20. In
the cold-walled case of Huang et al. (1995) a similar deviation level was found.
The equation of state in mean variables is
p = R
T, (2.18)
or in terms of the mean total energy
E
p = ( 1)
_
E
1
2
_
u
2
+ v
2
+ w
2
_
k
_
(2.19)
528 Barre et al.
1 10 100 1000
y
+
0.5
1.0
1.5
2.0
T
,
T
_
~
T
_
T
~
1 10 100 1000
y
+
0.000
0.005
0.010
0.015
0.020
T
T
_
~
(a) (b)
Figure 2: Mean temperature prole across supersonic boundary layer: (a) Favre
and Reynolds mean, (b) mean temperature dierence. (Temperatures scaled
with freestream temperature.)
where is the ratio of specic heats (c
p
/c
v
), and R is the gas constant. The
presence of the turbulent kinetic energy term k,
k =
ii
2
=
i
u
i
2
, (2.20)
suggests a strong coupling between the mean equations and the turbulent
transport equations. Use of the total energy as a primary variable facilitates
the use of shock-capturing numerical methods.
In the incompressible case, the only corresponding term requiring closure
would be the Reynolds stresses
ij
; however, the inclusion of the mean total
energy equation necessitates models for the turbulent heat ux c
p
u
j
T
, the
turbulent mass ux u
i
=
i
u
i
u
j
. While some or all of these terms may not be signicant in a majority
of compressible ows, it is important to include them at the outset in order
to recognize the assumptions invoked in obtaining more simple forms of the
equations.
2.2 Morkovins Hypothesis and the Strong Reynolds Analogy
Before proceeding onto the development and analysis of models for the un-
known correlations that arose in the development of the mean conservation
equations, it is useful to examine one of the most important concepts associ-
ated with the study of compressible ows. Morkovins hypothesis (Morkovin
1964) and the associated set of relations which constitute what is termed the
strong Reynolds analogy (SRA) have proved useful in the analysis of compress-
ible, supersonic ows. (The reader is referred to the review by Brashaw (1977)
and the study by Gaviglio (1987) as well as the book of Smits and Dussauge
[19] Compressible, high speed ows 529
(1996) for further discussion on this topic.) These ideas were developed from
an analysis of 2D supersonic boundary layer data valid for M < 5 boundary
layer ows at moderate values of wall heat transfer q
w
, but have been gener-
ally applied as a basic concept. In the discussion, the analysis will focus on a
simple boundary layer ow with velocity components (u, v) in the streamwise
x and wall normal y directions, respectively.
From an analysis of the data available at the time, Morkovin deduced that
the . . . essential dynamics of these supersonic shear ows will follow the in-
compressible form . . . This observation was quantied by the proposal that
_
w
_
u
u
2
and
_
w
_
u
2
u
2
(2.21)
would depend little on Mach number . . . This hypothesis was deduced from
Morkovins analysis of the solutions of the momentum and total enthalpy
equations. These solutions, which will now be derived, are what Morkovin
termed the strong Reynolds analogy (SRA).
The derivations to follow are based on a linearization of the governing equa-
tions which is justied by the early success of Kovasznays categorization (Ko-
vasznay 1953) of compressible turbulence into rotational, entropy, and acous-
tic modes. For moderate Mach number supersonic ows, the pressure uc-
tuations (acoustic mode) are neglected. From the perfect gas law (p = RT
= ( 1)c
v
T) and for small uctuations, the pressure uctuations are related
to both the density and temperature uctuations through
p
p
=
+
T
T
, (2.22)
so that if the pressure (acoustic) uctuations are negligible, the density and
temperature uctuations are related through
=
T
T
. (2.23)
Since this is dierent to the isentropic relation, the temperature uctuations
must be considered to be non-isentropic.
In the case of an adiabatic wall, the SRA as proposed by Morkovin (1964)
(see also Young 1953) is given by
T
t
=
T
tw
(2.24)
T
t
= 0, (2.25)
where the subscript t denotes total conditions (
T
t
=
T+ u
2
/(2c
p
)). The absence
of total temperature uctuations corresponds to neglecting uctuations due to
the entropy mode. When these equations are coupled with the total enthalpy
530 Barre et al.
equation and terms which are quadratic in the uctuations are neglected, the
temperature and velocity uctuations are related by
c
p
T
uu
. (2.26)
(Note that it has additionally been assumed here that uu
vv
.) This
instantaneous relation between the temperature and velocity uctuations can
be recast into more useful statistical relations. For example, by using the mean
total enthalpy relation (neglecting quadratic uctuations)
c
p
T
t
= c
p
T
tw
= c
p
T +
u
2
2
(2.27)
evaluated at the boundary layer edge, (2.26) becomes
_
T
2
T
tw
T
e
2
u
u
e
_
u
2
u
e
, (2.28)
or, by using the local Mach number (M
2
= u
2
/R
T),
_
T
2
T
( 1) M
2
_
u
2
u
. (2.29)
In terms of the correlation coecient R
u
T
, dened as
R
u
T
=
u
_
u
2
_
T
2
, (2.30)
(2.29) yields
R
u
T
1 (2.31)
As was discussed at the beginning of this subsection, the pressure uctuations
have been neglected in the analysis so that the density and temperature uc-
tuations are related by (2.23). Thus, both (2.29) and (2.31) can be rewritten
in terms of these density uctuations as
_
( 1) M
2
_
u
2
u
(2.32)
and
R
u
1, (2.33)
respectively. One remaining expression associated with the SRA can be found
from the turbulent Prandtl number dened by
Pr
t
=
t
c
p
k
T
. (2.34)
[19] Compressible, high speed ows 531
A relation between the vertical turbulent heat ux and the turbulent shear
stress can be obtained directly from (2.26), so that
u =
c
p
v
, (2.35)
and a relation between the mean temperature and velocity gradients can be
obtained from (2.27), so that
u =
c
p
T/y
u/y
. (2.36)
Equations (2.35) and (2.36) can be combined to yield the ratio
_
u
u/y
__
T/y
v
_
1 (2.37)
or
Pr
t
=
t
c
p
k
T
1. (2.38)
Equations (2.26), (2.28), (2.29), (2.31), and (2.38) are the ve relations ob-
tained by Morkovin (1964) that constitute the strong Reynolds analogy. Ad-
ditional relations have since been derived from this basic set. For example, a
very strong Reynolds analogy (VSRA) has been proposed which is the instan-
taneous counterpart to (2.29),
T
T
( 1) M
2
u
u
. (2.39)
In order to account for heat ux at the wall, an extended strong Reynolds
analogy has been proposed (ESRA) (Gaviglio 1987, Rubesin 1990)
T
T
( 1) M
2
u
/ u
1
Pr
t
1
_
1
T
t
/
T
_
(2.40)
where the proportionality to the turbulent Prandtl number Pr
t
has been point-
ed out by Huang et al. (1995). Gaviglio (1987) assumed an equality between the
characteristic length scales associated with both the velocity and temperature
small-scale uctuations to arrive at the form given in (2.40), and Rubesin
(1990) assumed a mixing length formulation to arrive at a relation for the
static enthalpy uctuations which eventually led to (2.40).
Both Morkovins hypothesis and the various forms of the Reynolds analogies
can be used to validate both experimental and numerical simulation data as
well as the predictive capability of RANS. In turn, experimental and numerical
simulation data can be used to validate the assumptions used in the derivation
of the relations comprising the Reynolds analogies.
532 Barre et al.
2.3 Turbulent transport equations
In the analysis of compressible, turbulent ow elds through either physical or
numerical experiments, the equations describing the transport of the turbulent
Reynolds stresses will be needed. As was seen in the development of the mean
ow equations, the turbulent stresses are needed to close the equations. The
turbulent stress equations lead directly to the turbulent kinetic energy equa-
tion which is especially useful in understanding the underlying ow dynamics
through an analysis of the component terms comprising the energy budget.
The transport equations for the turbulent stresses in the compressible case
are obtained in a manner analogous to the incompressible formulation although
now the dependent variables are decomposed into Favre-mean and uctuating
components and then Reynolds-averaging the equations (e.g. Gatski 1996).
This procedure leads to the Reynolds-averaged equations for the Favre corre-
lation tensor
ij
,
D
ij
Dt
=
ik
u
j
x
k
jk
u
i
x
k
+
ij
ij
+M
ij
+D
ij
, (2.41)
where
ij
= p
_
u
i
x
j
+
u
j
x
i
_
, (2.42)
ij
=
2
3
ij
+
D
ij
=
ik
u
j
x
k
+
jk
u
i
x
k
, (2.43)
M
ij
= u
i
_
jk
x
k
p
x
j
_
+u
j
_
ik
x
k
p
x
i
_
, (2.44)
D
ij
=
x
k
_
_
u
i
u
j
u
k
+p
(u
jk
+u
ik
)
. .
turbulent transport
(
ik
u
j
+
jk
u
i
)
. .
viscous diusion
_
_
. (2.45)
As the equation shows, the transport of the Reynolds stress is controlled by a
balance among its production, redistribution
ij
, destruction
ij
, mass ux
contribution M
ij
, transport and diusion D
ij
.
A simplication results when the turbulent kinetic energy equation is con-
sidered since it is simply the trace of the (2.41),
Dk
Dt
=
ik
u
i
x
k
+ +M +D, (2.46)
where
=
ii
2
= p
i
x
i
, (2.47)
=
ii
2
=
ik
u
i
x
k
, (2.48)
[19] Compressible, high speed ows 533
M =
M
ii
2
= u
i
_
ik
x
k
p
x
i
_
, (2.49)
D =
D
ii
2
=
x
k
_
_
u
i
u
i
u
k
2
+p
ik
ik
u
i
_
_
. (2.50)
The terms on the right hand side of (2.46) are the pressure-strain correlation
(which here appears as pressure dilatation), the compressible turbulent dissipa-
tion rate, the mass ux contribution, and the turbulent transport, respectively.
Theoretical models have been developed for all these terms, but with varying
degrees of validation. Some of the experimental and numerical calibration and
validation of models for these terms will be discussed in subsequent sections.
As in the incompressible case, the deviatoric part of the tensor dissipa-
tion rate, (2.43) is usually assimilated into the pressure-strain rate correla-
tion, which leaves the isotropic dissipation rate in the compressible Reynolds
stress or turbulent kinetic energy formulations to be obtained from a modeled
transport equation. In the compressible case, a partitioning of the isotropic dis-
sipation rate can be invoked (Zeman 1990; Sarkar et al. 1991; see also Wilcox
1998) to explicitly account for compressibility eects. The form of the parti-
tioning is
= +
d
(2.51)
where is the solenoidal (incompressible) dissipation and
d
is the dilatation
(compressible) dissipation.
A common high-Reynolds-number form of the isotropic solenoidal dissipa-
tion rate equation that is used is given by
D
Dt
== C
1
ik
u
i
x
k
C
2
2
k
+D
(2.52)
where
D
=
x
j
__
+C
ij
k
_
x
i
_
(2.53)
for a second-moment closure, or
D
=
x
j
_
_
+
t
_
x
j
_
(2.54)
for a two equation k- formulation. The rst and second terms on the right
side of (2.52) are the production and destruction of dissipation, respectively,
and the last term is the viscous diusion. The closure coecients C
1
, C
2
,
C
, and
T
, (3.1)
where l is the length scale of the large eddies in the radial (or transverse)
direction. In a supersonic boundary layer ow, M
g
is roughly constant ( 0.2);
whereas, in a mixing layer ow, for example, it is a linear function of the
convective Mach number (M
g
2.2M
c
; see Section 3.3.1).
3.1 Turbulent boundary layers
Just as in the previous section where the consequences of the strong Reynolds
analogy were presented, an equally important set of relations for the mean ow
behavior in turbulent boundary-layers can be derived. These are relations on
the variation of both mean velocity and mean temperature in the log-layer of
a compressible boundary-layer. This compressible law of the wall serves as a
necessary validation of both computational and experimental results as well
as the basis for compressible wall functions. Adaptations have been made for
ows other than at plate boundary-layers in zero pressure gradient; however,
for the purposes here in the derivation of the fundamental forms, the simple
case of planar ows in zero pressure-gradient will be considered.
3.1.1 Law of the wall
Whilst the very existence of a logarithmic portion of the mean velocity prole
near a wall is still a subject of debate in some quarters, and there is no con-
sensus on the exact values of the relevant constants, it is recognized that the
logarithmic law of the wall is a suciently good approximation to be useful for
modelling purposes. The compressible law of the wall (Van Driest 1951) denes
the behavior of both the mean velocity u and temperature
T, as well as the
turbulent shear stress
xy
in a region of the boundary layer. In zero pressure
gradient planar ow, the momentum and total energy equations (Cartesian
coordinates) near the wall become
tot
=
xy
xy
=
w
, (3.2)
where subscript w denotes conditions at the wall, and
q
tot
= q
y
+v
_
q
y
+c
p
_
u
w
= q u
w
= q
w
(3.3)
respectively. In extracting these equations from (2.11) and (2.13), advection
eects have been assumed to be small, and in the energy equation both mass
ux and turbulent and viscous transport have been neglected.
536 Barre et al.
In the log-law region the mean velocity and temperature proles are given
by
u
y
=
(
w
/)
1/2
i
y
, (3.4)
with
i
the (incompressible) Karman constant, and
T
y
= Pr
t
(q/c
p
)
(
w
/)
1/2
i
y
, (3.5)
respectively. In the region of the boundary layer where (3.4) and (3.5) hold,
the total heat ux is taken to be
q = c
p
t
Pr
t
T
y
(3.6)
so that
c
p
t
Pr
t
T
y
= q
w
+ u
w
(3.7)
where
t
=
w
_
u
y
_
1
(3.8)
The temperature distribution can then be (formally) extracted from (3.7) to
give
T
T
w
Pr
t
c
p
_
q
w
u
w
+
u
2
2
_
. (3.9)
Equation (3.9) should be viewed with caution since it is only an approximate
relationship because the integration limit on the temperature gradient ex-
tends directly to the wall (the expression is not valid in the viscous sublayer).
Equation (3.9), coupled with the compressible law of the wall (3.5), yields a
relationship between the wall temperature and the wall heat ux.
Since the pressure is assumed constant in the wall layer, the compressible
law of the wall for velocity (3.4) can be written in terms of the friction velocity
u
(=
_
w
/
w
) as
u
y
=
u
i
y
T
w
. (3.10)
This velocity log-law can then be combined with the temperature relation (3.9)
to obtain
u
vd
=
u
c
ln
_
y u
w
_
+B
c
(3.11)
or
u
vd
_
2c
p
T
w
Pr
t
_
1/2 _
arcsin
_
q
w
+ u
w
w
D
_
arcsin
_
q
w
w
D
__
, (3.12)
[19] Compressible, high speed ows 537
where the Van Driest velocity u
vd
is dened by
u
vd
=
_
_
T
w
T
_
1/2
d u, (3.13)
and
D =
_
q
2
w
2
w
+
2c
p
T
w
Pr
t
_
1/2
(3.14)
c
=
c
(M
, B
q
, Pr
t
), B
c
= B
c
(M
, B
q
, Pr
t
), B
q
= q
w
/
w
c
p
u
T
w
. (3.15)
These functional dependencies generally have minimal eect in the low super-
sonic regime so that the values for
c
and B
c
are close to their incompressible
ones (e.g. Huang et al. 1993).
3.2 Compressibility eects
In the evaluation of both experimental and numerical results of compressible
boundary layer ows under suitable conditions, compliance with the com-
pressible law of the wall is an important test which needs to be applied to the
results.
3.2.1 Mean density gradient
One of the best examples of how the log-law can be used to assess turbulent
model performance can be found in the study of Huang et al. (1994), where
it is shown how mean density gradients can signicantly alter two-equation
turbulence model calibrations for boundary layer ows.
In order to ensure that the law of the wall is maintained, a unique rela-
tion among turbulence model closure constants must exist. For incompressible
ows, this constraint leads to the condition that
=
2
i
(C
2
C
1
)
_
C
. (3.16)
Huang et al. 1994 showed that to recover the correct log-law in ows with
mean density gradients one should take
=
i
2
_
C
2
C
1
_
_
C
(3.17)
where
i
2
=
2
i
_
1 +
y
d
dy
3
2
y
2
d
2
dy
2
+ 3
_
y
d
dy
_
2
_
(3.18)
C
1
= C
1
_
1 +
2
i
_
C
K
_
y
d
dy
+
y
2
d
2
dy
2
3
2
_
y
d
dy
_
2
__
. (3.19)
538 Barre et al.
If a k- formulation is used, the mean density eect is present but does not
have as severe an impact on the log-law. (See Viala (1995) for the impact of
this eect on a variety of two equation models.)
3.2.2 Turbulent heat ux
As was shown in (2.17), the heat ux plays a prominent role in the mean (total)
energy equation. In the compressible formulation, it also directly aects the
pressure-velocity correlation p
i
which appears in the turbulent transport
term D
ij
in (2.45) since
p
i
p
=
i
T
T
+
(3.20)
or
p
i
= Ru
i
T
+R
i
= R
u
i
T
+R
i
. (3.21)
A representative distribution of the heat ux and mass ux contributions (Rai
et al. 1995) for the streamwise p
and cross-stream p
pressure-velocity
correlations is shown in Figure 3. (Variables are scaled with free stream velocity
and pressure. The spanwise correlation p
, (b) p
.
these are summarized in Gatski (1996). Gradient-diusion models are the most
popular and simplest closures for the turbulent heat uxes. However, more
recent attempts have taken into account variable Prandtl number behavior.
This has required the solution of transport equations for the temperature
variance as well as models for the dissipation of temperature variance. In
addition, analogous to the development of algebraic stress models in terms
of tensor representations involving the mean strain rate and rotation rate
[19] Compressible, high speed ows 539
tensors, explicit algebraic scalar ux models have also been developed (e.g.
Wikstr om et al. 2000). While these models are not specically developed for
compressible ows, their applicability to such ows is based on the assumption
that a variable mean density extension is sucient. This approach can be
partially validated from an analysis of direct numerical simulation results.
Huang et al. (1995) (using the DNS data of Coleman et al. 1995) showed
that in a cold wall, compressible boundary layer, at Mach numbers of 1.5 (Case
A) and 3.0 (Case B), the (vertical) turbulent heat ux is given by
v
=
v
. (3.22)
Huang et al. (1995) retained the triple uctuating product term in (3.22) since
its eect did extend outside the sublayer and had an eect on the heat ux
prole near its maximum value. Even then its contribution was only about 5%
of the total heat ux
v
(=
v
),
v
, and
was not
signicant indicating a near equality between the quantities
v
and v
1 10 100 1000
y
+
0.000
0.050
0.100
0.150
0.200
c
p
vT
c
p
vT
c
p
vT
1 10 100 1000
y
+
0.0000
0.0500
0.1000
0.1500
0.2000
c
p
vT
c
p
vT
c
p
vT
(a)
_____
___
_
_____
_____
___
_
_____
(b)
Figure 4: Distribution of turbulent heat ux quantities across channel. Re-
sults from Huang et al. (1995): (a) Case A, (b) Case B. (Quantities shown
normalized with wall values.)
3.3 Mixing Layers
The compressible mixing layer is studied as a canonical ow that is nevertheless
representative of many practical applications, such as supersonic mixing in
scramjet engines, infra-red and acoustic signature of hot jets, and afterbody
ows on aircraft. Studies have shown that plane mixing layers, annular mixing
layers and, in some cases, jets can be assumed to exhibit very comparable
behavior when dealing with mean or turbulent ow characteristics.
540 Barre et al.
3.3.1 Characteristic features
A schematic representation of a mixing layer is shown in Figure 5. Two paral-
lel ows (with corresponding mean velocities U
1
, U
2
; Mach numbers M
1
, M
2
;
speeds of sound a
1
, a
2
; and densities
1
,
2
) merge together after the trail-
ing edge of a splitter plate. As these ows merge, the Kelvin-Helmholtz (KH)
instability generates the structure shown in Figure 6 where shadowgraph visu-
alizations of subsonic variable density plane mixing layers at dierent Reynolds
number from Brown and Roshko (1974) are shown. The Reynolds number in
case (a) is four times the one in case (e). It is clear that increasing Reynolds
number leads to more and more small scale turbulent structures, but the initial
large scale structures resulting from the KH instability persist even at the high-
est Reynolds numbers. Large-scale structures are still present in supersonic
U
2
, M
2
, a
2
,
2
U
1
, M
1
, a
1
,
1
Figure 5: Schematic representation of a plane mixing layer.
ows as shown in Figure 7 where tomographic visualizations are presented for
four plane mixing layer ow congurations (from Debisshop 1993, and Cham-
bres 1997) covering a convective Mach number range from 0.525 to 1. These
visualizations are instantaneous pictures taken with an exposure time of 30ns.
The supersonic ow (upper ow on Figure 7) is seeded with ether droplets and
the picture is obtained by collecting Mie scattered light. These pictures can
be viewed as instantaneous images of the ow structure. In the M
c
= 0.525
case, 2D large scale structures appear clearly and are almost as sharp as in
the subsonic case of Brown and Roshko (1974). However, when the convec-
tive Mach number is increased up to 0.64, the ow appears less organized. In
particular, the upper edge of the shear layer seems to be more horizontal and
less disturbed by the large scale structures than in the M
c
= 0.525 case. The
decrease in spreading rate with increasing M
c
is also visible in these pictures.
When the convective Mach number is increased up to M
c
= 1, the visualiza-
tion appears to show a more uniform mixing zone with a sharp boundary and a
more homogeneous internal ow. Two pictures are given for the M
c
= 1 ow in
order to show the repeatability of the visualization process. Large scale struc-
tures are less organized here than in the low convective Mach number cases
and are highly three-dimensional, as was clearly shown in sequence of experi-
ments carried out by Clemens and Mungal (1992) and numerical simulations
[19] Compressible, high speed ows 541
(a)
(b)
(c)
(d)
(e)
Figure 6: Eect of Reynolds number on the structure of low speed plane mixing
layers. From (a), high Reynolds number to (e), low Reynolds number. (See
Brown and Roshko (1974) for details.)
by Sandham and Reynolds (1991). The trend towards three-dimensionality is
reproduced in the linear stability characteristics of compressible mixing layers
(Sandham and Reynolds 1990). Such calculations also reproduce the growth
rate reduction of mixing layers indicating that mean ow instabilities, large
scale structures and shear layer spreading rate are coupled by the same under-
lying ow physics. The persistence of large scale structure over a wide range
of Mach and Reynolds numbers is an important feature of mixing layers, and
needs to be taken into account when developing models of such ows.
When the Reynolds number is suciently large the ow reaches an asymp-
totic downstream state where self-similarity of the (properly normalized) mean
and turbulent elds is observed. The approximate (error function) shape of the
longitudinal mean velocity prole for low speed ow is shown in Figure 8. From
this prole one can dene the thickness of the shear layer. It is, in fact, pos-
sible to dene this thickness in many dierent ways that are not completely
542 Barre et al.
Bord de fuite
M
c
= 0.525
M
c
= 0.64
M
c
= 1
M
c
= 1
10 20 30
X(cm)
Y
Figure 7: Tomographic visualizations of three kind of compressible plane mix-
ing layers at M
c
= 0.525, 0.64, and 1 (from Chambres 1997).
U
1
U
2
=
U
_
U
y
_
max
, (3.23)
where U = (U
1
U
2
) is the velocity dierence between the external ows,
and (U/y)
max
is the maximum mean velocity gradient of the ow.
When the self-similar state is reached, the mixing layer spreads linearly
with a constant value of the spreading rate
x
_
0
= C
(1 r) (1 +
s)
2 (1 +r
s)
, (3.24)
with r = U
2
/U
1
, s =
2
/
1
, and C
1
/
2
0.20
0.15
0.10
0.05
0
.1 .2 .4 .6 .8 1
1 2 3 4 5 6
2 4 6 810
M
1
Figure 9: Eect of density ratio on spreading rate with U
2
= 0: , values
for incompressible ow. Other symbols are for compressible mixing layers: +,
Maydew and Reed (1963); , Ikawa (1973); , Siriex and Solignac (1966).
(From Brown and Roshko 1974)
it appears that an extra parameter must be introduced in order to quantify
the compressibility eect. The most useful parameter has been the convective
Mach number originally named by Papamoschou and Roshko (1988) who drew
upon linear stability analysis of a vortex sheet for justication of the particular
formulation. Figure 10 clearly shows the basis of the convective Mach number
concept. In a frame of reference xed to large scale structures traveling with
544 Barre et al.
a constant convective velocity U
c
, one can dene two Mach numbers of the
external streams
M
c
1
=
U
1
U
c
a
1
and M
c
2
=
U
c
U
2
a
2
. (3.25)
In the convected frame of reference there exists a saddle point between two
U
1
U
c
U
c
U
1
U
c
U
c
U
2
M
2
, U
2
,
2
M
1
, U
1
,
1
a
1
M
c
1
=
U
c
U
2
a
2
M
c
1
=
(a)
(b)
Figure 10: Convective Mach number concept (Papamoschou and Roshko 1988):
(a) laboratory frame of reference; (b) convective frame of reference.
adjacent eddies. If it is assumed that the static pressure is constant across the
mixing layer and that the compression of uid along particle paths leading to
the saddle points is isentropic, then the mechanical equilibrium of this saddle
point leads to the relation
_
1 +
(
1
1)
2
M
2
c
1
_
1
(
1
1)
=
_
1 +
(
2
1)
2
M
2
c
2
_
2
(
2
1)
, (3.26)
where
1
and
2
are the ratios of specic heats for the two streams of uid.
Using the denition of the two convective Mach numbers, an expression for
U
c
can be obtained as follows for the simplied case where
1
=
2
U
c
=
a
2
U
1
+a
1
U
2
a
1
+a
2
. (3.27)
Papamoschou and Roshko (1988) thus obtained a new functional form for the
mixing layer spreading rate,
_
x
_
= C
(1 r) (1 +
s)
2 (1 +r
s)
(M
c
) =
_
x
_
0
(M
c
) , (3.28)
[19] Compressible, high speed ows 545
where (M
c
) is a compressibility function which is the ratio between the ob-
served spreading rate of a compressible mixing layer and the incompressible
value at the same velocity and density ratio. Experimental data to support
this are given in the next section.
3.3.2 Experimental and DNS results for RANS model validation
Many measurements of mean and turbulent quantities have been carried out
using either hot-wires (Barre et al. 1994) or LDV (Elliott and Samimy 1990,
Goebel and Dutton 1991, Gruber et al. 1993, Debisschop 1993, and Chambres
1997) in both plane mixing layers and supersonic jet congurations (Lau 1981,
Bellaud 1999). The majority of these studies have been summarized in Lele
(1994) and more recently in Smits and Dussauge (1996).
Both experimental and numerical studies have conrmed that the spreading
rate decreases when the convective Mach number increases. Figure 11 shows
the reduction in the normalized spreading rate (M
c
) with convective Mach
number for a wide range of experiments. Between M
c
= 0.5 and M
c
= 1 there
Plane Mixing Layer
Annular Mixing Layer
Chambres (1997)
Debisschop (1993)
Elliot and Samimy (1990)
Goebel and Dutton (1991)
Papamoschou et al. (1998)
Lau (1981)
Mistral (1993)
Bellaud (1999)
0.20
0.00
0.25 0.50 0.75 1.00 1.25 1.50
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
1.10
1.20
M
c
(M
c
)
Figure 11: Normalized growth rate versus convective Mach number for dierent
mixing layer studies.
is a reduction by more than a factor of two. Above M
c
= 1 the normalized
spreading rate is relatively constant. The reduction in spreading rate appears
to be also accompanied in a majority of the experiments by a corresponding
reduction in the turbulent uctuations. Figures 12, 13, and 14 show results
from several dierent plane and annular mixing layer studies of the variation
of the normalized Reynolds stress components
u
(=
_
u
2
),
v
(=
_
v
2
),
and u
u
U
max
Figure 12: Streamwise velocity uctuations as a function of the compressibility
parameter M
c
for plane and annular mixing layers.
Plane Mixing Layer
Annular Mixing Layer
Chambres (1997)
Debisschop (1993)
Elliot and Samimy (1990)
Goebel and Dutton (1991)
Gruber et al. (1993)
Bradshaw et al. (1964)
Denis et al. (1998)
Mistral (1993)
Bellaud (1999)
0.04
0.00 0.20 0.40 0.60 0.80 1.00 1.20
0.06
0.08
0.10
1.12
1.14
1.16
M
c
v
U
max
Figure 13: Transverse velocity uctuations as a function of the compressibility
parameter M
c
for plane and annular mixing layers.
with increasing M
c
. This behavior has led to the conclusion by some that the
turbulent anisotropy, dened by
max
and
u
u
+
v
max
, (3.29)
exhibits a quasi-constant behavior over the convective Mach number range.
Figure 15 shows experimental results obtained for (
u
/
v
)[
max
. For the ma-
jority of these studies, the anisotropy data is relatively constant; however, the
studies of Goebel and Dutton (1991) and Gruber et al. (1993) actually show a
[19] Compressible, high speed ows 547
Plane Mixing Layer
Annular Mixing Layer
Barre (1993)
Chambres (1997)
Debisschop (1993)
Elliot and Samimy (1990)
Goebel and Dutton (1991)
Gruber et al. (1993)
Bradshaw et al. (1964)
Denis et al. (1998)
Lau (1981)
Mistral (1993)
Bellaud (1999)
0.02
0.00 0.20 0.40 0.60 0.80 1.00 1.20
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
M
c
u
(U )
v
2
max
Figure 14: Turbulent shear stress evolution as a function of the compressibility
parameter M
c
for plane and annular mixing layers.
growth with M
c
. The Gruber et al. (1993) study is limited to a single convec-
tive Mach number, but the result is consistent with the Goebel and Dutton
(1991) result. The reason for this growth in (
u
/
v
)[
max
is due to the fact
that Goebel and Dutton (1991) show a relatively constant value in
u
/U
but a decrease in
v
/U with M
c
(see Figs. 12 and 13). The result is an ef-
fective increase in (
u
/
v
)[
max
with an increase in the compressibility eects
(see Figure 15). This contradiction in the experimental results is worrying if
one wishes to obtain a physical interpretation of the eects of compressibility
and develop predictive models. Also the fundamental question still persists:
What are the real mechanisms responsible for the decrease of the mixing and
the turbulent activity at high Mach numbers?
Some studies were carried out almost a decade ago to try to answer this
question. From an analysis of numerical simulations, Zeman (1990) and Sarkar
et al. (1991) proposed models for the extra-dissipation due to dilatational
eects in order to explain the reduction in the turbulence. The functional
form for these models was given by
= +
d
= [1 +T(M
t
)], (3.30)
where the extra-dissipation was represented by the turbulent Mach number
term T(M
t
). This form of the partitioning allowed for the direct extension of
the incompressible (solenoidal) form of the dissipation rate transport equation.
Utilization of this type of model brought the calculations into agreement with
the experimental results available. An example of the signicant inuence this
term had on the predictive behavior of a Reynolds stress model is shown in
Figure 16 with the function T(M
t
) given simply by M
2
t
(M
t
is the turbulent
548 Barre et al.
Plane Mixing Layer
Annular Mixing Layer
Chambres (1997)
Debisschop (1993)
Elliot and Samimy (1990)
Goebel and Dutton (1991)
Gruber et al. (1993)
Bradshaw et al. (1964)
Denis et al. (1998)
Mistral (1993)
Bellaud (1999)
1.2
0.00 0.20 0.40 0.60 0.80 1.00 1.20
1.6
2.0
2.4
2.8
3.2
3.6
M
c
v max
Figure 15: Anisotropy evolution as a function of the compressibility parameter
M
c
for plane and annular mixing layers.
0
0.0
0.4
0.2
0.6
0.8
1.0
1 2 3 4
Without extra - dissipation
With extra - dissipation
"Langley Experimental Curve"
5
(M
c
)
M
c
Figure 16: Predicted compressibility function (M
c
) from a RANS calculation
using a Reynolds stress model. Langley Experimental Curve from Kline et
al. (1981).
Mach number). The RANS calculation is from Sarkar and Lakshmanan (1991)
who used a denition of (M
c
) which diered in detail (see their equation
(25)) from that given in (3.28); nevertheless, the gure clearly shows the same
qualitative trend of decreasing spreading rate with M
c
displayed in Figure
11. Such results, as well as those by Zeman (1990) who used a slightly more
complex form for T(M
t
), strongly suggested that the increase in compressible
dissipation and the resultant decrease in the turbulence was the underlying
dynamic reason for the reduction in spreading rate.
[19] Compressible, high speed ows 549
0.07
M
j
= 0.2
M
j
= 0.4
M
j
= 0.8
M
j
= 1.15
M
j
= 1.55
M
j
= 1.92
M
j
= 2.5
M
j
= 3.0
M
j
= 3.5
0.06
0.05
0.04
0.03
0.02
0.01
0
0 1 2 3
r/r
o
j
U
j
2
Figure 17: Normalized pressure uctuations for dierent Mach numbers.
However, with the emergence of direct numerical simulations, additional
features of such ows became evident. Sarkar (1995) identied turbulence
production as a key term for understanding the eect of compressibility on
homogeneous turbulence. For mixing layers Vreman et al. (1996) showed that
compressible dissipation remains negligible even at large Mach numbers. Ac-
cording to these simulations, it would seem that dierences in the level and
distribution of the pressure uctuations are central to the reduction in the tur-
bulent activity. With an increase in Mach number, one sees a reduction in the
pressure uctuations, an observation which has been conrmed in more recent
simulations by Freund et al. (1997) and Pantano and Sarkar (2001). Vreman
et al. use models of compressible vortices to conrm the origin of the reduction
of pressure uctuations. With this mechanism the reduction of growth rate is
not due to extra explicit compressibility terms in the equations, but due to an
implicit modication of existing terms (at second moment closure level this
would be the pressure strain). Since the pressure dilatation term remains small
there are no terms in the k or equations that can be modied. Instead, if we
follow the algebraic Reynolds stress modelling procedure to convert from sec-
ond moment closure to a two equation model, it is the eddy viscosity relation
that has to carry the compressibility eect.
Figure 17 (Freund et al. 1997) shows the pressure uctuation proles from
DNS of an annular mixing layer. Results for nine jet Mach numbers M
j
ranging
from 0.2 to 3.5 are given corresponding to convective Mach numbers ranging
from 0.1 to 1.8. It is clear from these direct numerical simulations that a strong
decrease in the maximum of
p
/
j
U
2
j
is observed with increasing Mach num-
ber. A rst consequence of an increase in Mach number across an eddy is to
limit the inuence of the pressure waves to a zone in space within which the
speed dierence is subsonic. This sonic-eddy concept was rst presented by
550 Barre et al.
Breidenthal (1990) and later extended by Papamoschou and Lele (1993). In-
deed, the inuence of this lack of communication by the pressure can be quan-
tied by the gradient Mach number (3.1) rst introduced by Sarkar (1995).
This Mach number can be interpreted as the ratio of the shearing velocity
across a large eddy to the sound speed. If this Mach number becomes too high
then the shearing will become too fast relative to the sound speed. This will
result in a decrease in the communication across the eddy and thus restrict
the structure size. Figure 18 shows the variation of gradient Mach number M
g
versus convective Mach number M
c
in the center of an annular mixing layer.
It is clear that, after a gradual increase, a saturation eect is reached when M
c
is close to unity. This trend is very similar to the one observed for spreading
rate and turbulent correlations. While this interpretation of the role played by
2.5
2.0
1.5
1.0
0.5
0
0 0.5 1.0 1.5 2.0
M
c
M
g
Figure 18: Gradient Mach number evolution versus convective Mach number
at the center of an annular mixing layer (DNS results from Freund et al. 1997)
the gradient Mach number as a measure of compressibility eects is certainly
plausible, most of the DNS results are not in agreement with the available
experimental results. The principal dierence is related to the structure of the
Reynolds stress tensor. With the exception of the Pantano and Sarkar (2001)
study, the numerical simulations (Sarkar 1995, Vreman et al. 1996, and Fre-
und et al. 1997) predict strong degrees of anisotropy when the convective Mach
number increases; whereas, the majority of the experimental results, excluding
those of Goebel and Dutton (1991) and Gruber et al. (1993), show that the
anisotropy of the Reynolds stress tensor is not strongly aected by compress-
ibility. Nevertheless, the changes in intensity of the pressure uctuations (and
pressure-dilatation) in the compressible case as compared to the incompress-
ible case is conrmed by the available compressible direct simulations.
All these results clearly suggest that the task of providing suitable modeling
for compressibility eects remains largely an open question. Recall that the
[19] Compressible, high speed ows 551
rst models for dilatation-dissipation (Zeman 1990, and Sarkar et al.1991),
postulated that the increase observed in the turbulent Mach number when the
level of compressibility increased could explain the decrease in turbulent activ-
ity for the compressible cases. This model gave completely compatible results
with experiments (Chambres 1997) for the turbulent intensities and anisotropy
parameters. On the other hand, direct numerical simulations (Vreman et al.
1996, Freund et al. 1997, Pantano and Sarkar 2000) now suggest that this rst
theory was inappropriate and that it must be replaced by a precise modeling of
the pressure-strain correlation. It appears that the Reynolds stress anisotropy
may be strongly aected by the initial condition and run time of DNS. In this
respect the results of Pantano and Sarkar (2001) now represent the closest
approximation that we have to a self-similar compressible mixing layer simu-
lation, which should be the starting point for physically-sound modelling work.
Another deciency is the restriction of DNS to ow Reynolds numbers which
are an order of magnitude lower than those of the experiments, thus making
generalizations of such DNS results questionable.
On the experimental side, several limitations still remain. For example, it is
not possible at this time to obtain reliable experimental data for some com-
pressibility terms, such as the pressure strain, which appears to play a principal
role in the dynamics. However, by using the Reynolds analogies and related
assumptions, it is now possible to establish budgets of turbulent kinetic energy
in supersonic mixing layers (Chambres 1997, Bellaud 1999). These results are
quite useful in validating closure models, and comparing the structure of the
Reynolds stress tensor with the one obtained in incompressible ows. Figure
19 shows the production and diusion terms in the kinetic energy budgets of
an annular mixing layer obtained from experiments of Bellaud (1999) and DNS
of Freund et al. (1997) at M
c
= 0.85. While these budgets are not completely
equivalent, a qualitative agreement is clearly shown. The most important fea-
ture revealed by both these budgets is the asymmetry of the diusion term
between the subsonic and the supersonic edge of the ow. At present, this
type of compressibility eect has no satisfactory explanation. Nevertheless,
with this type of agreement one can assume that these budgets across the
layer are a good approximation of the actual physical behavior which must be
modeled.
4 Shock Wave/Turbulence Interactions
The presence of a shock wave in a turbulent ow signicantly inuences the
turbulence energetics as well as the mean eld near and downstream of the
shock. The compressibility eects which inuenced the dynamics of the shock-
free ows discussed in the previous section are also expected to be relevant
in the presence of shocks. To make predictions, specic models have had to
be built in order to properly take into account the drastic modication of the
552 Barre et al.
Production, DNS
Production, Bellaud (1999)
Diffusion, DNS
Diffusion, Bellaud (1999)
1.20 0.80 0.40 0.00 0.40 0.80 1.20
0.20
0.40
0.60
0.20
0.00
0.40
0.60
0.80
1.00
Y
Figure 19: Comparison between experimental and DNS turbulent kinetic en-
ergy budgets across an annular mixing layer at M
c
= 0.85.
Flap
Over-expanded Nozzle
Transonic Flow
Separated flow
Sonic line
B.L. separation
Second shock reflection
Shock Reflection
Figure 20: Shock wave/turbulence interaction congurations.
turbulent eld subjected to strong pressure gradients.
Some types of ows of practical interest in which shocks are present are
illustrated by the sketches shown in Figure 20. Such shock wave/boundary
layer interactions have been extensively studied for many years both experi-
mentally and numerically. Review articles like those of Green (1970), Korkegi
(1974) and Fernholz and Finley (1981) summarize the main results for these
types of ow elds. In one set of experiments an oblique shock wave impinges
[19] Compressible, high speed ows 553
on a boundary layer. It has been shown by Smits and Muck (1987) and An-
dreopoulos and Muck (1987) that, in these conditions, if the shock is strong
enough the ow separates near the interaction. As for the evolution of the tur-
bulent quantities across the shock, it has been shown by Ardonceau (1984) and
Smits and Muck (1987), for example, that the turbulence intensities and the
anisotropy ratio are increased due to the shock interaction. The same kind of
results are available for oblique shock wave interaction with a free shear layer.
Data show a high amplication of the turbulence intensities which depends on
the shock strength (Dolling and Murphy 1982, Settles et al. 1982, Hayakawa
et al. 1984, Samimy and Addy 1985, and Jacquin et al. 1991).
Additional experimental studies have been performed either in a shock tube
in order to study the (pure) interaction between a traveling normal shock wave
reected on the end wall of the tube and the subsequent interaction with the
ow induced by the incident shock (Troiler and Duy 1985, Hartung and Duy
1986) or by the wake of a perforated plate (Honkan and Andreopoulos 1992,
Keller and Merzkirch 1990, Briassulis and Andreopoulos 1996, Briassulis et
al. 1999, Poggi et al. 1998). All these experiments show a strong increase in
density uctuations through the shock, but there is little data describing the
behavior of the velocity eld.
Unfortunately, in very complex congurations streamline curvature and un-
steady separation problems both occur. This makes the study and the pre-
diction of shock wave/boundary layer interaction problems dicult. Thus, it
is desirable to nd a simpler ow problem to study which has at least some
of the key dynamic features of the general problem, but without some of the
complicating features. One such basic ow problem is the interaction between
a normal shock and an isotropic turbulent eld. The main advantage of in-
vestigating such a ow is to isolate the net eect of the strong gradients of
mean quantities imposed by the shock wave. This basic conguration has many
advantages for both experiments and computations. For the experiments, it
reduces the number of geometrical parameters, while, for numerical predic-
tions, analytical theories as well as advanced simulation methods (LES, DNS,
EDQNM) can be applied and compared.
The remainder of this section will be divided in two parts describing the
current understanding of the shock wave/homogeneous turbulence interaction
and shock wave/boundary layer interaction problems.
4.1 Shock wave/homogeneous turbulence interaction
Early theoretical work on this problem was done in the 1950s, based on linear
analysis of a plane disturbance interacting with a shock wave. Ribner (1953,
1954, 1969), in particular, investigated how a shock wave is perturbed by an
impinging single wave. The theory, later called Ribners theory, is based on
the mode theory developed by Kovasznay (1953) and extended to the shock
problem by focusing mainly on the acoustic eld generated by the interaction
554 Barre et al.
with the shock wave. More recently, numerical simulations of such ows have
emerged. The Rotman (1991), Lee et al. (1993, 1997), and Mahesh et al. (1995,
1997) simulations have shown, among other results, that the vorticity ampli-
cations predicted are in good agreement with linear theories, and the DNS
studies of Lee et al. (1993, 1997), Jamme (1998) and Hannapel and Friedrich
(1992, 1995) have shown how isotropic turbulence becomes axisymmetric after
the shock, which is also consistent with the linear analysis. In addition, the
evolution of the turbulent spectra across the shock exhibits a bigger ampli-
cation at large wave numbers compared to lower ones which means that the
longitudinal integral length scales are decreased while the lateral ones seem
unaected by the shock interaction. In this section we consider in detail the
predictions of linear theory based on the work of Jamme (1998), followed by
comparisons with experiment.
4.1.1 Linear interaction analysis (LIA)
The problem under study is the interaction between a normal shock wave and
an isotropic incoming turbulent eld. It is assumed that this interaction takes
place in an innite medium with no boundaries. In this kind of ow eld the
mode theory rst proposed by Kovasznay (1953) can be applied. The theory
results from a linearization of the NavierStokes equations for moderately sup-
ersonic ows. In this case, if the turbulence level is suciently small to keep
the linear approximation valid, the uctuation eld can be reduced to a linear
combination of three fundamental modes described by
Vorticity Mode :
u
u
,
T
T
= 0,
p
p
= 0,
= 0 (4.1)
Entropy Mode :
T
T
,
u
u
= 0,
p
p
= 0,
=
T
T
(4.2)
Acoustic Mode :
=
p
p
=
1
( 1)
T
T
, with
u
u
= f
_
p
p
_
. (4.3)
Note that here we are using uctuations from conventionally averaged quan-
tities, rather than the mass-weighted averages introduced earlier. In the lin-
earized ow the two formulations are equivalent. Any turbulent eld satisfying
the initial assumptions can be reduced to a linear combination of these three
independent modes. Furthermore, if there is a non-random phase relationship
between these modes, one can then obtain (non-random) correlated turbulent
elds for some physical quantities of interest.
For the shock wave/turbulence interaction, the physics of the problem can
be modeled by using this modal decomposition for both the incoming and
the transmitted turbulence across the shock wave. This physical situation is
sketched in Figure 21, where the incoming and refracted turbulent eld are
[19] Compressible, high speed ows 555
shown. In the general case the incoming turbulent ow is three-dimensional
and consists of many waves with a continuous spectrum representative of a
large Reynolds number isotropic turbulence. However, for clarity the linear
interaction theory will be used in a simplied case.
Acoustic
Acoustic
Entropy
Entropy
Vorticity
Vorticity
Figure 21: Schematic of incoming and refracted turbulent eld with shock.
First, consider the two-dimensional interaction of a simple vorticity/entropy
wave with a two-dimensional normal shock wave. This simplied case is of great
interest because the vorticity/entropy wave structure is representative of the
classical turbulent eld encountered in many moderately supersonic turbulent
ows. The interaction between simple vorticity or entropy waves was rst done
analytically by Ribner (1953) for a vorticity wave and by Chang (1957) for an
entropy wave.
The analysis presented here is adapted from Jamme (1998) who studied the
problem shown in Figure 22. The shock wave is distorted by the incoming
wave eld. The incoming wave can be expressed as
Vorticity Mode :
_
_
u
1
u
1
= lA
v
e
i(mx+lyU
1
mt)
v
1
u
1
= mA
v
e
i(mx+lyU
1
mt)
(4.4)
Entropy Mode :
_
1
= A
e
e
i(mx+lyU
1
mt)
T
1
T
1
=
1
p
1
= 0,
(4.5)
where
l = sin
1
and m = cos
1
556 Barre et al.
Figure 22: Sketch of a linear shock/wave interaction.
where =
_
2
x
+
2
y
is the wavenumber of the incoming wave, A
v
is the
amplitude of the vorticity wave upstream of the shock, A
e
is the amplitude of
the entropy wave upstream of the shock, and the subscript 1 refers to upstream
conditions relative to the shock. The Euler equations downstream of the shock
can be written as
2
t
+U
2
2
x
=
2
_
u
2
x
+
v
2
y
_
u
2
t
+U
2
u
2
x
=
1
2
p
2
x
v
2
t
+U
2
v
2
x
=
1
2
p
2
y
s
2
t
+U
2
s
2
x
= 0
p
2
p
2
=
2
+
T
2
T
2
and
s
2
C
p
=
p
2
p
2
2
.
(4.6)
The initial conditions for the solution of Eulers equations across the shock
[19] Compressible, high speed ows 557
can be given by the linearized RankineHugoniot equations
u
2
(x = 0)
U
1
= (1 B
1
)
1
U
1
t
+B
1
u
1
U
1
+B
2
T
1
T
1
v
2
(x = 0)
U
1
=
v
1
U
1
+E
1
2
(x = 0)
2
= C
1
1
U
1
_
u
t
_
+C
2
T
1
T
1
p
2
(x = 0)
p
2
= D
1
1
U
1
_
u
t
_
+D
2
T
1
T
1
,
(4.7)
where B
1,2
, C
1,2
, D
1,2
, and E
1
are functions of and M
1
. With these starting
conditions, the eld downstream of the shock wave can be written as
u
2
u
1
= Fe
ix
e
i(lyU
1
mt)
+Ge
i(mrx+lyU
1
mt)
v
2
u
1
= He
ix
e
i(lyU
1
mt)
+Ie
i(mrx+lyU
1
mt)
p
2
p
2
= Ke
ix
e
i(lyU
1
mt)
2
=
K
e
ix
e
i(lyU
1
mt)
+Qe
i(mrx+lyU
1
mt)
T
2
T
2
=
1
Ke
ix
e
i(lyU
1
mt)
Qe
i(mrx+lyU
1
mt)
,
(4.8)
where r is ratio of the velocity upstream of the shock (U
1
) to the velocity
downstream of the shock (U
2
). The shock wave deformation equation can then
be given by
1
U
1
t
= Le
i(lymU
1
t)
y
=
l
m
Le
i(lymU
1
t)
.
(4.9)
The seven coecients F, G, H, I, K, Q, and L appearing in (4.8) and (4.9) are
the transfer coecients between the upstream and downstream shock ows.
They are obtained by solving a system of seven equations, composed of the
three equations of motion downstream of the shock and the four equations
resulting from the RankineHugoniot conditions across the shock. The int-
erested reader is referred to the reference of Jamme (1998) for details. This
558 Barre et al.
leads to two kinds of solutions for the pressure uctuations downstream of
the shock wave. These two families of solutions depend on the angle
1
of
the wave vector. There exists a critical angle
c
for the wave vector direction
which separates the two kinds of pressure wave solutions. The value of this
critical angle can be computed as a function of the upstream velocity U
1
, the
downstream velocity U
2
, and the speed of sound a
2
,
cot
2
(
c
) =
_
a
2
U
1
_
2
_
U
2
U
1
_
2
. (4.10)
Then, knowing
c
, the two families can be characterized as either (i) an evanes-
cent mode (
c
<
1
<
2
), where the acoustic wave after the shock decreases
exponentially or (ii) a non-evanescent mode (0 <
1
<
c
), where the acoustic
wave after the shock propagates downstream without attenuation.
The three-dimensional generalization of this theory can be obtained by inte-
grating (over physical and Fourier space) the results from all the contributing
waves included in the incoming turbulent eld. This can include initially
acoustic elds. An initial spectrum which gives a good representation of a real
turbulent eld (see Jamme (1998) for details) is
E() = C
4
e
2
, (4.11)
In this section, LIA results are presented for four kinds of initial elds corre-
sponding to typical cases which are most often encountered in real turbulence.
These are:
(1) Pure solenoidal turbulence (vorticity mode);
(2) Pure entropic turbulence (entropy mode);
(3) Mixed turbulence: in phase vorticity and entropy modes
_
2
1
=
_
T
2
1
T
= ( 1) M
2
1
_
u
2
1
U
1
= 0 = u
< 0
( is the phase shift between the vorticity and entropy waves);
(4) Mixed turbulence: out of phase vorticity and entropy modes
_
2
1
=
_
T
2
1
T
= ( 1) M
2
1
_
u
2
1
U
1
= = u
> 0.
[19] Compressible, high speed ows 559
For these four cases, Figure 23 shows the streamwise evolution of the LIA
computed turbulent kinetic energy (q
2
) amplication rate for a ow with M
1
=
1.5. The most striking result is that the nature of the incoming turbulence
plays a crucial role in the evolution of the amplication rate of q
2
. For Case
(1) (pure solenoidal turbulence) a 40% amplication of the initial turbulent
kinetic energy (TKE) is observed through linear mechanisms. By contrast the
pure entropy wave of Case (2) has its initial kinetic energy reduced by a factor
of ten. When these two waves are coupled together, it is possible to obtain a
larger TKE amplication if u
2
2
/v
2
2
(because of the
assumed isotropy of the incoming eld u
2
1
= v
2
1
). In contrast, the far eld (see
Figure 26) does not exhibit such an asymptotic behavior for the anisotropy
[19] Compressible, high speed ows 561
1.0
1.5
2.0
2.5
3.0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
M
1
2
2
v
'
2
1
v
'
2
2
q
2
1
q
2
2
u'
2
1
u'
Figure 26: Amplication coecients for turbulent velocity uctuations down-
stream of the shock (far eld) for Case (3) (in phase mixed turbulence), from
Jamme (1998).
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1.0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
M
1
Mixed upstream
turbulent field Case 4
Mixed upstream
turbulent field Case 3
Pure vorticity upstream
turbulent field
()
(
1
)
upstream
Figure 27: Amplication coecient for longitudinal Taylor microscale in the
far eld (from Jamme 1998).
coecient. The amplication ratio of the turbulent kinetic energy reaches a
value of about 2.6 beyond Mach 3.
Another measure of interest is the Taylor microscale. the LIA predicts a
decrease for both the longitudinal and lateral scales (see Figs. 27 and 28). For
the longitudinal Taylor microscale, the ratio can reach 0.4 for M
1
= 5 (Figure
27).
LIA theory also predicts the evolution of the thermodynamic characteristics
of the turbulent eld across the shock. The thermodynamic uctuations can be
characterized by an analogy with a pure sound eld or a pure vorticity-entropy
eld. As an example, consider the inuence of the pressure uctuations on the
density uctuation eld. The temperature and density uctuations are related
562 Barre et al.
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
1.0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
M
1
Mixed upstream
turbulent field Case 3
Mixed upstream
turbulent field Case 4
Pure vorticity upstream
turbulent field
(
2
)
upstream
(
2
)
Figure 28: Lateral Taylor microscale amplication coecient downstream (far
eld) (from Jamme 1998).
to the pressure and entropy uctuations through
T
2
T
2
=
( 1)
2
2
p
2
p
2
. .
Acoustic
+
s
2
C
2
p
..
Entropy
2
=
1
2
p
2
p
2
. .
Acoustic
+
s
2
C
2
p
..
Entropy
(4.12)
2
T
2
T
2
=
1
2
p
2
p
2
( 1)
2
2
p
2
p
2
=
( 2)
p
2
p
2
. (4.13)
A ratio between these quantities can be constructed and is given by
=
_
_
T
2
T
_
=
( 2)
n
2
p
n
T
(4.14)
with
n
p
=
_
p
2
/p
_
2
/
and n
T
= 1 +
_
T
2
/T
_
2
/
, (4.15)
where n
p
and n
T
are polytropic coecients.
For a pure isobaric turbulence only the vorticity and entropy modes are
active and we have
p
p
= 0 = = 0 =
_
n
p
= 0
n
T
= 2.
(4.16)
[19] Compressible, high speed ows 563
For a pure acoustic eld we may write:
s
s
= 0 = = 0.6 =
_
n
p
=
n
T
= .
(4.17)
The consequence of this is that a given turbulent ow will satisfy Morkovins
hypothesis when 0.6.
Figure 29 shows the evolution of the LIA computed polytropic coecients
(downstream far eld conditions) as a function of the incoming Mach num-
ber. The upstream turbulence was composed of a pure vorticity wave in this
computation. For small Mach number (M
1
< 1.2), the downstream turbulent
eld is isentropic, but as the Mach number increases the polytropic coe-
cients reach asymptotic values of n
T
= 1.9 and n
p
= 0.6. The resulting eld
1.6
1.2
0.8
0.4
0.0
2.0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
M
1
n
pT
n
p
Figure 29: Evolution of polytropic coecients computed from LIA in far eld
for pure vorticity upstream turbulence (from Jamme 1998).
now exhibits non-negligible entropy uctuations which were not present in the
incoming eld. In the case of composite upstream turbulence (vorticity and
entropy mode with u
1
T
1
< 0) and for M
1
= 1.5, the LIA results in the far
eld downstream of the shock are
n
p
= 0.4
n
t
= 1.96
_
= = 0.035 =
0.6
= 0.058. (4.18)
Thus, after the shock, the turbulent eld is a 5.8% pure sound eld. The
turbulence then satises Morkovin hypothesis even though it is not isobaric.
In this case p
2
p
= 0.4
_
. (4.19)
564 Barre et al.
This example illustrates the fact that it is possible to obtain turbulent elds
which satisfy Morkovin hypothesis without the necessity of having p
/p = 0.
4.1.2 Comparison with experimental results
Only a few attempts have been made to study such a ow. Blin (1993) and
Jacquin et al. (1991) generated a supersonic quasi-homogeneous turbulent ow
using a grid as a sonic throat. The ow downstream of the grid was super-
sonic with a relatively low Mach number (1.7), but with parasitic shock waves
which signicantly deteriorated the quality of the turbulent ow eld. It was
necessary to shock the ow by means of a second throat creating a pure
normal shock wave/free turbulence interaction. Unfortunately, experimental
diculties due to the very low turbulence level (close to the noise level of the
anemometer) led to signicant diculties in interpreting the results. On the
other hand Debieve and Lacharme (1986) obtained a quasi-homogeneous tur-
bulence by generating perturbations in the settling chamber of a supersonic
wind tunnel at Mach 2.3. They showed that, after the expansion in the nozzle,
the turbulent eld becomes strongly non-isotropic and the turbulence level is
drastically decreased, which leads to experimental diculties. A third kind of
experiment was performed by Barre et al. (1996) who used a multi-nozzle sys-
tem to generate a homogeneous and isotropic turbulent eld in a Mach 3 ow.
This turbulent eld then interacted with a normal shock wave generated at
the center of the test section of a supersonic wind tunnel by means of a Mach
eect. The main advantage of such a device is to obtain a shock wave free from
parasitic excitation, making this interaction quite close to an ideal interaction
and in agreement with the assumptions for linear theories. However, some lim-
itations still persist in this conguration. First, the turbulent eld issued from
the multi-nozzle contains some acoustic waves generated at the trailing edge
of the multi-nozzle. These waves, even if they are damped when they reach
the shock wave, make it increasingly dicult to interpret the results because
they have a large inuence on the interaction with the initial turbulent eld.
Second, the turbulence level is quite low immediately upstream of the shock
wave which leads to experimental diculties as well. However, as noted by
Barre et al. (1996), this experimental conguration yielded some interesting
results relating to the behavior of the turbulence in such a shock/turbulence
interaction. In order to obtain better quality results in the near future, im-
portant modications to this experimental device are in progress to suppress
these two main limitations just described. It is hoped that better quality for
both the incoming ow and the interaction itself can be obtained.
For the remainder of this section we focus on the experimental results pre-
sented in Andreopoulos et al. (2000) for shock interaction with turbulence.
Figure 30 shows the behavior of the amplication ratio for the velocity uctu-
ations G
u
2 = u
2
/u
2
0
across the shock from experiments, LIA, rapid distortion
theory (RDT, see Andreopoulos et al. for references), and DNS at several up-
[19] Compressible, high speed ows 565
stream Mach numbers M
s
. It is clear that the disagreement is very large due
to the experimental diculties with statistical convergence for the shock tube
results, and due to the dierences in Reynolds number compared to the DNS.
Other results have been obtained by Barre et al. (1996) at Mach 3 in a super-
sonic wind tunnel, where the turbulence amplication was measured by means
of a hot-wire. In Figure 31 the measured results compared to LIA are shown.
In the far eld agreement is quite good. However in the near eld there is
disagreement, probably due to the fact that the pressure uctuations, which
exist in the near eld turbulence, had to be omitted in the data reduction
of the hot-wire results because no measurements of the pressure uctuations
were available.
5
4
3
2
1
0
1.00 1.02 1.04 1.06 1.08 1.10 1.12
u
1
, 2x2 grid
u
1
, 3x3 grid
u
2
, 2x2 grid
u
2
, 3x3 grid
u
3
, 3x3 grid
u
3
, 2x2 grid
LIA, streamwise velocity
RDT, streamwise velocity
DNS, Lee et al (1993)
1.14 1.16 1.18 1.20
M
s
G
u
2
Figure 30: Velocity amplication ratio across a shock both from experiments,
LIA and RDT, and DNS. (From Andreopoulos et al. 2000): u
1
, longitudinal
velocity; u
2
, u
3
, lateral velocities.
There is only a limited amount of experimental data available for length
scale evolution across such shock interactions. Figure 32 shows results obtained
by Briassulis (1996) in a shock tube for dierent upstream Mach numbers and
dierent values of the upstream integral scale. Values of the longitudinal inte-
gral scale amplication rate G
L
1
(1) across the shock are plotted at dierent
positions (x/) where is the mesh spacing of the grid acting as the turbu-
lence generator. Results are given for three shock intensities (M
s
= 1.26, 1.37,
1.50), and at several integral scale initial values (recall that the initial integral
scale increases with x/). It is clear that in all cases, a decrease of the integral
scale is found, but it appears to depend strongly on both the shock strength
and the initial value of the upstream integral scale. At large x/, far from the
grid, the initial integral scales are the largest and the scale reduction across
the shock is much more important.
566 Barre et al.
20 10 0 10 20 30 40
6
5
4
3
2
1
0
X
c
(mm)
<u
0
>
2
<u>
2
LIA (near field)
LIA (far field)
Figure 31: Longitudinal turbulence amplication across a Mach 3 normal shock
(from Barre et al. 1996): X
c
corresponds to shock location.
100 90 80 70 60 50 40 30 20 10 0
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
x/
M
flow
=0.36, M
s
=1.26, M
r
=1.18
M
flow
=0.475, M
s
=1.37, M
r
=1.25
M
flow
=0.6, M
s
=1.5, M
r
=1.39
G
L
1
(1)
Figure 32: Ratio of the longitudinal integral length scale across the shock for
dierent upstream integral scale and Mach numbers (from Andreopoulos et al.
2000).
The attenuation of the longitudinal integral scale is also dependent on the
shock strength. For high Mach numbers the attenuation is larger for a xed
upstream scale. Figure 33 from Barre et al. (1996) conrms this fact. For
a Mach 3 shock, the attenuation is about 0.14 which is roughly equivalent
to the results obtained by Briassulis (1996) at Mach 1.5. Thus, it appears
that beyond Mach 1.5 an asymptotic behavior is obtained for the reduction
in longitudinal integral scale across the shock. However, it is clear that, at
the present time, no systematic study concerning the eect of the ratio be-
[19] Compressible, high speed ows 567
20 10 0 10 20 30 40
0.0
0.2
0.4
0.6
0.8
1.0
X
c
(mm)
Lf
Lf
0
Figure 33: Longitudinal integral scale attenuation across a Mach 3 normal
shock (from Barre et al. 1996): Lf, downstream shock value; Lf
0
, upstream
shock value.
tween the upstream integral scale and the shock thickness has been done.
Concerning the evolution of the lateral integral scale, Barre et al. (1996) have
shown experimentally that this scale is practically unaected by the shock
interaction.
A direct consequence of these results for the turbulent scales is the evo-
lution across a normal shock of the uctuating longitudinal mass ow rate
spectrum. Figure 34 shows the upstream and downstream hot-wire spectra
obtained across a Mach 3 normal shock by Barre et al. (1996). (In Figure 34
spectra are represented in dimensional form. Their integral is the local veloc-
ity variance, thus this representation takes into account the turbulent kinetic
energy amplication across the shock). It is clear that the downstream spec-
trum is shifted towards the high frequencies. This fact conrms the observed
reduction in longitudinal integral scale across the shock (cf. Figure 33).
The amplication ratio of each turbulent scale resolved by the experiment
can be obtained by computing the ratio of the two spectra. This ratio is
presented on Figure 35. It is clear that the most amplied frequencies are
higher than the ones corresponding to the incoming longitudinal integral scale.
The shock then performs an energy transfer from the incoming integral scale
towards smaller scales. This is conrmed by DNS data from Lee et al. (1993),
Hannappel and Friedrich (1995), and Jamme (1998).
Unfortunately, it has not been possible up to this point to reach any general
conclusions from the experimental studies on the following questions (among
others):
What is (are) the physical parameter(s) that govern(s) all aspects of this
interaction?
568 Barre et al.
Before Shock
After Shock
100
k (m
1
)
1000 10000
10
5
10
6
10
7
Figure 34: Upstream and downstream hot-wire signal spectra for a Mach 3
normal shock (from Barre et al. 1996).
0 500 1000 1500 2000 2500 3000
k (m
1
)
PSD
PSD
0
6.0
5.0
4.0
3.0
2.0
1.0
0.0
Figure 35: Ratio of hot-wire signal spectra across a Mach 3 normal shock (from
Barre et al. 1996): PSD, downstream shock power spectral density; PSD
0
,
upstream shock power spectral density.
How can the problem be parametrized? Should the shock thickness be
taken into account?
It is hoped that future experiments (some now in progress) can provide some
answers to these questions.
While progress has been slow, this does not diminish the importance of
understanding the dynamics in this simplied situation before attempting to
develop models for much more complex shock/turbulence interactions such as
shock/boundary-layer interactions.
[19] Compressible, high speed ows 569
4.2 Shock/boundary-layer interaction
It is only with recent improvements in computer hardware that direct numer-
ical simulations of supersonic wall-bounded ows with embedded shocks have
become feasible. An accurate DNS of such a spatially evolving ow eld is a
dicult task. Two obvious challenges are, rst, the large number of numer-
ical grid points required in a full spatial simulation and, second, the ability
to retain a high level of global numerical accuracy when shocks are present.
The issue of global numerical accuracy for shock capturing in such ows has
been the focus of some fundamental studies (Adams and Shari 1996, Casper
and Carpenter 1998). Casper and Carpenter analyzed the behavior of high-
order schemes for unsteady ows with shocks for aeroacoustic applications,
and concluded that in the presence of moving shocks it is dicult to achieve
globally high-order accuracy. Nevertheless, Adams and Shari developed a
hybrid scheme that uses a fourth order compact scheme in smooth regions
and a fourth order ENO (Essentially Non Oscillatory) scheme near shocks.
They showed that direct simulations of ows with steady shocks were feasible
provided a high-order scheme is used that satises certain shock capturing
properties. The methodology has been applied to the direct numerical simula-
tion of a compression ramp ow at low Reynolds number (Adams 1998). Data
relevant to turbulence modeling have yet to be extracted from this simulation,
but clearly there is a large potential for future work on this theme. However,
accurate computations at high Reynolds numbers comparable to experiments
will remain outside the range of DNS for many years to come.
The focus in this subsection will then be on the RANS predictions of super-
sonic boundary-layer ows with shock interactions. Several examples of such
ows have been extensively documented and discussed (Fernholz and Finley
1976, 1980, 1981; Fernholz et al. 1989; Settles and Dodson 1991; Knight and
Smits 1997). Since it is dicult to measure ow variables in close proxim-
ity to the wall, and to obtain highly accurate turbulence statistics in such
ows, detailed quantitative comparisons with RANS calculations are dicult.
Such comparisons mainly focus on distributions of mean variables across the
boundary layers as well as pressure and skin friction distributions along the
surface. Where available, comparisons with turbulence statistics provide a use-
ful, though qualitative, guide to model performance. Shock/boundary-layer
interactions can be classied on how the shock is generated (Smits and Dus-
sauge 1996). One such category is an incident shock interaction in which the
shock from an external source impinges on the layer. Another category is a
compression surface in which the shock is generated by the same surface that
generated the boundary layer.
Examples of the former category include the reection of an oblique shock
at a turbulent boundary layer (see Haidinger and Friedrich 1995, also Figure
20), and the shock eld generated by a double-n geometry. For the double-
n example, experimental results from Zheltovodov et al. (1994) have been
570 Barre et al.
used for comparison with computational results from turbulent models rang-
ing from an algebraic Baldwin-Lomax model to a full Reynolds stress model
(see Knight and Smits 1997, pp. 5.395.43). Comparisons with mean surface
quantities such as wall pressure and temperature, and surface heat transfer
have been made. While wall pressure variation has been successfully com-
puted by the models, wall temperature and surface heat transfer predictions
have not been as successful. As might be expected from such complex ow
elds, characterizing dynamic features such as shock/shock interactions as well
as shock/boundary-layer interactions make it dicult to assess and uniquely
identify model deciencies.
Examples of the compression surface category include shocks generated on
an airfoil, wing, bump, cone, (single) n, and ap/ramp surfaces (see Figure
20). A multitude of studies of such ow elds exist in the literature (e.g.
Haase et al. 1993, 1997; Leschziner et al. 1999, Loyau et al. 1998) utilizing
a variety of turbulence models ranging from simple algebraic models to full
second-moment closures. As noted previously, the range of experimental data
available for comparison with calculations may vary; however, wall pressure
variation (or pressure coecient) and skin-friction coecient are fundamental
and can provide important information about the general characteristics of
the ow eld as well as model performance. As an example, consider the ow
eld predictions of an RAE 2822 airfoil (Re = 6.2 10
6
, M
= 0.754 and
angle-of-attack 2.57
, 16
, 20
, 24
. Separation in the
vicinity of the at-plate/ramp juncture occurs at 16
(incipient), 20
, and 24
.
(Unsteady and three-dimensional eects are present in the 24
8 ramp
16 ramp
LRR
SSG
20 ramp
_
_
(a) (a) (a)
0.2 0.1 0.0 0.1 0.2
(ss
0
)/s
0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
C
f
x
1
0
3
Experiment
LRR
SSG
(b)
Figure 37: Streamwise distribution along surface of compression ramp: (a)
normalized pressure for 8
, 16
, and 20
ramp.
Another measure of the inuence of the at-plate/ramp juncture on the ow
eld can be seen from the streamwise evolution of the skin-friction coecient.
Figure 37b shows the distribution of the measured skin-friction coecient C
f
along the surface for the 16
ramp
case was examined by Abid et al. (1995). Figure 38 shows the evolution of the
normalized second-moment
ss
along a streamline downstream through the
shock. (The
ss
component is the streamwise component and the normaliza-
tion is with the value at the inow plane.) The initial streamline location was
y
+
2200 which is well away from any direct eects of the wall functions used.
Neither the LRR nor the SSG pressure-strain models were able to correctly
predict the location or amplication level of the experimental results.
These examples were intended to show that simple mean properties of the
ow, such as surface pressure and skin friction, can be used as a guide to the
[19] Compressible, high speed ows 573
0.1 0 0.1 0.2
s
0
2
4
6
8
s
s
/
s
s
0
Experiment
LRR model
SSG model
Figure 38: Evolution of
ss
second-moment along streamline in compression
ramp ow.
ow dynamics, and that even well established closure models straightforwardly
applied may not be able to replicate these mean properties. Extending such
comparisons to the turbulence quantities (such as shown in Figure 38) becomes
even more challenging and poor predictive performance may not be surprising.
While the predictions of the compression ramp ows are disappointing, sev-
eral other shock/boundary layer studies have been conducted with more suc-
cess. These include airfoils, blunt-ns, bumps (plane, axisymmetric, skewed),
and wings. Some examples are discussed in the recent papers by Leschziner and
colleagues (Batten et al. 1999, Leschziner et al. 1999, Loyau et al. 1998) as well
as reports on organized attempts at predicting a cross-section of aerodynamic
ows (Haase et al. 1993, 1997).
5 Concluding Remarks
The dynamics of compressible, turbulent ows, with or without shocks, is quite
complex. In this chapter some of the characterizing features of these ows
have been presented, as well as some of the methodologies used in the analysis
and prediction of such ows. These methodologies have included experiments,
numerical simulations, and RANS calculations.
From an analysis of the various experimental results from both shock tube
and supersonic wind tunnels, it appears that the amplication ratio of tur-
bulent kinetic energy as turbulence passes through a shock wave is strongly
574 Barre et al.
inuenced by the nature of the incoming turbulence. However, it is still dicult
to clearly parameterize the amplication ratio behavior. Eects of incoming
Mach number, turbulent Reynolds number, compressibility level, turbulence
intensity and integral scales have not been separated and quantied by the
available experiments up to this point. The same conclusions also hold for the
evolution of the turbulent scales and spectra across the shock.
There have been several direct numerical simulations of homogeneous tur-
bulence interacting with a shock wave during the last decade. These show that
turbulence kinetic energy amplication is consistent with the results of linear
theories. Longitudinal length scales are reduced across the shock, although
it is still unclear what the proper parameter or range of parameters are to
describe this (upstream integral length scale, upstream turbulence level, tur-
bulence Reynolds number, shock thickness etc.) Lateral scales are essentially
unaected by passage through the shock, while dissipation rate and dissipation
scales both increase.
For mainly shock-free turbulent ows progress has been made in parame-
terizing the eect of compressibility and sorting out the causes of the changes.
For attached boundary ows up to Mach 5 the strong Reynolds analogy holds
and the essential turbulence dynamics follows that of incompressible ow. The
gradient Mach number may be used to distinguish these ows (with low M
g
)
from mixing layer and jet ows where true compressibility eects become im-
portant at lower absolute Mach numbers. For mixing layers the gradient Mach
number is proportional to the convective Mach number, which has been found
to collapse a wide range of experimental data. Direct numerical simulations
are now showing that statistical terms such as the pressure strain are central
to explaining the reduction of mixing layer growth rate with Mach number.
Simulations are additionally providing data to form models of explicit com-
pressibility terms such as dilatation dissipation and pressure dilatation.
While simulation techniques (DNS and LES) are appealing on the grounds
that they make few assumptions about the turbulence, practical calculations
and predictions of complex compressible ow elds are usually done by RANS
calculations. As was outlined in this chapter, the mean conservation and tur-
bulent transport equations written in Favre variables provides an eective
framework for developing models for the higher-order correlations which ap-
pear in these equations. In the compressible case, new correlations directly
associated with the compressibility appear and these terms now require mod-
eling. In the absence of shocks, or even in the presence of shocks under certain
circumstances, the utilization of variable mean density extensions to the in-
compressible formulation is justied. When strong compressibility eects arise,
the modeling situation becomes less clear. As has been discussed, several ques-
tions about the underlying physics remain, and without this more complete
picture it will not be possible to develop an accurate array of closure models
for the compressible correlations. Currently, many RANS approaches revert
[19] Compressible, high speed ows 575
to the simple variable density extensions just mentioned, and most attempts
to include compressibility models are on a case-by-case basis. Clearly, more
focused eorts are required to develop a somewhat more general approach to
predicting these supersonic ows.
It is hoped that this chapter has provided the reader with a more complete
understanding of the broad scope of research that has been undertaken in an
eort to successfully predict and subsequently control supersonic turbulent
ows. Many challenges and opportunities remain in the study of the dynamics
of such ows.
Acknowledgements
The authors thank Professor P.G. Huang for providing the direct numerical
simulation data used in Figure 4, and Dr. S. Jamme for providing the LIA
results.
A Identities for Mass-Weighted Averages
f = f +
, = +
(A.1)
f =
f +f
= f +f
, f
= 0 (A.2)
f =
f +f
=
f +f
=
f +f
, f
= f
f (A.3)
f
= f
, f
(A.4)
f
= f
f +f
= f
+f
= f
(A.5)
f
= f
= 0,
(A.6)
f
=
f
(A.7)
f
=
_
f
__
g
_
(A.8)
f
= (
f
) = f
_ _
(A.9)
576 Barre et al.
References
Abid, R., Gatski, T.B., and Morrison, J.H. (1995). Assessment of pressure-strain
models for predicting compressible turbulent ramp ows, AIAA J. 33, 156159.
Adams, N.A. (1998). Direct numerical dimulation of turbulent compression ramp
ow, Theoret. Comput. Fluid Dynamics 12, 109129.
Adams, N.A., and Shari, K. (1996). A high-resolution hybrid compact-ENO scheme
for shock-turbulence interaction problems, J. Comput. Phys. 127, 2751.
Andreopoulos, J., and Muck, K.C., (1987). Some new aspect of the shock wave bound-
ary layer interaction in compression ramp corner, J. Fluid Mech., 180, 405428.
Andreopoulos, J., Agui, J.H., Briassulis, G.K. (2000). Shock wave-turbulence inter-
actions, Ann. Rev. Fluid Mech. 32, 309345.
Ardonceau, P. (1984). The structure of turbulence in a supersonic shock wave/bound-
ary layer interaction, AIAA J. 22, 12541262.
Bardina, J.E., Huang, P.G., and Coakley, T.J. (1997). Turbulence modeling valida-
tion, testing, and development, NASA TM 110446.
Barre S. (1993). Action de la compressibilite sur la structure des couches de melange
turbulentes supersoniques. Th`ese de Doctorat, E.N.S.A.E. Toulouse (France) jan-
vier.
Barre, S., Alem, D., and Bonnet, J.P. (1996). Experimental study of a normal shock/
homogeneous turbulence interaction, AIAA J. 34, 968974.
Barre, S., Dupont, P., Dussauge, J.P. (1992). Hot-wire measurements in turbulent
transonic ows, Eur. J. Mech. B 11, 439454.
Barre S., Quine, C., and Dussauge, J.P. (1994). Compressibility eects on the struc-
ture of supersonic mixing layers: experimental results, J. Fluid Mech. 259, 4778.
Batten, P., Leschziner, M.A., and Craft, T.J. (1999). Reynolds-stress modeling of
afterbody ows. In Turbulence and Shear Flow Phenomena 1 S. Bannerjee and
J.K. Eaton (eds.), Begell House, 215220.
Bellaud, S. (1999). Mesures et analyses detaillees des champs turbulents en couches
de melange annulaire supersoniques. Th`ese de Doctorat Universite de Poitiers.
Blin, E. (1993). Etude experimentale de linteraction entre une turbulence libre et une
onde de choc. Th`ese de doctorat de lUniversite Paris VI.
Bonnet, J.P., Chambres, O., Lammari, M., Barre, S., Braud, P. (1996). Mesures 3D
en couches de meange planes supersoniques, Contrat DRET 94/2586 A., Octobre.
Bonnet, J.P., Gresillon, D., Taran, J.P. (1998). Nonintrusive measurements for high-
speed, supersonic and hypersonic ows, Ann. Rev. Fluid Mech. 30, 231273.
Bradshaw, P. (1977). Compressible Turbulent Shear Layers, Ann. Rev. Fluid Mech.
9, 3354.
Bradshaw, P., Ferriss D.H., and Johnson, R.F. (1964). Turbulence in the noise-
producing region of a circular jet, J. Fluid Mech. 19, 591624.
Briassulis, G.K. (1996). Unsteady nonlinear interactions of turbulence with shock
waves. PhD Thesis, City College, City University of New York.
[19] Compressible, high speed ows 577
Briassulis, G.K., and Andreopoulos, J. (1996). High resolution measurements of
isotropic turbulence interacting with shock waves, AIAA 34th Aerospace Sciences
Meeting, Paper 96-0042.
Briassulis, G.K., Agui, J.H., and Andreopoulos, J. (1999). The structure of weakly
compressible grid generated turbulence, J. Fluid Mech., submitted.
Breidenthal R. (1990). The sonic eddy: a model of compressible turbulence, AIAA
28th Aerospace Sciences Meeting, Paper 90-0495.
Brown, G.L., and Roshko, A. (1974). On density eects and large structure in tur-
bulent mixing layers, J. Fluid Mech. 64, 775816.
Casper, J., and Carpenter, M.H. (1998). Computational considerations for the sim-
ulation of shock-induced sound, SIAM J. Sci. Comput. 19, 813828.
Chambres, O. (1997). Analyse experimentale de la modelisation de la turbulence en
couche de melange supersonique. Th`ese de Doctorat Universite de Poitiers.
Chang, C.T. (1957). Interaction of a plane shock and oblique plane disturbances with
special reference to entropy waves, J. Aero. Sci. 24, 675682.
Clemens, N.T. and Mungal, M.G. 1995 Large scale structure and entrainment in the
supersonic mixing layer, J. Fluid Mech. 284, 171-216.
Coleman, G.N., Kim, J., and Moser, R. (1995) A numerical study of turbulent su-
personic isothermal-wall channel ow, J. Fluid Mech. 305, 159183.
Coleman, G.N., and Mansour, N.N. (1991). Modeling the rapid spherical compression
of isotropic turbulence, Phys. Fluids A 3, 22552259.
Cooke, P., McDonald, M., and Firmin, M. (1979). Airfoil RAE2822 pressure dis-
tributions and boundary layer wake measurements, AGARD AR-138.
Debi`eve, J.F., and Lacharme, J.P. (1986). A shock wave/free turbulence interaction.
In Turbulent Layer/Shock Wave Interactions, J. D`elery (ed.), Springer.
Debisschop, J.R. (1993). Comportement de la turbulence en couches de melange su-
personiques. Th`ese e Doctorat Universte de Poitiers.
Denis S., Delville J., Garem J.H., Barre S., and Bonnet J.P. (1998). Etude du contr ole
des couches de melange planes et axisymetriques-(partie subsonique), Rapport nal
contrat CNRS 780-441 (commande Dassault) decembre.
Dolling, D.S., and Murphy, M. (1982). Wall pressure in a supersonic separated
compression ramp oweld, AIAA and ASME 3rd Joint Thermophysics, Fluids,
Plasma, Heat Transfer Conference, Paper 82-0986.
El Baz, A.M., and Launder, B.E. (1993). Second-moment modelling of compressible
mixing layers. In Engineering Turbulence Modelling and Experiments 2 W. Rodi
and F. Martelli (eds.), Elsevier, 6372.
Elliott, G.S., and Samimy, M. (1990). Compressibility eects in free shear layers,
Phys. Fluids A 2, 12311240.
Favre, A. (1965).
(x, t)
+d
for = 1, . . . , N. Clearly
p(, ) is a multi-dimensional quantity and it represents the PDF for a single
realization of a turbulent ow, so that if
(x, t)
+ d
is not
satised for all , then p(, ) = 0, otherwise p(, ) = 1. Consequently, the
ne-grained density can be expressed as
p(, )
N
=1
(x, t)
. (1)
In contrast, the conventional PDF P() is representative of an ensemble
of realizations, such that its integral over a nite region of its sample space
represents the likelihood of occurrence for an event. It is a straightforward
matter to show, see Pope (1985), that
P() = p(, )). (2)
The details of the derivation of the evolution equation for P() can be found in
OBrien (1980) and Kollmann (1992). Kollmann (1989, 1990) derives the PDF
equation in terms of the characteristic function, which is the Fourier transform
of the PDF. Whatever the method of derivation the evolution equation for the
[20] The joint scalar probability density function 585
density weighted scalar PDF) is given by
P()
t
. .
I
+ u
k
P()
x
k
. .
II
N
=1
()
P()
_
. .
III
=
x
k
_
u
k
[ = )
P()
_
. .
IV
+
N
=1
__
J
i,
x
i
=
_
P()
_
. .
V
(3)
where
and J
i,
represent the net formation rate per unit volume and molec-
ular diusive ux, respectively, of species and where indicates a density
weighted quantity.
The terms have the following signicance:
Term I: Rate of change in physical space;
Term II: Mean convection in physical space;
Term III: Chemical source production in compositional space;
Term IV: Turbulent transport in physical space;
Term V: Micromixing in compositional space.
Assuming Fickian diusion the micromixing term can be expressed as
=1
N
=1
__
Sc
x
k
x
k
=
_
P()
_
, (4)
where Sc is the Schmidt number.
With the assumption of constant it can also be shown, see Kollmann
(1992), that the term can be expressed in terms of a two-point PDF
Sc
N
=1
lim
x
2
x
k
x
k
__
P(,
; x, x
, t)d
_
, (5)
where the primes denote terms based on conditions at x
.
586 Jones
2.1 Terms Appearing in the PDF Equation
Inspection of equation (3) reveals that the chemical source production term
appears in exact and closed form. In other words it is known in terms of
and
P(). This feature of equation (3) has always constituted the main reason
for using the PDF approach in reacting ows. The implication is that realis-
tic chemistry involving nite-rate kinetics can be incorporated explicitly in
turbulent ow calculations without an assumed ame structure to character-
ize the thermochemistry. In comparison with other forms of thermochemical
closure, this is clearly a novel, advantageous and unique feature. However,
the micromixing term appears in unclosed form due to the appearance of the
conditional expectation. To evaluate this term, the joint PDF of the scalars
and their gradient (equation (4)) at the single-point level, or the two-point
joint PDF of the scalars (equation (5)) is required. Needless to say, evolu-
tion equations for such higher-order PDFs contain further unknown terms
and dramatically increase the dimensionality and/or computational cost of
the problem. The micromixing term involves diusion coecients and com-
position gradients, which stem directly from the molecular diusion term in
the instantaneous species conservation equations, and an order of magnitude
argument demonstrates that it cannot be neglected, even at high Re. Since
diusion is important predominantly amongst the small scales of turbulent
motion, an accurate representation of these scales is clearly essential to the
determination of the evolution of the PDF and hence of the averaged scalar
eld. In hydrocarbon combustion, this is particularly true for species such H
2
and H which are formed and consumed by reaction (in the ne scales of tur-
bulent motion) and which possess much larger diusivities than other scalars.
As a consequence it can be argued that the classical closure problem asso-
ciated with turbulent combustion, which manifests itself as a mean reaction
rate in moment closures, is simply re-expressed in terms of the micromixing
term in PDF closures. In fact the appearance of the chemical source term
in exact form in the PDF formalism is merely a consequence of the mathe-
matical manipulations employed to derive such an equation. As pointed out
by Kollmann (1992), in PDF methods nonlinear terms are transformed into
linear terms with variable coecients. This is achieved by converting the as-
sociated dependent variables (which make the nonlinear terms unclosed upon
averaging) into independent variables of the PDF at the cost of increasing
the dimensionality of the formalism. The results of such a manipulation do
not overcome the basic physical problems inherent in any statistical approach
to turbulence, whether it be ReynoldsFavre averaging, spatial averaging or
transported PDF modelling. The PDF
P() is only a function of the scalar
eld and includes no velocity information; consequently the turbulent diu-
sion of the PDF (term V in equation (3)) appears in unclosed form. The joint
velocity-composition PDF allows this term to be expressed in exact form, but
at the expense of introducing further unknowns and extra independent vari-
[20] The joint scalar probability density function 587
ables to the equation. Finally it is to be noted that, at the level of closure
corresponding to equation (3), u
k
is unknown. This quantity is conventionally
supplied through the use of a turbulence model.
3 Solution Aspects of PDF Evolution Equations
At this stage it may appear somewhat premature to discuss the solution al-
gorithm for the PDF evolution equation, particularly since the modelling of
the unclosed terms has not yet been considered. However, the reality is that
the method of numerical solution and the modelling under this formalism are
strongly interconnected. For problems involving a single scalar dimension and
in simple geometries, solutions of PDF equations using conventional nite-
dierence techniques have been obtained (Janicka et al. 1978, 1979). How-
ever, in order to exploit the full potential of PDF equations, it is necessary
to solve for a multi-dimensional PDF involving several non-conserved scalars.
Pope (1981) estimated that the number of computer operations at each node
and at each time step, if nite-dierence equations were to be employed, was
of the order of exp(6N), where N is the number of scalar variables. In the
same paper he described a solution algorithm for performing a Monte Carlo
(particle-method) simulation of the scalar PDF equation. In this approach
the computational cost rises only linearly with N and the PDF is simulated
via an ensemble of stochastic particles. That a continuous PDF
P() has a
discrete counterpart is a widely used principle. The histogram constructed
from
(i)
(x, t), (where
(i)
represents the ith realisation or measurement) is
an approximate discrete representation of the PDF. In the limit of an in-
nite number of realizations (with the sampling bin sizes tending to zero), the
histogram tends to the PDF this result is formalized by equation (2).
With regard to the closures for the mixing term in equations (3) and (4),
which we will discuss later, the implication is that, given a suitable approxima-
tion (written in PDF form), an equivalent particle model representation needs
to be constructed. Thus in the section dealing with the mixing models, both
interpretations will be given, usually for the case of a single scalar. For certain
mixing models the extension to multi-scalar problems is simple; in other mod-
els it is still not clear how this can be achieved in a rigorous and/or tractable
manner. However, the existence of a link between the particle method and the
PDF approach is essential if the advantages of the latter method are to be
exploited. Such a link and the necessary criteria for its use have been investi-
gated by Pope (1979, 1981, 1985); it is argued that the fundamental basis for
equivalence of these two methodologies is that the resulting statistical mo-
ments (mean, variance, . . . ) arising from the particle-method approach should
evolve in an identical manner to its deterministic counterpart in the limit of
an innite number of notional particles.
588 Jones
4 Basic Modelling Requirements of the Molecular
Mixing Term
Closure approximations constructed to represent the micromixing term must
satisfy a number of properties of term V in equation (3), requirements which
are discussed in Kakhi (1994) where various mixing models are also reviewed.
The most basic requirements are that the approximation leaves the means
of the scalars unchanged, i.e. (x, t)) (x, t + t) (x, t)) = 0, and
that the model gives rise to PDFs which correspond to a set of physically
possible sequences of events. In addition to the general realizability condition,
P(
, x, t) = 0 for
,= 1 or
Y
< 0. (6)
While the satisfaction of these requirements by any mixing model is essential if
anything useful is to result, they are not exhaustive and there are many other
conditions which could be considered and imposed. These include for example
the linearity and independence properties of a set of passive scalar quantities,
see Pope (1983), and the localness concepts introduced by Subramanian and
Pope (1998). However these have not so far found widespread use and will not
be considered in detail here. Instead we turn to the constraints imposed on
mixing models by the need to produce results in agreement with measurements
in simple congurations involving passive scalar quantities.
Of these requirements probably the most important are the scalar variance
decay rates, i.e. the scalar dissipation rates, predicted by the model. All mixing
models proposed in current use have utilised the assumption that the ratio of
the scalar to mechanical turbulence integral time scale is constant so that
D
Dt
2
) = C
d
2
), (7)
where
2
C
d
is the time scale ratio, r, and where k/ is used here to represent
the integral time scale. For inert ows and in situations where the mechanisms
responsible for generating the velocity and scalar uctuations have roughly the
same length scales, for example in thin shear layers where the generation and
dissipation rates of turbulence energy and scalar variance are in approximate
balance, then the constant C
d
has a value around 2.0. While the assumption
[20] The joint scalar probability density function 589
has been very widely used in the context of RANS closures and gives rea-
sonable results in many near equilibrium shear layer ows, it is not of general
applicability since the time scale ratio will be dierent in dierent ows: values
in the range
1
3
< r < 3 have been observed, see Warhaft and Lumley (1978).
The real test of any mixing model lies in its ability to predict correctly the
evolution of the scalar PDF and in discussing the statistical behaviour of the
various models, reference will be made to the following parameters: skewness,
atness and superskewness. These are the normalized moments of the PDF,
dened as follows:
n
=
_
( ))
n
P()d
_
_
( ))
2
P()d
_n
2
_
_
n = 3 skewness
n = 4 atness
n = 6 superskewness.
(8)
For a reference Gaussian PDF all the odd moments are zero, because of sym-
metry, and the values of the normalized fourth and sixth moments are 3 and
15 respectively.
A popular test case utilised by many workers is that based on isotropic ho-
mogeneous turbulence with initial conditions corresponding to segregated uid
parcels (in composition space), with a PDF idealized as a double -function dis-
tribution. Here it is to be expected that the process of mixing/homogenization
will result in the initial PDF relaxing to a smooth, bell-shaped distribution
and, in the limit of t , converging to a -function at the mean composition.
Though there is no formal proof that it should be so, it is conventional to as-
sume that the bell-shaped distribution is Gaussian. Eswaran and Pope (1988)
performed a DNS of decaying passive scalar elds in inert, homogeneous and
isotropic turbulence. A statistically stationary velocity eld was maintained by
forcing the simulations i.e. energy was added to the velocity eld at low wave
numbers, allowing the simulations to reach a quasi-equilibrium state where
the rate at which energy was dissipated at the small scales was equal to the
rate at which energy was added at the large scales. The results demonstrated
that, as time progressed, P(; t) evolved from a double -function distribution
to an inverted parabola and subsequently to a bell-shaped distribution. The
evolution of the PDF shapes appeared to be independent of initial conditions
(such as the ratio of the mechanical and scalar turbulence length scales, which
had a signicant eect on the scalar variance decay rate). It was argued that
this observation should considerably simplify the task of constructing a realis-
tic mixing model and it was found that after about eight eddy turnover times
(a relatively long period of time), the skewness, atness and superskewness of
tended very closely to the Gaussian values of 0, 3 and 15 respectively. The
tendency to Gaussianity was accompanied by an ever increasing lack of correla-
tion between the scalar dissipation rate and (x, t). This is an important
result since the micromixing term (equation (4)) is eectively conditioned
590 Jones
on (x, t). As a modelling recommendation, it is argued that the assumption
of being independent of (x, t), although valid at long times, is not valid in
the short-term range of mixing: this raises questions concerning the validity of
the statistical independence assumption which is often invoked in combustion
and turbulence modelling. Additional evidence of the tendency to Gaussian
PDFs was provided by Tavoularis and Corrsin (1981) who measured the evo-
lution of the temperature eld in a nearly homogeneous turbulent ow with
uniform mean velocity and mean temperature gradients. The overheat of the
uid by heating rods was considered small enough to have negligible eect on
the turbulence, and hence the temperature can be viewed as a passive scalar.
It was found that typical values of the skewness and atness factors of the
temperature measured on the centre-line of the tunnel, far downstream, were
indistinguishable from the values for a normally distributed random variable
(which are 0 and 3, respectively); it was found that a Gaussian with the same
mean and variance tted the PDFs nearly perfectly. The velocity-scalar joint
PDF also closely resembled a jointly-normal distribution.
Despite the above comments concerning the evolution of a decaying scalar
eld to Gaussianity, it must be borne in mind that the scalars considered here
are bounded quantities, e.g. species mass fractions, and so their PDFs are
bounded and cannot be exactly Gaussian. The implication is that the PDF
approaches a Gaussian distribution only in the limit of zero variance where
most of the interacting uid parcels are very close to the mean composition.
However the distribution near the bounds of the scalar are unlikely to resem-
ble Gaussian behaviour, but in the cases considered above this is immaterial
given the degree of homogenization at that stage of the mixing process. The
provision of a suitable model for micromixing is a central requirement in the
PDF evolution equation approach. A number of proposals have been made and
these are described and reviewed below. In discussing mixing models atten-
tion will be restricted to homogeneous turbulent ows without reaction; this
is usually the basis on which the models are derived. However in all cases the
models are intended for application, without modication, to inhomogeneous
turbulent ows both with and without reaction.
4.1 Linear Mean Square Estimation (LMSE) closure
Using equation (5) and the denition of conditional probability, P(,
) =
P(
P(; t)
t
=
Sc
N
=1
lim
x
2
x
k
x
k
__
P(
[ = )d
P()
_
=
Sc
N
=1
lim
x
2
x
k
x
k
_
E(
[ = )P()
. (9)
[20] The joint scalar probability density function 591
Dopazo (1973, 1979) and Dopazo and OBrien (1974) considered the case of
the single scalar, where E(
, t) at a point x
[ = ) = ) +r(x, x
, t) ( )) , (10)
where ) is the expectation (mean) of the random variable (x, t) and
r(x, x
, t):
r(x, x
, t) =
_
(x, t) (x, t))
_
(x
, t) (x
, t))
_
(x, t)
(x
, t)
, (11)
where
, t), given (x, t), to a certain prescribed behaviour. Later OBrien (1980)
reformulated the problem in terms of linear mean square estimation, which
involves less restrictive assumptions but gives rise to the same result. The
requirement of homogeneity and isotropy together with equations (10) and
(11) leads to the following result for the RHS of equation (9):
3
_
2
r
2
_
=0
[ (x, t))]
6
2
r/
2
_
1
=0
, where = [x
2
)
dt
= 2A
2
). (14)
Comparing this result with a conventionally modelled scalar variance equation
in homogeneous, isotropic and inert turbulence implies the identity
2
). (15)
Referring to equation (13) as the LMSE model, the particle method simulation
corresponding to this closure becomes
d
(p)
dt
=
C
d
2
k
_
(p)
)
_
, (16)
592 Jones
where A has been replaced with C
d
k
consistent with equation (7). The model
constitutes a linear, deterministic simulation of the mixing phenomenon, con-
tinuous in time. Analytical solutions for equation (13) are discussed in Dopazo
and OBrien (1976), OBrien (1980) and Pope (1985); the results of such anal-
ysis show that if the initial condition is a normal distribution, the model pre-
serves the Gaussianity which is considered as a satisfactory result. However,
the drawback is that the initial functional form of the PDF is never relaxed
and consequently, in the case of an initial double -function specication for
the PDF, the subsequent time evolution remains as two -functions which
approach each other in composition space. Pope (1985) points out that such
behaviour is to be expected since the model contains no information about the
shape of the distribution only the mean appears in equation (16). Despite
this drawback, Dopazo (1979) argues that the modelling is applicable provided
the turbulence is given a chance to induce randomness and smear out the
initial discontinuous state. In other words it is applicable if the PDF is con-
tinuous. Predictions based on LMSE were compared against measurements in
Dopazo (1979) for grid-generated homogeneous turbulence with an asymmet-
ric mean temperature prole. In the far downstream region good agreement
was found for the conditional expected value of temperature versus tempera-
ture.
4.2 Coalescence-Dispersion Mixing Models
4.2.1 Curls Mixing Model
The work of Curl (1963) was concerned with the analysis of the mixing of
clouds of droplets in a two-liquid phase chemical reactor. It was assumed that
all the droplets in the system were of the same size and that coalescence took
place between two droplets having a solute concentration of c
1
and c
2
. Redis-
persion occurred immediately to produce two equal droplets of concentration
1
2
(c
1
+c
2
). The total concentration range was divided into L intervals of con-
centration width c and P(l)c was dened as the fraction of the droplets with
concentrations within the lth, 1 l L), interval. The basic balance equation
constructed for P(c) stated that the rate of change of the number of droplets
in a particular interval is inuenced by:
the rate of creation of new droplets in the interval (c, c + c) due to
collisions of droplets from concentration intervals located symmetrically
about c;
the rate of loss of droplets by collision of droplets from the interval
(c, c +c) with those from all other concentration intervals.
[20] The joint scalar probability density function 593
This led to
P(c)
t
= 4Q
_
_
_
c
_
0
P(c +)P(c )d P(c)
_
_
_
. (17)
Curl dened Q as twice the collision rate divided by the total number of
droplets in the system and solved equation (17) for the case of inert and react-
ing ows using conventional nite-dierence techniques. Subsequently Spiel-
man and Levenspiel (1965) proposed a particle (Monte Carlo) method for the
solution of such an equation and used the methodology to investigate the ow
characteristics of an ideal stirred tank reactor involving a dispersed phase sys-
tem. Curls mixing closure has also been used in the context of continuous
systems, to describe the turbulent mixing of gases. The work of Flagan and
Appleton (1974) and the references cited therein may be consulted as examples
of such applications.
As noted by Dopazo (1979) the direct extrapolation of Curls model to
the mixing of a single-phase gaseous scalar eld is questionable, particularly
since it was originally formulated to describe the interactions of droplets. An
implication would be that two uid parcels at dierent temperatures, say,
would mix instantaneously to form uid parcels with a temperature equal to
the average of the two original temperatures. This feature of the model is
clearly unrealistic and has also been highlighted by Pope (1982) for stochastic
simulations; considering the test situation discussed in section 4, an initial
condition corresponding to a double -function was specied and the mixture
allowed to evolve. Of the N
p
particles in the system, N
p
/2 were assigned
(p)
=
1 and the remainder as
(p)
= 1. In any particular time step, pairs of particles
were selected, and the number chosen was the nearest integer to
N
c
= Bt
k
N
p
. (18)
This expression was chosen to ensure a variance decay rate consistent with
equation (7) and, for Curls model, B 2. The particles were then allowed
to interact along the straight line connecting them in composition space, such
that their new concentrations jump from their initial values to their joint
mean value. The values of particles not selected remain unaltered. The results
demonstrated that, after the rst time step, took the values 1, 0 and 1;
after the second step they were 1, 1/2, 0, 1/2 and 1; after the jth step they
were all integer multiples of 2
1j
. Therefore a continuous distribution is never
achieved since certain values of the scalar sample space (those not correspond-
ing to multiples of 2
1j
) cannot be reached. In spite of this, after a sucient
period of time, a bell-shaped (but non-Gaussian and discrete) distribution is
obtained. In contrast to the LMSE closure however, the PDF does relax from
its initial shape.
594 Jones
4.2.2 Modied Curl
Dopazo (1979) and Janicka et al. (1979) independently proposed modica-
tions to Curls model in order to avoid some of its shortcomings. Although
the starting point motivating their modelling was dierent, the end results
were very similar indeed, largely because their approach followed the ideas be-
hind the derivation of Curls model. Essentially the dierence between Curls
original model and the modied forms lies in the treatment of the disper-
sion of the material point properties. In discussing these dierences it is use-
ful to introduce the concept of the transition PDF. The transition PDF is
dened such that T
r
(, [)d is the probability that the interaction of a
material point = with a point = produces + d and
( + ) ( + ) + d. It is clear from the denition that the
transition PDF has to satisfy the following properties:
In coalescence-dispersion models the mixing process is eectively rep-
resented by the pairwise interaction of material points. These material
points are only allowed to mix along the line joining them in composition
space and this is consistent with the notion of physical mixing. There-
fore, if is outside the interval (, ), it follows that T
r
(, [)d=0.
This implies that the interval (, ) denotes a locally convex domain for
.
We require T
r
(, [) = T
r
(, [+). This eectively states that
after pairwise interaction, the new scalar properties and +
are formed with the same likelihood. This condition, together with the
previous requirement, ensures that the mean of the dependent scalar, ,
remains unchanged by the micromixing process.
Necessarily T
r
(, [) is a PDF with respect to . Therefore the normal-
ization condition follows; when the integration is performed over the do-
main of consistent with the comments made above
_
T
r
(, [)d = 1.
In terms of the transition PDF, coalescence-dispersion models can be ex-
pressed in the general form (using a density-weighted PDF)
P()
t
B
k
_
_
_
_
d
_
d
P()
P()T
r
(, [)
P()
_
_
_
, (19)
where B is a constant which can be evaluated by multiplying equation (19)
by ( ))
2
, integrating over space, and then comparing the result with
equation (7). In terms of Curls original specication, where material points/
particles coalesce and disperse with properties involving their mean value, the
transition PDF takes the following form:
T
r
(, [) =
_
1
2
( + )
. (20)
[20] The joint scalar probability density function 595
In modied forms of Curls model the material points that interact are not
constrained to mix to their common mean values. In fact the degree of mixing
can vary between none and complete (up to the mean composition). If it is
assumed that all composition states between these limits can be attained with
equal likelihood then the transition PDF will then be a uniform distribution,
see Janicka et al. (1979), over the range of the locally convex domain (, ):
T
r
(, [) =
_
_
_
1
||
[, ]
0 , [, ].
(21)
In this case discontinuous PDFs cannot arise (in the sense that Curls original
model produces) since all possible composition states can be reached. How-
ever it must be noted that mixing is still represented as a jump process and
such closures are essentially discontinuous representations of the mixing phe-
nomenon. Also, when applied to the isotropic homogeneous turbulence test
problem with initial conditions corresponding to a PDF comprising a double
-function distribution, the resulting atness (equation (8)) and higher even
moments of the resulting PDF increase indenitely with time, whilst for a
Gaussian PDF these values are constant.
In the particle method analogy of the generalised coalescence-dispersion
model, a random selection of a pair of particles
(p)
(t) and
(q)
(t) is followed
by mixing such that the new concentrations are given by
(p)
(t +t) =
(p)
(t) +
1
2
x
_
(q)
(t)
(q)
(t)
_
(q)
(t +t) =
(q)
(t)
1
2
x
_
(q)
(t)
(q)
(t)
_
, (22)
where x is a uniformly distributed random number and where the number of
particles selected for mixing is given by equation (18).
There have been a number of studies aimed at deriving improved forms
of coalescence-dispersion models. Various alternative transition probabilities
have been examined, for example by Pope (1982), Kosaly (1986) and Kosaly
and Givi (1987). Pope (1982) formulated an age-biased model such that con-
tinuous PDFs with nite atness and superskewness could be obtained. This
was subsequently investigated by Chen and Kollmann (1991). Attempts have
also been made to overcome the limitations arising from the discontinuous
nature of coalescence-dispersion models and these include the reaction zone
conditioning approach of Chen and Kollmann and the ordered pairing method
of Norris and Pope (1991). However no signicant advantage appears to arise
from these various approaches and so they will not be considered further here.
Further details may be obtained from the original papers and from the review
of Jones and Kakhi (1996).
596 Jones
4.3 Langevin Models
Nonlinear integral models (such as coalescence-dispersion models) can be in-
terpreted stochastically as Poisson processes which form the basis of particle
interaction (jump) models, see Pope (1985). Random diusion processes can
also be used to simulate the behaviour of micromixing, and these form the
basis of Langevin models. Pope (1985) proposed a model for the mixing term
using a Langevin equation:
(t) = (t +t) (t) = g [(t) (t))]
t
+
_
b
2
)
_
1
2
W
t
. (23)
The rst term on the RHS is a linear deterministic term followed by a
random diusive term, g and b are constants to be determined, and B is
a non-negative diusion coecient for this process. The term is a char-
acteristic turbulent time scale and W
t
is the increment of a Wiener pro-
cess which is dened by the properties W
t
) = 0 and W
2
t
) = t where
W
t
= W
t+t
W
t
. A function which satises these properties is W
t
(t)
1/2
N
s
i=1
i
, where
i
, i = 1, . . . , N
s
, represents a set of independent stan-
dardized (i.e. zero mean, unit variance) normal random variables. The func-
tion W
t
is a continuous and non-dierentiable function of time. The mean
is preserved in this formulation, and the requirement for the correct vari-
ance decay rate leads to a relation between g and b leaving one constant to
be specied. Pope points out that from any initial condition and in the ab-
sence of any disturbing inuences, the PDF relaxes to Gaussian with this
model. The rate at which the PDF relaxes depends on the value of b: for
b = 0, LMSE is recovered for which there is no relaxation. A drawback of
the model is that the term involving b can lead to the violation of scalar
bounds and, for this reason, the expression shown in equation (23) is not
considered as useful for representing scalar mixing. However, for the veloc-
ity eld which is not bounded, such a Langevin model can be and has been
employed.
4.4 Binomial Models
Vali no and Dopazo (1990) proposed a sequential LMSE and random jump pro-
cess model capable of producing a Gaussian relaxation for bounded scalars.
During a time step t, two additive subprocesses take place involving
(n)
(t):
rst all N
p
particles modify their values according to an LMSE variation and
then N
p
B
is a positive di-
mensionless constant whose value strongly aects the relaxation rate of the
PDF and its value requires calibration against measurements). The remain-
ing N
p
(1 B
t/) particles do not alter their values, whilst those that are
[20] The joint scalar probability density function 597
randomly sampled modify their values according to
(p)
(t +t) (t)) =
(p)
( ))
2
)
1
2
, (24)
where
(p)
is either a standardized Gaussian random variable if is an un-
bounded scalar, or a standardized binomially distributed random variable if
is bounded. Since the latter case is relevant to reactive scalars, we will con-
centrate on its representation:
(p)
=
MT
_
MT(1 T)
. (25)
Here is a binomially distributed integer random variable with mean MT and
variance MT(1T), T is a real number in [0, 1] and can take integer values
between 0 and M. To ensure boundedness, the minimum and maximum values
of correspond to equal to zero and M, respectively. With this requirement
it can be shown that
M =
()
min
) (
max
))
2
)
(26)
T =
()
min
)
(
max
min
)
(27)
In homogeneous turbulence
2
) decreases with time and thus, for xed
bounds, equation (26) implies that M increases. Furthermore as M , the
de MoivreLaplace theorem shows that
(p)
tends to a normal continuous ran-
dom variable; thus a distribution close to Gaussian at long times is implied.
For the test case corresponding to an initial double -function with
min
= 0
and
max
= 1, diculties arise in the early stages of mixing with regard to
equation (26). Vali no and Dopazo show that, if the result of equation (26) is
truncated to the nearest integer (M is an integer), no relaxation of the PDF
is observed until M 2. This diculty was overcome by allowing for ran-
dom samplings of stochastic particles around each of the initial -functions.
As
2
) decreases, these two samplings converge towards that originally pro-
posed with a single M and T specication, given by equations (26) and (27).
Monte Carlo simulations of homogeneous turbulence demonstrated that the
moments of the PDF evolved correctly: i.e., to within statistical error (based
on nite sample size), the mean and skewness were invariant, the variance de-
cayed exponentially and the atness tended asymptotically to 3.0 the Gaus-
sian value. However, the rates at which the moments evolved were strongly
aected by the value of B
; for B
(t)
_
2
_
t
1
2
2
)
_1
2
(p)
. .
Random diusion term
_
1 +K
_
1
2
)
2
__
t
((t) )) ,
. .
Linear deterministic term
(28)
where
2
is equal to
2
M
whenever
(p)
(t) > 0, and is equal to
2
m
for
(p)
(t) < 0; the terms
2
M
and
2
m
are the maximum (positive) and minimum
(negative) uctuations allowed, respectively, K is a constant to be adjusted
to yield the correct evolution of the PDF shape and to be consistent with the
scalar variance decay rate, equation (7). It is worth noting that the binomial
model described in Vali no and Dopazo (1990) is sequential but is neither con-
tinuous nor dierentiable; the present (non-sequential) model is continuous
but non-dierentiable, and is therefore a more physically realistic representa-
tion of mixing. Equation (28) is similar to the proposal of Pope (1985), namely
equation (23), except that the Wiener process (which may generate unbounded
scalar values) is now replaced by a standardized binomially distributed random
variable,
(p)
, and dened to ensure boundedness as in the previous binomial
model. As time increases
2
) decreases, so that
2
)/
2
0,
(t)/
0
and for large values of t, equation (28) asymptotically yields relaxation to
a Gaussian PDF. Comparison of the PDF forms (using K = 2.1) with the
DNS data of Eswaran and Pope (1988) demonstrated very good agreement.
However, the sensitivity of the results to the value of K was not discussed.
Binomial models have been demonstrated to work well for PDFs of sin-
gle scalars with xed bounds. However, the ability of the models to handle
multi-scalar PDFs is not clear. The bounds of any particular reactive scalar
are not xed but depend on the values of the other scalars; each scalar prop-
erty, e.g. the specic mole number of H
2
, CO, . . . , has its own bounds xed
by conservation principles and the allowable scalar domain is a complicated
hypervolume in multi-dimensional composition space. Consequently the exten-
sion of binomial models to combustion applications, where a number of species
and temperatures have to be included, is not straightforward. Ensuring sat-
isfaction of the realizability condition, equation (6), while maintaining mean
values unchanged by mixing appears problematic. However some progress in
this direction has been made by Hulek and Lindstedt (1996).
[20] The joint scalar probability density function 599
4.5 Mapping Closures
In the literature dealing with mixing models, the quest for predicting Gaussian
PDFs in homogeneous turbulence appears to have culminated in the proposal
for mapping closures. This closure has received a considerable amount of at-
tention over a relatively short period of time, partly because of its ability to
predict (in analytical form in certain cases) the correct evolution and asymp-
totic behaviour of passive scalar mixing in homogeneous elds. Chen et al.
(1989) addressed the problem of closure for the mixing term by constructing
(x, t) as a time-dependent mapping of a Gaussian reference eld; the evolu-
tion of the mapping characterizes the evolution of the scalar eld for which the
PDF is sought. The mapping technique is capable of handling strongly non-
Gaussian (x, t). For homogeneous turbulent scalar elds the PDF equation
can be written as follows:
P(; t)
t
+
_
E
_
D
2
[
_
P(; t)
= 0, (29)
where D is the diusivity and E
_
D
2
[
_
is the expectation of D
2
condi-
tioned on (x, t) = . The cumulative distribution function (CDF) of P(; t)
is dened as
F(; t)
_
P(
; t)d
. (30)
Equation (30) taken together with (29) leads to the evolution equation for the
CDF:
F(; t)
t
+ E
_
D
2
[
_
F(; t)
= 0. (31)
The closure establishes the following mapping relation, (Pope 1991, Kollmann
1992):
s
(x, t) X[(z), t], (32)
where (x, t) is the dependent scalar eld, (z) is a time-independent stan-
dardized Gaussian reference eld residing at a spatial location z, and
s
(x, t)
is referred to as the surrogate eld, see Pope (1991), and is eectively dened
through equation (32). The mapping X does not determine the relation be-
tween the locations of (x, t) and (z). However, if the relation between the
locations in physical and reference space is restricted to a stretching transfor-
mation uniform in physical space, we have
z
m(t), (33)
600 Jones
where m(t) accounts for the average rescaling produced by the turbulent
stretching and molecular diusion, see Gao (1991b). The mapping X is con-
structed as follows: the value of the CDF of a turbulent eld at X(, x, t)
(where is the sample space of the Gaussian reference eld) is presumed
equal to the value of the standardized Gaussian CDF at (z) = ,
F(X(, x, t)) = F
G
(), (34)
implying the relation = X(, x, t). A direct consequence of the monotonicity
of the CDFs is that the mapping also increases monotonically, i.e. X(
1
) <
X(
2
) for
1
<
2
. Dierentiating equation (34) with respect to
F
=
F
G
we nd
P(; t) = P
G
_
X
_
1
. (35)
Performing chain dierentiation of equation (34) with respect to time yields
F
t
+
F
X
t
= 0. (36)
This result employs the concept of time independence of F
G
, the reference
CDF. Vali no et al. (1991) consider the case of time-dependent statistics, i.e.
s
(x, t) X[(z, t), t] (cf. equation (32)). In this case the equation for P
G
()
satises
P
G
(, t)
t
=
__
=
_
P
G
(, t)
_
. (37)
A closure hypothesis for
t
X(, x, t) = E
_
D
2
s
[
s
= X(, x, t)
_
. (38)
It is relatively straightforward to show that the Laplacian of the surrogate
eld is
s
(x, t) =
2
X((z), t) = m
2
_
X
z
i
z
i
+X
z
i
z
i
_
, (39)
[20] The joint scalar probability density function 601
where X
,= z has
no inuence on X((z), t). Thus, the condition (x, t) = is the image of
the condition (z) = . This, taken with the results for Gaussian statistics of
the reference eld, see Pope (1991), leads to the evolution equation for the
mapping:
X
t
= m
2
D
_
z
i
z
i
__
X
2
)
X
_
, (40)
where
2
) = 1 and D is treated as spatially uniform. Since equation (29)
makes use of the mixing term employing Fickian diusion, this strictly implies
that D cannot vary from one specie to another thus it appears that this
level of closure is constrained to an equal diusivity assumption. The mapping
closure described above relates to single point statistics. As in all other one-
point, one-time closures, length scale information is not available;
z
i
z
i
) in
equation (40) requires external information and is similar in form to a scalar
dissipation. Vali no and Gao (1992) suggest the use of a characteristic time
k/, which is what is used in other mixing models. Another possibility is to
model the joint scalar and its gradient, see Vali no and Dopazo (1991b). In this
case
2
)/D
s
[
s
=
_
. .
E
s
= E
_
2
[ =
_
. .
E
or E
_
(
s
)
2
[
s
=
_
= E
_
()
2
[ =
_
. (41)
Gao (1991b) tested this assumption by comparing predictions of E
s
(
s
)/E
s
(
1
2
),
for which an analytical expression exists (using the initial double -function
specication) against DNS results. Excellent agreement was obtained suggest-
ing that the underlying assumption behind equation (41) is sound. In section 4
it was pointed out that the DNS results of Eswaran and Pope (1988) predicted
Gaussianity after a relatively long period of time, implying the existence of a
602 Jones
substantial interim period of non-Gaussianity. Gao (1991b) demonstrated that
the evolution of E
s
was closely related to its initial value, which can be aected
by dierent initial length scales, consistent with DNS results. The observed
persistence of non-Gaussianity in the atness and superskewness is clearly
related to the behaviour of the tails of the PDF. When near the mean, the
behaviour is strongly Gaussian, while near the bounds it is not (but after sub-
stantial mixing, very little probability, i.e. area under the PDF, exists near the
tails). Miller et al. (1993) used mapping closures together with other mapping
relations (referred to as JohnsonEdgeworth translations) and DNS results to
show that the level of agreement of the normalized conditional dissipation of
the scalar at the bounds was rather poor. The deciency was attributed to the
inability of the closures to allow for variations in the scalar bounds as the inert
mixing process evolved. The DNS results they employed for validation demon-
strated that the scalar maxima and minima migrated towards the mean with
time; these were used as an empirical input for calculations where the bounds
were allowed to relax, and the subsequent improvement in the predictions at
the scalar extremities was noted.
How useful mapping closures will be for prediction methods in practical
ows is still an open question and the problems associated with incorporat-
ing extra scalars (inert and reactive) in a mathematically and computationally
tractable manner need to be resolved. Furthermore, although such a technique
yields excellent results in homogeneous turbulence, it is not clear whether it
will necessarily yield better results in inhomogeneous ows where the turbulent
time scales can vary and are not known. The use of k/ as the only character-
istic time scale may limit its usefulness relative to simpler and cruder models
previously discussed.
4.6 EMST Mixing Model
The Euclidean minimum spanning tree (EMST) mixing model formulated by
Subramanian and Pope (1998) retains some of the properties of mapping clo-
sures and overcomes some of the problems related to their extension to multiple
scalars. The model can also be viewed as a generalisation of the ordered pair-
ing model of Norris and Pope (1991). It utilises the concept of localness, the
satisfaction of which requires mixing models to be local in the sample space.
In terms of particles it implies that only particles with adjacent properties
can interact and mix; for example a particle with property can mix only
with particles with properties . Since all of the scalars considered are
continuous functions of position and time, this is clearly a property of the
exact scalar conservation equations. In the EMST model a Euclidean mini-
mum spanning tree constructed in scalar space is used to dene neighbouring
particles and mixing is represented by a pairwise interaction of particles with
adjacent properties. A straightforward application of the model to a simple
test problem led to stranding patterns in scatter plots which was eliminated
[20] The joint scalar probability density function 603
by the incorporation of an intermittency-like aspect through the introduction
of a particle age property. This is then used to distinguish a subset of particles
for mixing. An integral turbulence frequency
k
essentially determines the
rate at which mixing occurs. Although it complies with localness concepts, the
model does not satisfy linearity and independence principles, probably a re-
ection of the fact that it seems to be impossible to satisfy all exact properties
simultaneously while retaining a useful model at the single-point single-time
level of closure. The model has been shown to yield encouraging results for
inert scalar mixing and, in contrast to other binary interaction models, to be
consistent with the ame sheet approximation in diusion ames. More re-
cently Xu and Pope (2000) and Tang et al. (2000) have applied the model in
a joint velocity-scalar PDF formulation to obtain extremely impressive results
for a piloted turbulent diusion ame displaying local extinction. However at
the present time it is uncertain whether this arises as a consequence of the
mixing model or from the relatively detailed and realistic chemical reaction
mechanism used in their calculations.
4.7 Models for the Turbulent Transport of the PDF
Traditional modelling practice suggests that in many ows turbulent transport
can be adequately represented through a gradient transport hypothesis. For
many of the applications of the PDF method to thin shear ows it appears to
yield acceptable results for quantities relating to gross mixing patterns (e.g.
mixture fraction). Hence we can write
x
k
_
u
k
[ = )
P()
_
. .
IV
x
k
_
P()
x
k
_
, (42)
where
t
= 0.7 and
t
is given by the k- model. With a second moment
closure being used for the velocity eld, Chen and Kollmann (1988) used the
following approximation:
x
k
_
u
k
[ = )
P()
_
. .
IV
x
k
_
c
s
k
k
u
P()
x
j
_
, (43)
with c
s
= 0.3.
In the past, other proposals for closing the turbulent diusion term have
been made. Dopazo (1975) tentatively proposed an LMSE-type approxima-
tion and such an expression can be expected to work well in nearly Gaussian
cases. Janicka et al. (1978) used a modelled transport equation for u
k
[ = )
(i.e. Reynolds averaging and for a single conserved scalar) though the mod-
elling involved a formidable task. Certain terms involving the pressure were
604 Jones
neglected, through ad hoc assumptions or order of magnitude analysis, and
conditional Reynolds stresses were modelled using a gradient transport hy-
pothesis. However, these methods have now been largely replaced by the gra-
dient models referred to above.
5 Stochastic Methods for Solving the Scalar PDF
Equation
Because of the large number of independent variables involved in the PDF
equation, nite dierence techniques are not feasible and stochastic (Monte
Carlo) methods have to be resorted to. In the methods to be described it will
be assumed that turbulent transport of the PDF is represented by the gradient
transport model as discussed in section 4.7 but no restriction will be placed
on the mixing models to be used, other than that a particle method can be
constructed to represent it. In the present context it is important to note that
the Monte Carlo solution method is not a statistical simulation of turbulence
but rather of the closed form of the PDF evolution equation. Stochastic particle
methods can be formulated in either an Eulerian or Lagrangian framework and
both approaches will be outlined here.
5.1 Eulerian Particle Method
In the Eulerian particle method the PDF is represented by an ensemble of ^
particles located at each node point with each particle having N properties
corresponding to the random variables of the PDF. To utilise these particles
the closed form of the PDF equation is rst discretised using an explicit, i.e
forward time, approximately factored nite dierence approximation which
may be written as
P(; x, t + t) = (I + t)
. .
Source
(I + t)
. .
Mixing
(I +Tt)
. .
Transport
P(; x, t + t
T
) = (I +Tt)
P(; x, t), where t
T
is a notional time step
indicating that transport has been applied. A similar procedure is applied,
rst to represent mixing and then to account for chemical reaction.
On completion of these operations an approximation to
P(; x, t + t) is
obtained. Obviously the discretisation must be stable and consistent and the
time step must be suciently small for the factorisation errors to be negligible.
[20] The joint scalar probability density function 605
In general this requires the time step to be such that the Courant and diu-
sion numbers are small, (compared with unity). The transport contribution to
equation (44) at position p can be written as
P
p
(; x, t + t
T
) = t
(x, t
n
)P
(; x, t
n
), (45)
where the summation is over the points involved in the discretisation.
To simulate transport the ensemble of particles representing the PDF at po-
sition p and t +t
T
is now obtained by selecting at random tA
^ particles
from the ensemble at each node involved in the discretisation. Mixing is then
simulated by allowing the particles within each ensemble to mix according to
the specied model, and chemical reaction is described by allowing each parti-
cle to react according to the chemical mechanism being used. A more detailed
description of the method is provided by Pope (1981), Kakhi (1994) and Jones
and Kakhi (1996).
The method oers the advantage that, in a single step, the particles are ex-
changed only between neighbouring grid nodes to form a new ensemble. Hence
the numerical implementation is fairly straightforward since the stochastic
particle locations can be described by integer pointers. However it also suers
from a number of major disadvantages which severely restricts its use. First
the dierence coecients A
^) < 1, (46)
is violated then the simulated PDF at that node will be completely unaected
by transport processes! Furthermore proper account of convection and diu-
sion requires that the number of particles at each node must satisfy, for all
A
,= 0,
(tA
^) > m (47)
where m is a number at least greater than 1. Since the maximum value of t
is determined by stability considerations and the A
p=1
(p)
(
(p)
)( x
(p)
x), (48)
where ^ is the total number of particles,
(p)
is the weight of pth particle,
x
(p)
is the position of pth particle and the mean weight is dened as
=
1
^
N
p=1
(p)
. (49)
The discrete dierence approximation is then simulated stochastically by
allowing particle properties to evolve in a specied manner with their positions
being tracked as they move through the solution domain. Each particle obeys
certain equations which govern its transport. The particles move in physical
space by convection due to the mean ow, and by diusion due to the molecular
and eddy diusivity.
5.2.1 Transport
In Lagrangian PDF formulations convection is represented deterministically by
translation of particle positions, while gradient-diusion may be represented by
a classical random walk in physical space. Convection and gradient transport
diusion processes can be described by the following stochastic dierential
equation
d x
(p)
(t) = D( x
(p)
(t), t) t +E( x
(p)
(t), t) dW, (50)
where x is the position of a stochastic particle, D and E are the drift and
diusion coecient, respectively, and W is an isotropic Wiener process.
The Wiener process (Chandrasekhar 1954, Cox and Miller 1965) is a very
useful and widely used stochastic process for representing the random forces
[20] The joint scalar probability density function 607
corresponding to a diusion process. The Wiener process involves the random
variable
i
; and is dened by
W(t
n
) t
1/2
n
i=1
i
, (51)
where t
n
= n(t) and
i
is the sequence of n independent normal random
variables with zero mean and a unit variance. The increment of this process,
written as
W(t
n
) = W(t
n
) W(t
n1
), (52)
has the following properties:
W(t
n
)) = t
1/2
n
) = 0; (53)
W
2
(t
n
)) = t
2
n
) = t. (54)
That is, the increment of the Wiener process is a Gaussian random variable
with zero mean and a variance equal to the time step t. Equations (51) and
(52) show that the Wiener process is a continuous process, since W(t
n
) 0,
that is W(t
n
) W(t
n1
), as t goes to zero. It also shows that the
Wiener process is not dierentiable because, as t becomes innitely small,
the derivative of the process, W(t
n
)/t, becomes innite.
Comparing the FokkerPlanck equation corresponding to equation (50) with
the PDF evolution equation, we nd the drift and diusion coecients become
E
_
2
t
(55)
D u +
1
t
. (56)
Hence the stochastic dierential equation which represents the spatial trans-
port of the PDF is
d x
p
(t) =
_
u +
1
t
_
t +
_
2t
t
_
1/2
dW, (57)
and hence the spatial transport during interval time t is accounted for by
moving the stochastic particles according to
x
(p)
(t +t) = x
(p)
(t) +
_
u +
1
t
_
t +
_
2t
t
_
1/2
d, (58)
where is a standard normal random vector and x
(p)
denotes the position of
pth particle.
608 Jones
5.2.2 Mixing and Reaction
The modelled PDF equation written in Lagrangian form, (3), can be discre-
tised using an explicit forward-time, approximately-factored nite-dierence
scheme similar to that adopted in the Eulerian method. This is then used to
simulate the inuence of mixing and reaction on the evolution of the PDF by
rst sorting particles into their respective control volumes and then allowing
the resulting ensemble of particles in each cell to mix according to the specied
mixing model. Each particle is then allowed to react via the selected reaction
mechanism.
The variable particle weights introduced in equation (48) allow the number
of particles in each cell to be maintained roughly constant by the splitting
into two or removal of particles as appropriate. This removal and splitting
appears to be essential, particularly if the method is to be applied to two- and
three-dimensional ows with recirculation, and must be conducted in such a
manner that the statistical properties remain unchanged. For mixing models
based upon binary interactions, e.g. coalescence-dispersion models, and some
others, the particle representation must also be modied to account for the
diering particle weights for further details see Wouters (1998).
The Lagrangian particle method is not restricted to any particular coordi-
nate system and is thus quite general, and is relatively free from numerical
diusive type errors. However the implementation of the method in a compu-
tationally ecient manner, particularly for complex ows, is non-trivial and to
date it does not appear to have been widely used for the scalar PDF equation
methods described here but see Jones and Weerasinghe (2000), Weerasinghe
(2000) and Lindstedt et al. (2000). It has been applied most extensively in the
context of joint velocity-scalar evolution equation formulations.
5.3 Stochastic Field Methods
The recently proposed stochastic eld method of Vali no (1998) and Vali no
et al. (2000) represents an extremely promising method for solving the joint
scalar PDF evolution equation. In this formulation the PDF is represented
by a set of stochastic elds rather than particles. The method is purely Eu-
lerian. The evolution of the stochastic elds are obtained from solution of
standard conservation-type partial dierential equations to which a stochastic
source term is added. The resulting elds are continuous and dierentiable
in space and continuous but non-dierentiable in time. As a consequence the
eld equations can be solved using the relatively standard techniques found
in most current CFD codes. The convergence properties of the method also
appear to be signicantly better than most particle-based methods. So far the
method has been utilised only in conjunction with the LSME mixing model
and further investigation is needed into its use with alternative mixing models.
[20] The joint scalar probability density function 609
The ease with which the method can be incorporated into any Eulerian CFD
code makes it very attractive indeed.
5.4 Chemical Reaction
The chemical source terms appear in exact and closed form in the PDF equa-
tion and as a consequence this process can be simulated deterministically. In
the solution methods described above this is implemented by allowing each
particle or eld to react over a time interval t according to the specied
reaction mechanism. This in turn requires solution of a system of ordinary
dierential equations (ODEs):
d
dt
(p)
()
()
where = 1, . . . , N. (59)
For heat releasing reactions the resulting system of ODEs is sti and their
solution is non-trivial, see Hindmarsh (1980). Furthermore a PDF equation
solution will involve a large number of particles, typically of order 10
6
, and for
each of these the system of equations (59) will have to be integrated over a
time interval corresponding to several mean ow through times. Consequently
limitations in computer speed and memory preclude the implementation of
detailed kinetic mechanisms, with the possible exception of thin shear layers,
i.e. simple jet ames, and reduced mechanisms represent the only practical
approach. A number of techniques are available for reducing detailed kinetic
mechanisms and these include traditional methods involving combinations of
steady state and partial equilibrium assumptions to determine minor species
as a function of major species (see, for example, Peters and Rogg 1993, Sung
et al. 1998), computational singular perturbation methods (Lam and Goussis
1988, Massias et al. 1999) and intrinsic low-dimensional manifold techniques
(Maas and Pope 1992, Blasenbrey and Maas 2000). In all these approaches the
number of independent species, for which equations of the type (59) have to be
solved, is reduced to a small number, typically between 4 and 12. Generally
speaking the more species retained, the wider is the range of applicability
of the reduced mechanism. However a detailed review of reduced chemical
mechanisms is beyond the scope of the present chapter.
A further strategy for reducing computational costs is to tabulate the chem-
istry. A simple way of doing this is to calculate the integrals of the reaction
rates for a specied time interval for all possible chemical states and then to
store the results in a look-up table, see Chen et al. (1989). The scalar bounds
(these determine the allowable scalar space) for the interpolation tables are
constructed using mass conservation principles and the assumption of con-
stant carbon-to-hydrogen (C/H) and oxygen-to-nitrogen (O/N) atom ratios
(see, for example, Chen et al. 1989, Sion and Chen 1992), consistent with the
equal diusivity assumption implicit in the described turbulence and mixing
610 Jones
models. During Monte Carlo simulations, the changes of the scalar properties
as a result of the reaction are obtained from multilinear interpolation based on
the previously generated look-up table, thereby replacing the repeated time
integration of the sti ODEs. This strategy is computationally ecient for a
small number of species, particularly if only major stable species are involved
in the table. However the overriding constraint becomes the size of the look-
up table. As the complexity of the reduced mechanism increases, the number
of independent scalars required to specify the thermo-chemistry increases. For
interpolation tables including more than six scalars, the demands on memory
can become quite unmanageable if a suciently ne resolution of the chem-
istry is required; coarse grids may lead to solutions casting doubts on the
accuracy of the results, see Sion and Chen (1992). A further limitation of the
approach is that, in any ame calculation, only a small part of the allowable
space will be accessed. This accessible part of the space, though inuenced to
some degree by the mixing model, is essentially determined by the chemical
reaction mechanism; for any given inlet and initial conditions there will be no
path to certain parts of the allowable space.
One extremely attractive method for overcoming the shortcomings of the
tabulation method described above is the in situ adaptive tabulation method
(ISAT) formulated by Pope (1997). Here the chemistry is tabulated on the
y using a binary tree and an adaptive technique is used to maintain tabula-
tion/interpolation errors below a specied tolerance. The method can result
in up to three orders of magnitude reduction in computer time compared with
direct integration of the equations (59) and with this approach it is possible
to incorporate quite detailed chemistry into the PDF equation method. Re-
cently Xu and Pope (2000) and Tang et al. (2000) have, with impressive results,
utilised the reduced mechanism of Sung et al. (1998), which involves 16 species
and 12 reaction steps, in a joint velocity-scalar PDF equation computation of
a series of piloted methane-air jet ames.
6 Applications
In order to illustrate the PDF equation approach, some results taken from
Jones and Kakhi (1998) and Jones and Prasetyo (1996) will be presented. In
both cases the combustion mechanism applied was the four-step global scheme
of Jones and Lindstedt (1988). With the assumption of equal diusivities this
involves four independent scalars. The chemistry was tabulated and a stochas-
tic particle method used to solve the PDF equation. The cases considered cor-
respond to a turbulent non-premixed ame involving a jet of propane burning
in stagnant air surrounds, measured by Godoy (1982), and an opposed jet
counterow conguration. The turbulent counter ow ame has been the sub-
ject of extensive experimental studies and premixed, partially-premixed and
non-premixed cases have been measured, see for example, Mastorakos (1993),
[20] The joint scalar probability density function 611
Mastorakos et al. (1993), Abbas et al. (1992), Kostiuk et al. (1993a, b), and
Mounaim-Rousselle and Gokalp (1993). In the present case the calculation
method has been applied to the conguration of Mastorakos et al. It consists
of two opposed pipes of diameter 25.4mm, the exits of which are separated by
20mm. These are surrounded by concentric pipes of diameter 51.8mm. The
results of two cases are presented here: a premixed ame resulting from the
burning of opposed jets of a methane-air mixture, and the ame arising from
opposed jets one of which comprised a methane-air mixture and the other,
pure air. Further details of the calculation methods, etc., may be obtained
from the original papers.
6.1 Propane-Air Jet Diusion Flames
The calculated and measured radial proles for the propane-air ame are
shown in gures 1 to 3. The general features of the ame are quite well repro-
duced by the calculation and the measured and predicted proles of mixture
fraction and C
3
H
8
are in close accord. The level of agreement displayed in the
proles of CO
2
and O
2
is also reasonable though there are some discrepancies.
The measurements indicate that O
2
does not penetrate to the centreline at
x/D = 20 and 40 whereas the calculations show small but appreciable levels
and the maxima in the predicted CO
2
proles occur slightly further from the
centreline compared with the measurements. However, gure 2 displays a sig-
nicant degree of over-prediction in the levels of CO. The primary reason for
this is probably associated with the chemical reaction mechanism being used.
This has been shown previously to yield accurate results for laminar ames but
in this case accurate transport properties were used. It now appears that the
mechanism is unduly sensitive to the inuence of diusivity under fuel rich
conditions compared to detailed reaction schemes. This, combined with the
equal diusivity assumption implicit in the mixing models, results in exces-
sively high CO concentrations. A striking observation however is that the two
mixing models, LMSE and modied Curl, yield results which are practically
indistinguishable.
6.2 Counterow Flames
The counterow conguration considered is illustrated in gure 4. A steady
ow solution of the non-reacting methane-air mixture was used to provide
initial conditions for the burning PDF calculation and the ame was then ig-
nited by setting the mixture in the vicinity of the stagnation region to fully
burnt conditions. The solution was then marched suciently forward in time
for steady ow conditions to prevail. Computations of the ame were carried
out using both the coalescence-dispersion model and the LMSE closure. How-
ever only the coalescence-dispersion generated plausible results and the ame
extinguished when the LMSE model was used.
612 Jones
Figure 1: Radial proles of the measurements, Godoy (1982), and predictions
for and C
3
H
8
in a propane-air jet diusion ame issuing into stagnant sur-
rounds. The k- model and the scheme of Jones and Lindstedt (1988) are
employed. Predictions using the LMSE and coalescence-dispersion closures
are shown; Re 24000.
The predicted and measured centreline proles of the major species are
shown in gures 5 to 7 for the two premixed jets, = 0.9, and in gures
8 to 10 for the single premixed jet, = 1.3. For both ames, the measured
proles of CH
4
are reproduced to a good accuracy and, in conformity with the
measurements, the predicted methane concentrations reduce to zero in the air
[20] The joint scalar probability density function 613
Figure 2: Radial proles of the measurements, Godoy (1982), and predictions
for CO and CO
2
in a propane-air jet diusion ame issuing into stagnant
surrounds. The k- model and the scheme of Jones and Lindstedt (1988) are
employed. Predictions using the LMSE and coalescence-dispersion closures are
shown; Re 24000.
stream for the single premixed jet and to zero in the region between the two
ames of the two premixed jets.
In the latter case the minimum O
2
concentrations in the burning region, g-
ure 6, are also well reproduced. This behaviour is consistent with the observed
complete combustion and suggests that the fuel breakdown has been properly
614 Jones
Figure 3: Radial proles of the measurements, Godoy (1982), and predictions
for O
2
and N
2
in a propane-air jet diusion ame issuing into stagnant sur-
rounds. The k- model and the scheme of Jones and Lindstedt (1988) are
employed. Predictions using the LMSE and coalescence-dispersion closures
are shown; Re 24000.
simulated. A reasonable level of agreement is also evident in the proles of
CO
2
, see gures 7 and 9, where the prole shapes and maximum values are re-
produced to an acceptable accuracy. However in the case of CO, gure 10, the
level of agreement with measured values is less satisfactory. For the single pre-
mixed jet the predicted maximum CO concentration exceeds by around 30%
[20] The joint scalar probability density function 615
D
r
Do
H
Hp
Hp
stream 2
stream 1
z
Figure 4: Counterow jet.
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
M
e
a
n
C
H
4
m
o
l
e
f
r
a
c
t
i
o
n
Mastorakos
CoalescenceDispersion
Figure 5: Counter ow ame: mean methane mole fraction prole.
616 Jones
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
0.20
0.30
0.40
0.50
M
e
a
n
O
2
m
o
l
e
f
r
a
c
t
i
o
n
Mastorakos
CoalescenceDispersion
Figure 6: Counter ow ame: mean oxygen mole fraction prole.
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
0.20
0.30
0.40
0.50
M
e
a
n
C
O
2
m
o
l
e
f
r
a
c
t
i
o
n
(
x
5
)
Mastorakos
CoalescenceDispersion
Figure 7: Counter Flow Flame: mean carbon dioxide mole fraction prole.
[20] The joint scalar probability density function 617
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
0.20
M
e
a
n
C
H
4
m
o
l
e
f
r
a
c
t
i
o
n
Mastorakos
CoalescenceDispersion
Figure 8: Counter ow ame: mean methane mole fraction prole.
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
0.20
0.30
0.40
0.50
M
e
a
n
C
O
2
m
o
l
e
f
r
a
c
t
i
o
n
(
x
5
)
Mastorakos
CoalescenceDispersion
Figure 9: Counter ow ame: mean carbon dioxide mole fraction prole.
618 Jones
0.0 5.0 10.0 15.0 20.0
z (mm)
0.00
0.10
0.20
M
e
a
n
C
O
m
o
l
e
f
r
a
c
t
i
o
n
(
x
5
)
Mastorakos
CoalescenceDispersion
Figure 10: Mean carbon monoxide mole fraction prole.
the measured value of approximately 2.8%. As was the case for the propane-air
ame the main reason for this is probably related to the limitations associated
with the reaction mechanism and its interaction with the mixing model.
7 Concluding Remarks
In the preceding sections the joint scalar PDF evolution equation method for
calculating the properties of combusting ows was reviewed. Attention was
restricted to the equation for the joint PDF of the set of scalars (at the single-
point, single-time level) describing reacting ows. This constitutes about the
simplest level of closure in terms of the PDF formulation applied to combusting
ows, and has the merit that chemical source (reaction rate) terms appear in
closed form. In the method, the mean velocity and turbulence characteristics
are obtained from a conventional turbulence model approach, and accurate
mean velocity and turbulence elds are a clear prerequisite for success. Also,
if the method is to be applied successfully to turbulent combustion, then an
accurate mechanism for describing chemical reaction must be provided.
As in all statistical approaches, the PDF transport equation is unclosed
and approximations must be provided for the unknown terms. These comprise
terms representing turbulent transport in physical space and micromixing in
scalar space. For the former, a simple gradient transport model will probably
[20] The joint scalar probability density function 619
suce under many circumstances, provided of course that the mean velocity
eld is accurately reproduced by the turbulence model. Micromixing on the
other hand presents a closure problem distinctly dierent from that arising
in conventional moment closures. In fact the closure of the micromixing term
poses a serious diculty and, together with the chemical reaction mechanism,
is central to turbulence-chemistry interactions primarily because it involves
molecular diusivities and scalar gradients. These scale with the dissipative
motions in turbulent ow and this is precisely where combustion occurs. The
most widely used classes of mixing models are those based on the binary in-
teraction of uid parcels (coalescence-dispersion and EMST closures) and on
linear mean square estimation (LMSE). While these are all capable of repro-
ducing measured statistics, including variance decay rates, etc., they all have
limitations in terms of their ability to represent PDF evolution. For example,
in homogeneous isotropic turbulence they are incapable of reproducing the
observed relaxation of an arbitrary initial (passive) scalar eld to a Gaussian
PDF. This deciency has motivated the formulation of binomial Langevin and
mapping closures, both of which allow a relaxation to a Gaussian PDF un-
der the appropriate conditions. However initial eorts have concentrated on
the evolution of a single, usually passive, scalar and the extension of these
improved models to the case of several reactive scalars appears dicult and
further eort is needed. Also, in the case of reactive scalars (species mass frac-
tions) the PDF is likely to be far from Gaussian because the mass fractions
must lie in [0, 1] and because the PDF will be strongly aected by reaction
and under these circumstances the ability of mixing closures to describe re-
laxation to a Gaussian PDF, although desirable, may not be a primary factor
in accurately predicting turbulent ame properties. Also the deviation from
Gaussian in homogeneous isotropic turbulence may not be large, particularly
with binary interaction closures: the skewness, atness and superskewness,
which are often used to provide a measure of PDF shape, emphasise the be-
haviour far from the mean where probabilities are inevitably small.
A potentially more serious limitation may arise from the fact that the ma-
jority of current mixing models involve use of a single time scale, k/, which
clearly implies the ratio of mechanical to scalar turbulence time scale is con-
stant. As is well known, this assumption is of limited generality. As an al-
ternative the solution of an equation for the joint PDF of the scalar and its
gradient has often been suggested and this alleviates the problem of evaluating
turbulent scalar time scales. However for a set of reactive scalars, the increase
in complexity would then be considerable indeed. An interim step would be
to introduce a single scalar time scale based on the solution of a modelled
equation for the scalar dissipation rate, such as that formulated by Jones and
Musonge (1988). The precise denition of the scalar would have to be chosen
with care in each case but for non-premixed ames the mixture fraction could
be used as a quantity representative of scalar mixing. A further implication of
the use of an integral time scale is that the models are based on the notion of
620 Jones
an equilibrium turbulence spectrum whereby scalar turbulence is created in
the large scale motions, transferred through an inertial subrange to the ne
scales where diusivity acts to smooth out the scalar uctuations, i.e. where
scalar dissipation occurs. Thus, at least for an inert ow, the rate of transfer
of scalar turbulence through the inertial subrange will equal the scalar dissi-
pation rate. The mixing models in current use are in fact global models which
represent the rate of transfer through the inertial subrange rather than the
mixing process itself; they include no direct information on the molecular dif-
fusion processes that occur within the ne scale dissipative scalar uctuations;
amongst the assumptions implied is that of equal diusivities.
On an instantaneous basis, burning is likely to occur in the ne scales of
turbulent motion. If burning is in the laminar amelet regime then (instanta-
neous) ame thicknesses will be smaller than the smallest scale dissipative uc-
tuations, i.e. the ame thickness will be smaller than the Kolmogorov length
scale. Even if the conditions for amelet burning are not satised then the
fast reaction rates associated with heat release will ensure that burning takes
place in layers which are thin compared with the (inertial) length scales repre-
sentative of energy containing turbulent motions. In both situations molecular
diusion processes are likely to exert an important inuence on burning. Now,
species with low molar mass, such as H and H
2
, diuse at a much greater rate
than others and this is likely to have important consequences. In hydrocarbon-
air ames H and H
2
are, essentially, produced and consumed in the burning
region, where they have maximum concentration, and are intimately linked to
OH and CO formation; high levels of H will give rise to high levels of OH. If no
account is taken of diusion processes in devising models for micromixing and
the assumption of equal molecular diusivities is consequently implicitly in-
voked, then the predicted implied instantaneous levels of H and H
2
are likely to
be higher than those observed. In reality the higher diusivities will result in H
and H
2
diusing away (on an instantaneous basis) from the burning region at
a greater rate than other species with a consequence that maximum H, H
2
and
OH concentrations will be reduced and this will impact on CO levels. Some
indication of the magnitude of this eect can be obtained from laminar ame
computations. Recently Hasko (2000) has used GRI-Mech 2.11 to compute a
range of laminar counterow ames using alternatively the equal diusivity
assumption and accurate transport properties. With the equal diusivity as-
sumption the computed mass fractions of H and OH were too high by a factor
of 2 to 3 and 1.4 to 2 respectively, depending on the strain rate. The eect on
CO levels was less pronounced, with the errors being typically around 10%. If
the above description of turbulent burning is correct then it is important to
note that the inuence of molecular transport properties on micromixing and
burning will be present at all Reynolds numbers. If the eects are signicant
what will be needed is a mixing model which provides some description of the
ne scale mixing processes in the presence of reaction. However, exactly how
this is to be constructed is at present unclear though the linear-eddy model
[20] The joint scalar probability density function 621
of Kerstein (1991) may provide the basis for doing so. The results of direct
numerical simulations of turbulent ames, which are becoming available, may
also provide guidance here.
In spite of the above remarks the PDF equation method incorporating the
modied coalescence-dispersion model and a simple global reaction mecha-
nism has been demonstrated to lead to results in general agreement with mea-
sured mixture fraction, fuel and CO
2
proles for both jet diusion ames and
counterow premixed and partially premixed turbulent ames. In cases where
burning is predicted, the LMSE and a modied coalescence-dispersion closure
produce practically identical results. However for the counterow ames it was
not possible to obtain a steady burning solution with the LMSE closure and
the implication is that the model predicts too low a value of critical strain
rate above which extinction occurs. The high predicted levels of CO com-
pared with measured values in all the ames considered are almost certainly a
consequence of limitations in the global reaction mechanism used, particularly
under fuel rich conditions. The recent results of Lindstedt et al. (2000) suggest
that better chemistry will bring signicant improvement in this regard.
Acknowledgements
It is a pleasure to acknowledge my indebtedness to Dr Maziar Kakhi, Mr Yudi
Prasetyo and Dr Rohitha Weerasinghe for their contributions to this work.
References
Abbas, N.S.T., Lockwood, F.C., Hussein, A.M.M., and Hassan, M.A. (1992). A study
of extinction and NO
x
in stretched premixed ames. J. Institute of Energy 65, 77.
Blasenbrey, T., and Maas, U. (2000). ILDMs of Higher Hydrocarbons and the Hier-
archy of Chemical Kinetics. In Proc. Combustion Institute 28, 16231630.
Chandrasekhar, S. (1954). Stochastic problems in physics and astronomy. In N. Wax
(ed.), Selected Papers on Noise and Stochastic Processes. Dover Publications, Inc.,
New York.
Chen, H., Chen, S., and Kraichnan, R. (1989). Probability distribution of a stochas-
tically advected scalar eld. Physical Review Letters 63(24), 26572660.
Chen, J.-Y., and Kollmann, W. (1988). PDF modelling of chemical non-equilibrium
eects in turbulent non-premixed hydrocarbon ames. In Proc. Combustion Insti-
tute 22, 645653.
Chen, J.-Y., and Kollmann, W. (1991). Mixing models for turbulent ows with exother-
mic reactions. In Turbulent Shear Flows 7, Springer-Verlag, 277292.
Chen, J.-Y., Kollmann, W., and Dibble, R.W. (1989). PDF modelling of turbulent
non-premixed methane jet ames. Combustion Science and Technology 64, 315
346.
Cox, D.R., and Miller, H.D. (1965). The Theory of Stochastic Processes. Methuen.
622 Jones
Curl, R. (1963). Dispersed phase mixing: 1. Theory and eects in simple reactors.
A.I.Ch.E. Journal 9(2), 175181.
Dopazo, C. (1973). Non-isothermal turbulent reactive ows: stochastic approaches.
PhD thesis, State University of New York at Stony Brook.
Dopazo, C. (1975). Probability density function approach for a turbulent axisym-
metric heated jet: centreline evolution. Physics of Fluids 18(4), 397404.
Dopazo, C. (1979). Relaxation of initial probability density functions in the turbulent
convection of scalar elds. Physics of Fluids 22(1), 2030.
Dopazo, C., and OBrien, E.E. (1974). An approach to the autoignition of a turbulent
mixture. Acta Astronautica 1, 12391266.
Dopazo, C., and OBrien, E.E. (1976). Statistical treatment of non-isothermal chem-
ical reactions in turbulence. Combustion Science and Technology 13, 99122.
Eswaran, V., and Pope, S.B. (1988). Direct numerical simulations of the turbulent
mixing of a passive scalar. Physics of Fluids A 31(3), 506520.
Flagan, R.C., and Appleton, J.P. (1974). A stochastic model of turbulent mixing
with chemical reaction: nitric oxide formation in a plug-ow burner. Combustion
and Flame 23, 249267.
Gao, F. (1991a, April). An analytical solution for the scalar probability density func-
tion in homogeneous turbulence. Physics of Fluids A 3(4), 511513.
Gao, F. (1991b, October). Mapping closure and non-Gaussianity of the scalar prob-
ability density function in isotropic turbulence. Physics of Fluids A 3(10), 2438
2444.
Godoy, S. (1982). Turbulent diusion ames. PhD thesis, University of London.
Hasko, S.M. (2000). Private Communication.
Hindmarsh, A.C. (1980). LSODE and LSODI, two new initial value ordinary dier-
ential equation solvers. ACM SIGNUM Newsletter 15(4).
Hulek, T., and Lindstedt, R.P. (1996). Computations of steady-state and transient
premixed turbulent ames using PDF methods. Combustion and Flame 104(4),
481504.
Janicka, J., Kolbe, W., and Kollmann, W. (1978). The solution of a PDF transport
equation for turbulent diusion ames. In Proc. Heat Transfer and Fluid Mechanics
Institute, Stanford University Press, 296312.
Janicka, J., Kolbe, W., and Kollmann, W. (1979). Closure of the equations for the
probability density function of turbulent scalar elds. Journal of Non-Equilibrium
Thermodynamics 4, 4766.
Jones, W.P., and Kakhi, M. (1996). Mathematical modelling of turbulent ames. In
Unsteady Combustion, M.H.F. Culick and J.H. Whitelaw (eds.), Kluwer Academic
Publishers, 411491.
Jones, W.P., and Kakhi. M. (1997). Application of the transported PDF approach
to hydrocarbon turbulent jet diusion ames. Combustion Science and Technol-
ogy 129, 393430.
Jones, W.P., and Kakhi, M. (1998). PDF modelling of nite-rate chemistry eects in
turbulent non-premixed jet ames. Combustion and Flame 115, 210229.
[20] The joint scalar probability density function 623
Jones, W., and Kollmann, W. (1987). Multi-scalar PDF transport equations for tur-
bulent diusion ames. In F. Durst, B. Launder, J. Lumley, F. Schmidt, and
J. Whitelaw (eds.), Turbulent Shear Flows 5, Springer-Verlag, 296309.
Jones, W.P., and Lindstedt, R.P. (1988). Global reaction schemes for hydrocarbon
combustion. Combustion and Flame 73, 233249.
Jones, W.P., and Musonge, P. (1988). Closure of the Reynolds stress and scalar ux
equations. Physics of Fluids 31(12), 35893603.
Jones, W.P., and Prasetyo, Y. (1996). Probability density function modelling of pre-
mixed turbulent opposed jet ames. In Proc. Combustion Institute 26, 275282.
Jones, W.P., and Weerasinghe, W.M.S.R. (2000). Probability density function mod-
elling of an axisymmetric combustion chamber. In Advances in Turbulence VII,
C. Dopazo (ed.), CIMNE, Barcelona, 497500.
Kakhi, M. (1994). The transported probability density function approach for predicting
turbulent combusting ows. PhD thesis, University of London.
Kerstein, A.R. (1991). Linear-eddy modelling of turbulent transport. Part 6. Mi-
crostructure of diusive scalar mixing fields. Journal of Fluid Mechanics 231,
361394.
Kollmann, W. (1989). PDF transport equations for chemically reacting ows. In
R. Borghi and S. Murthy (eds.), Turbulent Reactive Flows, Spinger-Verlag, 715730.
Kollmann, W. (1990). The PDF approach to turbulent ow. Theoretical and Compu-
tational Fluid Dynamics 1, 249285.
Kollmann, W. (1992). PDF transport modelling. Modelling of Combustion and Tur-
bulence, Von Karman Institute for Fluid Dynamics Lecture Series.
Kosaly, G. (1986). Theoretical remarks on a phenomenological model of turbulent
mixing. Combustion Science and Technology 49, 227234.
Kosaly, G., and Givi, P. (1987). Modelling of turbulent molecular mixing. Combustion
and Flame 70, 101118.
Kostiuk, L.W., Bray, K.N.C., and Cheng, R.K. (1993a). Experimental study of pre-
mixed turbulent combustion in opposed stream: Part I: Non-reacting ow elds.
Combustion and Flame 92, 233.
Kostiuk, L.W., Bray, K.N.C., and Cheng, R.K. (1993b). Experimental study of pre-
mixed turbulent combustion in opposed stream: Part II: Reacting ow elds and
extinction. Combustion and Flame 92, 396.
Lam, S.H., and Goussis, D.A. (1988). Understanding complex chemical kinetics with
computational singular perturbation. In Proc. Combustion Institute 22, 931941.
Lindstedt, R.P., Louloudi, S.A., and V aos (2000). Joint scalar PDF modelling of
pollutant formation in piloted turbulent jet diusion flames with comprehensive
chemistry. In Proc. Combustion Institute 28, 149156.
Lundgren, T.S. (1967). Distribution functions in the statistical theory of turbulence.
Physics of Fluids 10(5), 969975.
Maas, U., and Pope, S.B. (1992). Simplifying chemical kinetics: intrinsic low dimen-
sional manifolds in composition space. Combustion and Flame 88, 230264.
624 Jones
Massias, D., Diamantis, D., Mastorakos, E., and Goussis, D.A. (1999). An algorithm
for the construction of global reduced mechanisms with csp data. Combustion and
Flame 117, 685708.
Mastorakos, E. (1993). Turbulent combustion in opposed jet ows. PhD thesis, Uni-
versity of London.
Mastorakos, E., Taylor, A.M.K.P., and Whitelaw, J.H. (1993). Mixing in turbulent
opposed jet ows. In Proc. 9th Symp. Turbulent Shear Flows, Kyoto, Japan.
Miller, R.S., Frankel, S.H., Madnia, C.K., and Givi, P. (1993). JohnsonEdgeworth
translation for probability modelling of binary scalar mixing in turbulent ows.
Combustion Science and Technology 91, 2152.
Mounaim-Rousselle, Ch., and Gokalp, I. (1993). Turbulent premixed combustion in
counterow geometry: the inuence of coow.
Norris, A.T., and Pope, S.B. (1991). Turbulent mixing model based on ordered pairing.
Combustion and Flame 83, 2742.
OBrien, E.E. (1980). The probability density function (PDF) approach to reacting
turbulent ows. In Turbulent Reacting Flows, P. Libby and F. Williams (eds.),
Springer-Verlag. 185218. Appeared under Topics in Applied Physics, vol.44.
Peters, N. (2000). Turbulent Combustion. Cambridge University Press.
Peters, N., and Rogg, B. (eds.) (1993). Reduced Kinetic Mechanisms for Applications
in Combustion Systems. Lecture Notes in Physics M15. Springer-Verlag.
Pope, S.B. (1979). The relationship between the probability approach and particle
models for reaction in homogeneous turbulence. Combustion and Flame 35, 4145.
Pope, S.B. (1981). A Monte Carlo method for the PDF equations of turbulent reactive
ows. Combustion Science and Technology 25, 159174.
Pope, S.B. (1982). An improved mixing model. Combustion Science and Technol-
ogy 28, 131135.
Pope, S.B. (1983). Consistent modelling of scalars in turbulent flows. Physics of Flu-
ids 26, 404408.
Pope, S.B. (1985). PDF methods for turbulent reactive ows. Progress in Energy and
Combustion Science 11, 119192.
Pope, S.B. (1991). Mapping closures for turbulent mixing and reaction. Theoretical
and Computational Fluid Dynamics 2, 255270.
Pope, S.B. (1997). Computationally ecient implementation of combustion chemistry
using in situ adaptive tabulation. Combustion Theory and Modelling 1(1), 4163.
Sion, M., and Chen, J.-Y. (1992). Scalar PDF modelling of turbulent non-premixed
methanol-air ames. Combustion Science and Technology 88, 89114.
Spielman, L.A., and Levenspiel, O. (1965). A Monte Carlo treatment for reacting and
coalescing dispersed phase systems. Chemical Engineering Science 20, 247254.
Subramanian, S., and Pope, S.B. (1998). A mixing model for turbulent reactive flows
based on Euclidean minimum spanning trees. Combustion and Flame 115(4), 487
514.
[20] The joint scalar probability density function 625
Sung, C.J., Law, C.K., and Chen, J.-Y. (1998). An augmented reduced mechanism
for methane oxidation with comprehensive global parametric validation. In Proc.
Combustion Institute 27 295.
Tang, Q., Xu, J and Pope, S.B. (2000). PDF calculations of local extinction and
no production in piloted-jet turbulent methane/air flames. In Proc. Combustion
Institute 28, 133139.
Tavoularis, S., and Corrsin, S. (1981). Experiments in nearly homogeneous turbulent
shear ow with a uniform mean temperature gradient. Part 1. Journal of Fluid
Mechanics 104, 311347.
Vali no, L. (1998). A eld Monte Carlo formulation for calculating the probability
density function of a single scalar in a turbulent ow. Flow, Turbulence and Com-
bustion 60, 157172.
Vali no, L., and Dopazo, C. (1990). A binomial sampling model for scalar turbulent
mixing. Physics of Fluids A 2(7), 12041212.
Vali no, L., and Dopazo, C. (1991a). A binomial Langevin model for turbulent mixing.
Physics of Fluids A 3(12), 30343037.
Vali no, L., and Dopazo, C. (1991b). Joint statistics of scalars and their gradients in
nearly homogeneous turbulence. In Advances in Turbulence 3, A. Johansson and
P. Alfredsson (eds.), Springer-Verlag, 312323.
Vali no, L., and Gao, F. (1992, September). Monte Carlo implementation of a single-
scalar mapping closure for diusion in the presence of chemical reaction. Physics
of Fluids A 4(9), 20622069.
Vali no, L., Larroya, J.C., and Cazalens, M. (2000). A eld Monte Carlo formulation
for solving multiscalar probability density functions in turbulent ows: Application
to the study of a turbulent non-premixed methane ame using detailed chemistry.
submitted for publication.
Vali no, L., Ros, J., and Dopazo, C. (1991, September). Monte Carlo implementa-
tion and analytic solution of an inert-scalar turbulent mixing test problem using a
mapping closure. Physics of Fluids A 3(9), 21952198.
Vervisch, L. (1991). Prise en compte deets de cinetique chimique dans les ammes
de diusion turbulentes par lapproche fonction densite de probabilite. PhD thesis,
University of Rouen.
Vervisch, L. (1992). Applications of PDF turbulent combustion models to real non-
premixed ame calculations. Modelling of combustion and turbulence. Von Karman
Institute Lecture Series.
Warhaft, Z., and Lumley, J.L. (1978). An experimental study of the decay of tem-
perature uctuations in grid-generated turbulence. Journal of Fluid Mechanics 88,
659684.
Weerasinghe, W.M.S.R. (2000). Application of Lagrangian probability density function
approach to turbulent reacting ows. PhD thesis, University of London.
Wouters, H.A. (1998). Lagrangian models for turbulent reacting ows. PhD thesis,
Delft University of Technology.
Xu, J., and Pope, S.B (2000). PDF Calculations of Turbulent Nonpremixed Flames
with Local Extinction. Combustion and Flame 123(3), 281307.
21
Joint Velocity-Scalar PDF Methods
H.A. Wouters, T.W.J. Peeters and D. Roekaerts
In this chapter the joint velocity-scalar PDF approach is described. This ap-
proach was mainly developed by S.B. Pope and includes from the start the
complete one-point joint statistics of velocity and scalars. This is conceptually
appealing because it delivers in one framework closure models for Reynolds
stresses, Reynolds uxes and chemical source terms. The present text is based
on the PhD thesis of H.A. Wouters (Wouters 1998), which can be consulted
for further details and other applications.
The outline of this chapter is as follows. First the exact transport equation
for the velocity-scalar PDF is introduced and the closure problem is discussed.
Next the Monte Carlo solution method that is used to model and solve the PDF
transport equation is described. Modelling of the unclosed terms describing
acceleration in the turbulent ow are treated in some detail. The closure of
the micromixing terms can be done along parallel lines as is the case in the
scalar PDF method discussed in [20] and is not elaborated here. A description
of some methods for handling complex chemistry is included.
In the nal section results of test calculations are presented for a challeng-
ing test case, combining a complex ow pattern with eects of high mixing
rates and chemical kinetics. For this blu-body-stabilized diusion ame, the
relative importance of modeling of the velocity, mixing and chemistry terms
is studied.
1 PDF transport equation
From the full conservation equations given in [10], the transport equation for
the joint velocity-scalar PDF f
U
(V , ; x, t) can be derived (Pope 1985).
By integrating this quantity over a range of values of V and the proba-
bility that U and take values in these ranges is obtained. It is also useful
to consider the joint velocity-scalar mass density function (MDF) dened as
T
U
(V , , x; t) = ()f
U
(V , ; x, t). The transport equation for the MDF
reads
t
T +
x
i
V
i
T =
V
i
_
1
()
_
p)
x
i
T
ij
)
x
j
_
T
_
(1.1a)
V
i
_
1
()
_
x
i
ij
x
j
U = V , =
_
T
_
(1.1b)
626
[21] Joint velocity-scalar PDF methods 627
__
1
()
J
i
)
x
i
+S
()
_
T
_
(1.1c)
_
1
()
_
i
x
i
U = V , =
_
T
_
, (1.1d)
in which the rst two terms (1.1a) and (1.1b) on the right-hand side describe
the evolution in velocity space and the last two terms (1.1c) and (1.1d) describe
the evolution in scalar space. (The notation A[ B) denotes the conditional
expectation value of A upon condition B). Terms (1.1a) and (1.1c) occur in
closed form whereas the unclosed terms (1.1b) and (1.1d) contain conditional
averages because these eects cannot be expressed in terms of the one-point
distribution of U and .
The terms (1.1a) that describe the evolution in velocity space contain the
mean pressure gradient and the mean viscous stress tensor. Both terms can
be expressed in terms of the mean velocity eld and thus occur in closed form.
Assuming high Reynolds number, eects of mean shear T
ij
) are usually ne-
glected. The unclosed terms (1.1b) describe the mean eects of the uctuating
pressure gradient and the uctuating viscous stress tensor, conditional on the
values of velocity and scalars.
The terms (1.1c) that describe the evolution in scalar space contain the
eects of the mean molecular diusion ux and of reaction. The reaction rate
can be expressed as a function of the scalar variables and therefore it is
contained in closed form. The unclosed term (1.1d) describes the eects of
molecular mixing or micro-mixing given by the conditional mean eects of the
uctuating diusion ux. Note that the conditioning here is on both scalars
and velocity. In practice the same micromixing models are usually used in the
velocity-scalar PDF approach as in the scalar PDF approach (discussed in
[20]) and we will not elaborate on micromixing models here.
Starting from the transport equation for the joint velocity-scalar mass den-
sity function it is straightforward to derive the equation for the joint velocity-
scalar Favre PDF
f
U
. Both formulations are equivalent. The Eulerian trans-
port equations for the moments of
f
U
can be derived by multiplying each
term by a function of V
i
and
(e.g., V
i
V
j
, V
i
V
j
V
k
,
, or V
i
) and inte-
grating over the phase space (V , ). In this way also a transport equation for
the Reynolds stresses can be derived and it is a good question to ask which
Eulerian second-moment closure model is implied by a choice of closure of the
PDF transport equation, or the other way round, to ask whether a proposed
second moment closure model has a Lagrangian counterpart. Such questions
are discussed in detail in Pope (1994b) and Wouters (1998).
628 Roekaerts
2 Monte Carlo solution method
The PDF transport equation (1.1ad) is a partial dierential equation in many
dimensions. In the case of d spatial dimensions and n independent scalars, solv-
ing the stationary equation by means of a nite-dierence technique would
require a suciently accurate discretization in 2d + n dimensions. Alterna-
tively, in a Monte Carlo solution method, the 2d+n dimensional mass density
function is represented by a large ensemble of N notional uid particles. The
particle properties evolve according to particle models such that the evolution
of the statistics of the particle ensemble corresponds to the modeled PDF evo-
lution. The Monte Carlo method is computationally more ecient already for
the case d = 2, n = 1. Furthermore it oers additional opportunities for model
development. The basics of the Monte Carlo method and several of the numer-
ical algorithms involved are described in detail by Pope (1985), Haworth and
El Tahry (1990), Correa and Pope (1992) and Wouters (1998). This involves
the method of fractional steps, the estimation of mean elds and mean eld
gradients, particle boundary conditions and the integration of particle evolu-
tion equations. The method used here is based on the hybrid Monte Carlo
method of Correa and Pope (1992).
We consider systems whose one-point statistics depend on two spatial co-
ordinates only (d = 2). All Eulerian quantities, like mean velocity and scalar
elds and the mass density function, are dened on a two-dimensional grid.
The Eulerian joint velocity-scalar mass density function is represented by a
spatially equally distributed ensemble of Lagrangian notional uid particles
with properties like position, velocity and composition. Here the term no-
tional particles is used to distinguish from real physical uid elements.
Each of the particles represents an amount of mass m and the particles
are distributed such that the distribution of mass in space corresponds to the
actual density eld. An estimate of the Eulerian Favre joint velocity-scalar
PDF or, equivalently, the joint velocity-scalar mass density function in a cell
is given by the velocities and scalar properties of the particles present in that
cell. Note that the mass density function is dened on a grid but that the uid
particles are not. The particles each have their own position and velocity and
are not bounded to a grid. The estimation of the Eulerian PDF from the
ensemble of Lagrangian particles is illustrated in gure 1.
The particle properties evolve according to a modeled Lagrangian system
of the following general form
dx
i
= U
i
dt (2.1a)
dU
i
= K
u
i
dt +D
u
dW
u
i
(2.1b)
d
= K
dt +D
dW
, (2.1c)
in which K
u
and K
are independent.
The specic expressions for the drift and diusion terms for velocities and
scalars dene the model. Examples of particle velocity evolutions are given
below. Examples of scalar evolutions were given in [20].
The coecients of the particle models contain basically two dierent kinds
of mean quantities. First there are the Eulerian mean quantities, like
U
i
and
. These can be estimated from the particle properties e.g. by the following
expression
=
1
N
c
N
c
i=1
(i)
, (2.2)
in which the sum is over all particles present in a cell. (In practice also smooth-
ing algorithms are used) Second, the modeled terms contain a characteristic
mean turbulent frequency that determines the rate at which the modeled pro-
cesses take place. This frequency cannot be estimated from the particle proper-
ties in the joint velocity-scalar description and it has to be supplied externally.
Another special term is the mean pressure gradient. The mean pressure eld
can be obtained from the velocity and density elds and, in principle, it is
closed. Direct calculation of the mean pressure eld would require a pressure
solver that uses the particle velocity and mass distributions as input. In this
approach, stochastic uctuations in the particle ensemble may cause spurious
pressure uctuations. Here, the mean pressure eld is not calculated directly.
A hybrid Monte Carlo method is employed in which both the pressure and
turbulent frequency elds are solved by an external Eulerian nite-volume
method according to the method of Correa and Pope (1992).
630 Roekaerts
Chemistry Scheme
Lagrangian Algorithm
solve for
particle properties:
x
(i)
, U
(i)
,
(i)
LAGRANGIAN SUBMODEL
Finite Volume Method
solve for:
mean elds
U, p,
u
i
u
j
, k,
EULERIAN SUBMODEL
?
(i)
t
6
(i)
(i)
T
(i)
-
U
p
x
i
k
k
Figure 2: Sketch of the hybrid Monte Carlo method. Left side: Lagrangian
submodel with chemistry solver. Right side: Eulerian submodel that solves for
mean ow eld.
3 Hybrid ow eld model
The hybrid Monte Carlo method of Correa and Pope (1992) consists of an
Eulerian nite-volume submodel and a Lagrangian Monte Carlo submodel.
The hybrid algorithm is illustrated in gure 2.
The Eulerian model solves for the mean velocity, pressure, turbulent kinetic
energy and turbulent dissipation. Given these mean elds, the Lagrangian
Monte Carlo method solves for the joint velocity-scalar mass density function.
Eects of thermo-chemistry are solved in the Lagrangian algorithm. Mean
thermo-chemical properties that inuence the ow eld, like density and molec-
ular viscosity elds, are then coupled back to the nite-volume method which
solves for the new turbulent ow eld. Subsequent submodel calculations are
performed until a stable solution is reached.
In this algorithm, conventional modeling of the Eulerian ow eld equations
is needed also, and closer inspection is needed of the relation between closure
in the Lagrangian and the Eulerian framework.
During the Monte Carlo calculations, by an overall shift and scaling of
particle velocities in each cell, the Eulerian mean velocities and turbulent
kinetic energy are imposed onto the Lagrangian uid particles. As a result of
[21] Joint velocity-scalar PDF methods 631
the hybrid method, the mean velocity and pressure eld correspond to each
other but both are based on the density eld of the previous Monte Carlo
submodel calculation. Density changes seen during a Monte Carlo submodel
calculation have their full impact only after coupling with the nite-dierence
submodel. This can cause slow convergence when many switches between the
submodels are needed and because, in general, Monte Carlo calculations are
very time consuming.
In more recent algorithms, which also calculate the turbulent dissipation
by a Lagrangian model (see section 5), the mean turbulent frequency is de-
termined from the particle properties and the particle representation of the
density eld is directly used to solve for the mean pressure. In ows with
a strong density-ow-eld coupling, these new methods may converge faster
even though these models are more complex. An intermediate method is to
use only mean velocity elds from the nite dierence code and obtain second
moments from the Monte Carlo particles, see Muradoglu et al. (1999).
A basic requirement for the Monte Carlo algorithm is that the particle
ensemble gives a physically correct estimate of the mass density function.
This is expressed mathematically by the requirement that the sum of particle
masses divided by the cell volume, gives a correct representation of the density
eld.
Using a two-dimensional rectilinear grid, large cells correspond to a large
volume in three dimensions. Moreover, in axisymmetrical congurations, cells
near the axis represent a small volume in three dimensions whereas cells far
from the axis correspond to a larger three-dimensional volume. If all particles
are to represent the same amount of mass m, the particle number in a cell
must depend on the local density and on the three-dimensional cell volume.
For example, in cells with a small volume or with low density the number of
particles is lower and the statistical accuracy in these cells is lower. As a result,
the statistical accuracy varies throughout the domain which is undesirable. To
remedy this eect an ensemble with variable particle weights w
(i)
m is used
allowing the use of a nearly constant number of particles per cell. The system
of adjustable particle weights is described in detail by Haworth and El Tahry
(1990). In particular the average in equation (2.2) has to be replaced by a
weighted average.
As explained, the particle density representation on the one hand, and the
mean velocity and pressure eld on the other, are not coupled directly in the
hybrid method. During the Lagrangian submodel calculation there is no phys-
ical mechanism, by means of changes in the mean pressure gradient, to ensure
that the particle weight-to-volume ratio remains constant. Therefore, to en-
force the correct weight to volume ratio, a correction algorithm is needed for
particle weights (or positions). The correction algorithm used in the applica-
tions described below, transfers particles to adjacent cells if the total weight
in the cell is too high.
632 Roekaerts
4 Modeling velocity evolution
4.1 Langevin models
The unclosed terms in the joint velocity-scalar mass density function equation
that describe the evolution in velocity space read
V
i
__
x
i
ij
x
j
U = V , =
_
f
_
, (4.1)
here written in terms of the Favre joint PDF instead of the mass density
function. To illustrate the closure procedure, we rst consider the Simplied
Langevin Model (SLM) (Haworth and Pope 1986). A Langevin model is a
stochastic model for uid particle velocities. The modeled particle velocity
equation consists of a linear drift towards the local Favre mean and an added
isotropic diusion term, and reads
dU
i
=
_
1
2
+
3
4
C
0
_
(U
i
U
i
)dt + (C
0
)
1/2
dW
i
, (4.2)
in which = /k is the mean turbulent frequency, U
i
is a stochastic veloc-
ity realization, C
0
a constant, and dW
i
a stochastic Wiener increment. The
physical meaning of C
0
and the properties of the Wiener process are explained
below, after equation (4.5). Using Ito calculus, the Langevin model then cor-
responds to an evolution of the velocity PDF given by (Risken 1984, Gardiner
1990)
f
U
t
=
_
1
2
+
3
4
C
0
_
V
i
_
(V
i
U
i
)
f
U
_
+
1
2
C
0
f
U
V
2
i
. (4.3)
The basic properties of the SLM are that it yields the correct evolution of
turbulent kinetic energy and that it yields isotropization of the velocity dis-
tribution. To describe complex ow elds an extended model is needed: the
Generalized Langevin model (GLM) which is considered next. Earlier studies
used Langevin equations to model turbulent dispersion and Langevin equa-
tions for particle velocities were developed by assuming an analogy between
uid particles and particles undergoing a Brownian motion. In contrast, the
GLM was derived taking the NavierStokes equations as a starting point. For
an extensive review of the use of Langevin models for turbulence modeling
see Haworth and Pope (1986), Pope (1983) or Dopazo (1994) and references
therein.
From the NavierStokes equations, the equations of motion for a Lagrangian
uid element can be derived. The exact equations read
dx
i
= U
i
dt (4.4a)
dU
i
=
_
p)
x
i
T
ij
)
x
j
_
dt (4.4b)
+
_
x
i
ij
)
x
j
_
dt,
[21] Joint velocity-scalar PDF methods 633
with the Eulerian mean elds p(x, t)) and U(x, t)) evaluated at the uid
element position x(t). The rst term on the right-hand side of equation (4.4b)
can be expressed in terms of the one-point Eulerian velocity statistics but
the second is unknown. The velocity increments are modeled by a Langevin
equation according to
dU
i
=
_
p)
x
i
T
ij
)
x
j
_
dt (4.5)
G
ij
(U
j
U
j
)dt + (C
0
)
1/2
dW
i
,
in which C
0
is a positive constant, W(t) an isotropic Wiener process, the
dissipation of turbulent kinetic energy and G
ij
is a second-order tensor which
is modeled as a function of local mean quantities. A general form of G
ij
in
terms of the local mean velocity gradients, Reynolds stresses and dissipation
was derived by Haworth and Pope (1986). The Wiener process is a stochastic
process with zero mean and covariance
dW
i
dW
j
) = dt
ij
. (4.6)
The rationale of using the Langevin model to model turbulent velocities
is found in the consistency of the model with Kolmogorovs hypothesis of
turbulence (see Monin and Yaglom 1971). For high Reynolds number there is
a separation between the large energy containing scales and the small scales
at which viscous dissipation takes place. The cascade of energy from the large
to the small scales occurs in the inertial subrange where the energy dissipation
rate is the only relevant parameter. The eects modeled by the tensor G
ij
are characterized by the time-scales of mean deformation [U)/x[
1
and
energy dissipation T = k/. The Wiener process models eects that take place
on a shorter time-scale. The Wiener or diusion process is continuous but not
dierentiable and has the properties of a Markov process. This means that
the stochastic velocity increments only depend on the present state of the
uid element and not on its history. The use of a diusion process is justied
because the modeled small-scale processes are not dynamically important as
stated by Kolmogorovs theory. The specic form of the diusion coecient
is determined by the modeled Lagrangian structure function which, for the
Langevin model, reads
U
i
U
j
)
L
(U
i
(t +s) U
i
(t))(U
j
(t +s) U
j
(t)))
L
= C
0
s
ij
, (4.7)
with U
i
the velocity increment of a Lagrangian uid element. This is con-
sistent with Kolmogorovs inertial range scaling with C
0
being a universal
constant. A value C
0
= 2.1 was determined experimentally from diusion
measurements in grid turbulence by Anand and Pope (1985). For a detailed
treatment of dynamical properties of the Langevin model and consistency with
634 Roekaerts
Kolmogorovs theory see Haworth and Pope (1986), Pope (1994a) and Girimaji
and Pope (1990).
An important feature of the Lagrangian models is that they describe the
evolution of individual uid element velocities. This means that, for nite
model coecients, the Langevin model always describes physically realizable
distribution of velocities (e.g., positive normal stresses and turbulent kinetic
energy). Consequently, the Langevin model is always realizable by construction
as are all moment equations derived from the Langevin model.
Because the Langevin model describes the evolution of the full joint velocity
distribution it also provides a modeled evolution equation for the second mo-
ments of velocity. In particular, in the case of an isotropic dissipation tensor
(
ij
=
2
3
ij
), the exact expression for the pressure strain correlation in the
Reynolds stress equation,
ij
=
1
_
u
i
p
x
j
+u
j
p
x
i
_
, (4.8)
is represented by
ij
= G
ik
u
j
u
k
+G
jk
u
i
u
k
+
_
2
3
+C
0
_
ij
. (4.9)
In the approach followed here, following the lines of Pope (1994b), the
model parameters in the expression for G
ij
are specied such that the modeled
pressure-strain correlation equals that of several well-known second-moment
closures.
A basic second-moment closure model for
ij
is the combination of a linear
return-to-isotropy or Rotta term and a linear isotropization-of-production (IP)
term. The model (with xed model constants C
1
= 1.8 and C
2
= 0.6) is re-
ferred to as the Isotropization-of-Production Model (IPM) and the correspond-
ing GLM is called the Lagrangian IP model (LIPM). The Simplied Langevin
model (SLM) corresponds to the Rotta model (only linear return-to-isotropy
C
1
= 4.15, C
2
= 0). For the (L)IPM and the SLM/Rotta model, hybrid Monte
Carlo calculations were performed in a blu-body stabilized diusion ame
using both the Eulerian second-moment model and the Lagrangian Langevin
models (Wouters et al. 1996). Some results of these calculations are presented
below.
4.2 Third-moment and turbulent scalar-ux equations
Since the full joint probability of velocities and scalars is known the Lagrangian
models also provide modeled evolution equations for the third-moments of
velocity and the turbulent scalar uxes. The third-moment equations are
governed by the Langevin model for velocities whereas the combination of
Langevin model and micro-mixing model provides a model for the scalar uxes.
[21] Joint velocity-scalar PDF methods 635
The modeled equations were published by Pope (1994b) and are briey re-
viewed here. The modeled equations are looked upon to assess the validity of
the hybrid Monte Carlo methods.
Modeled triple velocity correlations
The evolution of the triple velocity correlations as modeled by the GLM reads
D
Dt
q
u
r
u
s
=
x
i
i
u
q
u
r
u
s
+
x
j
i
u
j
(
iq
u
r
u
s
+
ir
u
q
u
s
+
is
u
q
u
r
) (4.10)
+
_
G
ij
U
i
x
j
_
(
iq
j
u
r
u
s
+
ir
j
u
q
u
s
+
is
j
u
q
u
r
).
An algebraic expression for the triple moments can be derived from equa-
tion (4.10) by neglecting the mean convective term (the left-hand side) and
approximating the fourth order term by
i
u
q
u
r
u
s
=
u
i
u
r
u
s
+
u
i
u
q
u
s
+
u
i
u
q
u
r
. (4.11)
The resulting algebraic expression can be found in Pope (1994b) and is not
repeated here. For small anisotropies and low mean shear rates the full expres-
sion simplies to the model of Daly and Harlow (1970) (see [10], equation (27))
with the constant C
s
given by 1/(3
1
). For the Simplied Langevin model,
1
= (1/2 + 2/3C
0
) and C
s
takes the value 0.16.
Modeled turbulent scalar-ux equations
The evolution equation for the turbulent scalar ux, as modeled in the La-
grangian method, is given by the combination of GLM and a micro-mixing
model. The concept of micro-mixing was already introduced in [20]. Here
the modeled scalar ux equations are given assuming that the interaction-
by-exchange-with-the-mean (IEM or LSME [20]) mixing model is used. Other
mixing models predict the same evolution of scalar mean and variance. Dier-
ences between micro-mixing models occur in the evolution of third- and higher
moments. The eects on the modeled scalar ux equation are expected to be
small however.
Equating the exact scalar-ux equations with the modeled equations, in the
same way as was done for the Reynolds-stress equations, the modeled terms
read
i
=
_
G
ij
1
2
C
ij
_
u
, (4.12)
with the exact term given by
i
=
x
i
. (4.13)
636 Roekaerts
Here, the assumptions have been made that the mean pressure gradient term
can be neglected and that local isotropy prevails. The assumption of local
isotropy implies that the ux dissipation term
i
is zero.
Alternatively, the micro-mixing term, which models the decay of scalar vari-
ance, can be interpreted to model the dissipation of the scalar ux which would
contradict the assumption of local isotropy (Pope 1994b). By replacing the lo-
cal mean scalar value ) with the conditional mean [U) the scalar ux
model agrees with the local isotropy assumption. More generally speaking,
mixing models must be formulated such that particles interact only with par-
ticles that are within the same range of velocities. This concept agrees with
observations in DNS data where it is seen that mixing occurs in lamella-like
structures that move with the same velocity. For a more detailed treatment of
velocity-conditioned mixing see Fox (1996).
Returning to the scalar-ux pressure-scrambling term, a standard Eulerian
model for the pressure-scrambling term is given by
i
= C
1
+C
2
U
i
x
j
, (4.14)
with standard values of the constants C
1
= 3.0 and C
2
= 0.5 (see Jones
1994). The Lagrangian SLM/IEM model, with C
0
= 2.1 and C
= 2.0, corre-
sponds to the slow part of this Eulerian model but with C
1
= 3.075. Exam-
ples of a scalar ux model using the GLM, including rapid pressure eects, are
given in Pope (1994b) and Wouters et al. (1996) In the hybrid Monte Carlo
method however, scalar transport is governed completely by the Lagrangian
models and the relationship with standard Eulerian models is of no practical
importance.
4.3 Consistency
When solving the modeled velocity PDF transport equation with the hybrid
Monte Carlo method some caution is needed. In the hybrid method the tur-
bulent ow eld is dened twofold: on the one hand by the Eulerian mean
velocity and turbulent kinetic energy elds of the nite-volume submodel,
and on the other hand by the velocity distribution of the Lagrangian Monte
Carlo submodel. Although a unique denition of the turbulence model is not
guaranteed, both submodels use the same solution for the mean velocity and
turbulent kinetic energy elds. Turbulent transport of mean momentum and
production of turbulent kinetic energy are governed by the Reynolds stresses
and consistency between the two submodels on the level of mean velocities
and kinetic energy inevitably requires a consistent treatment of the second
moments.
Several studies, using the hybrid method, have used the k- model in combi-
nation with the SLM (Correa et al. 1994, Nooren et al. 1997). In the Eulerian
[21] Joint velocity-scalar PDF methods 637
k- model, the Reynolds stresses are given by an algebraic relation and the
second-moment equations are not modeled explicitly. No equivalent model in
terms of a Langevin model exists. As a result, the k- model cannot be part
of a consistent hybrid model. In the hybrid method employed here, a consis-
tent hybrid turbulence model is obtained by using the same second-moment
equation in both submodels with the Langevin model chosen equivalent to
the nite-volume turbulence model for the pressure-strain correlation. The
third-moments are still treated inconsistently in the two submodels. In the
Lagrangian method the third-moments are exact whereas in the second mo-
ment closure models these are given by a generalized gradient diusion model.
Eects of this inconsistency on the level of the third moments are assumed to
be small.
The inconsistency at the level of the second moments in the SLM/k-
model and the performance of two consistent hybrid models, SLM/Rotta and
LIPM/IPM, were studied in more detail in a blu-body-stabilized diusion
ame and the corresponding inert ow. Some results of the blu-body ow
calculation are presented in section 7.
Summarizing, in reacting ows where dierences in scalar elds couple back
to the ow eld through the mean density, large dierences can occur between
the standard nite-volume method (using an assumed-shape PDF method),
and the hybrid Monte Carlo method; even when using the same ow model
in the nite volume part of the algorithms. The point made here is that these
dierences do not necessarily stem from the dierent treatment of scalar uc-
tuations but can be caused by a simple dierence of the turbulence models.
In the hybrid model dierent representations for one and the same physical
quantity are used. Firm conclusions on specic model assumptions at the level
of second moment closure can only be made if the two representations that
are used are consistent at that level.
5 Developments in Lagrangian turbulence modeling
In the previous section modeling of the velocity PDF evolution equation has
been treated. This approach to modeling of turbulent ows has its limitations.
First, the turbulent dissipation or turbulent frequency has to be provided. In
the approach used here, the mean frequency is provided by solving the Eulerian
transport equations for the turbulent kinetic energy and turbulent dissipation.
Second, the eects of molecular viscosity and uctuating pressure gradients on
the velocity distribution occur in unclosed form and are modeled by means of
a Langevin model. In this section two extensions to the velocity PDF approach
are reviewed. First the joint velocity-dissipation method is discussed. Second,
the extension of this model by inclusion of the joint statistics of the velocity
wave vector is discussed.
Modeling of homogeneous turbulent ows in terms of the joint velocity-
dissipation PDF was proposed by Pope and Chen (1990). A Langevin equa-
638 Roekaerts
tion for the turbulent dissipation or turbulent frequency is solved. The model
dissipation satises two properties: (1) the dissipation obtains a log-normal
distribution, and (2) the mean dissipation evolves according to the standard
-equation. The model was extended to inhomogeneous ows by Pope (1991).
The main advantage of the model is that it provides a turbulent time-scale to
the modeled velocity evolution now that the joint PDF of velocities and dis-
sipation is known. Second, in inhomogeneous ows the model provides a way
to describe small-scale intermittency eects. In principle this set of modeled
Lagrangian equations is self-contained in the sense that no additional Eule-
rian transport equations have to be solved for turbulent ow quantities. In
an actual Monte Carlo simulation the mean pressure eld is still solved by
means of a pressure solver but the pressure eld is completely determined by
the Lagrangian velocity distribution.
A further development in Lagrangian turbulence modeling is the description
of the ow eld in terms of the joint PDF of velocity dissipation and wave
vector by Van Slooten and Pope (1997). In the rapid distortion limit, where
the eects of mean strain are much larger than viscous dissipation eects,
an exact solution of the NavierStokes equations can be written in terms of
Fourier modes of the velocity eld. To be more specic, an exact solution
exists in terms of the velocity and its wave number vector. The wave number
vector is dened as the unit vector in the direction of the wave vector. In
the predicted second-moment equations the rapid pressure-strain eects now
occur in closed form whereas the slow term has to be modeled. Future studies
have to show if the new model will yield better results in complex ows where
current statistical models have their limitations (e.g. swirling ows).
In principle the PDF description can, like moment closures, be extended
further and sensible choices of included quantities may allow for a better
description of certain physical processes. Of course the demand of practical
applicability narrows down the possible model choices. At this moment the
practical application of the PDF method is restricted to the velocity PDF and
joint velocity-dissipation PDF methods, the former method being a stable and
reliable tool in turbulent reacting ow modeling. The second method allows
for a better ow eld description but is more complicated and has yet to show
its value in complex ow congurations.
6 Chemical reaction
Although the Monte Carlo PDF method allows for an exact treatment of
chemistry from the point of view of turbulence modeling, the use of a full
reaction mechanism would require the integration of a sti system of many
coupled equations. For example a methane-air or natural gas-air diusion ame
requires equations for about 30 species obeying more than 200 chemical reac-
tions. The inclusion of NO-formation or the use of fuels with higher hydro-
carbon species complicates things even further. In the Monte Carlo Method
[21] Joint velocity-scalar PDF methods 639
chemical reaction has to be followed in many thousands of particles and to
make the calculation feasible judicious simplication is needed.
This section deals with some basic modeling of chemistry for methane-air
or natural gas-air diusion ames. A detailed treatment of turbulent combus-
tion chemistry is given in standard textbooks (see Gardiner (1984), Williams
(1985), Warnatz et al. (1996) and references therein).
Making certain assumptions about the ame conditions, full mechanisms
can be simplied to so-called skeletal mechanisms containing about 15 species
and 40 reactions. Usually, skeletal mechanisms are simplied further by making
certain assumptions about some of the elementary reactions. By assuming that
some reactions are in chemical equilibrium or assuming that certain species are
in steady state, reaction can be simplied to several steps (e.g. four or ve step
schemes containing ve or six species respectively). This way of reducing chem-
ical kinetics is described in detail by Peters and Rogg (1993) and Seshadri and
Williams (1994). For the specic case of methane-air combustion see Smooke
(1991). In the rest of this section, conserved-scalar models for methane-air dif-
fusion ame chemistry are discussed. A constrained-equilibrium model, which
is used in this study, is given in more detail. Then, a three-scalar simplied ki-
netics scheme, obtained from the Intrinsic Low-Dimensional Manifold method,
is discussed.
Constrained-equilibrium (CE) model
One of the simplest models to describe diusion ame chemistry is the ame-
sheet model of Burke and Schumann (1928). The model assumes an innitely-
fast irreversible global reaction of fuel and oxidizer to products which results
in piecewise linear relations between composition and mixture fraction. The
ame-sheet model however does not include the formation of intermediate
species like CO and H
2
or radical species like O, H, and OH.
Another simple model that can be expressed in terms of mixture fraction is
the chemical equilibrium model. This model assumes high Damkohler number
so that the reactions are fast enough to reach full chemical equilibrium. For
the case of methane-air combustion this assumption is valid in the high tem-
perature regions of the ame but in the low temperature regions on the rich
side of the ame the slow burn-out of CO does not reach equilibrium and the
CO concentration is underpredicted. This eect can be remedied by assuming
that slow three-body reactions, involving three reactants and two products
or vice versa, do not reach equilibrium. These so-called partial-equilibrium
models perform well in the low temperature regions of the ame.
In diusion ames reaction occurs mainly in a thin reaction zone around
stoichiometric mixture fraction and chemistry is frozen outside this zone. For
hydro-carbon combustion the constrained-equilibrium model of Bilger and
Starner (1983) combines the partial-equilibrium assumptions with the assump-
tion of a reaction zone around stoichiometric mixture fraction. The model as-
sumes a fuel breakdown or pyrolysis sheet at =
ig
>
st
. There all fuel
640 Roekaerts
Figure 3: Relation between fuel mass fraction Y
fu
, intermediate species
mass fraction Y
int
, oxidizer mass fraction Y
ox
, and mixture fraction for the
constrained-equilibrium model. In this example
st
= 0.2,
ig
= 0.38 and the
maximal intermediate species concentration Y
int,max
= 0.2.
reacts by a one-step irreversible innitely fast reaction to some intermediate
species. At the stoichiometric mixture fraction =
st
all intermediate species
are consumed. For methane combustion, Bilger and Starner take C
2
H
4
as the
intermediate species and the pyrolysis sheet is located at
ig
=
st
+0.018. The
relationship between fuel, intermediate species, oxidizer and mixture fraction
is depicted in gure 3.
In the application described below a simplied version of the constrained
equilibrium model is used. Basically, only the ame zone assumption of the
model is used. More details about this constrained-equilibrium model and its
performance in natural-gas diusion ames can be found in Peeters et al.
(1993a) and Peeters (1995).
Intrinsic Low-Dimensional Manifold method
In contrast to the traditional way of simplifying detailed kinetics schemes, the
Intrinsic Low-Dimensional Manifold (ILDM) method (Maas and Pope 1992)
provides a way of simplifying detailed schemes without making a-priori as-
sumptions. Conventional methods assume that certain reactions reach equi-
librium or that certain species are in steady state. This involves processes that
occur on scales much shorter than the scales of turbulence and that can be
decoupled from the rest of the system. The ILDM method performs a time-
scale analysis locally in composition space and automatically denes the slow
and fast time-scales. Several time scales are denoted as slow thereby speci-
fying the dimension of the reduced kinetics scheme. The ILDM method then
assumes that the remaining fast scales are in equilibrium.
[21] Joint velocity-scalar PDF methods 641
Consider a detailed kinetics scheme involving n
s
species and, given the time-
scales of turbulence, the n
f
fastest time-scales can be decoupled. The number
of describing variable or degrees of freedom is then given by n
g
= n
s
n
f
. The
ILDM method performs a time-scale analysis for every point in composition
space based on the eigenvalue-eigenvector representation of the local Jacobian
of the system. If the eigenvalues of the system are in descending order then the
rst n
g
eigenvectors of the system dene the low-dimensional manifold. The
other eigenvalues of the system have a lower negative real part which causes
a fast relaxation of the system to the manifold.
More details about the application of the ILDM method to methane-air
combustion and the validation of the technique in laminar ames are given
in Schmidt et al. (1996).
In Situ Adaptive Tabulation method
Another development in turbulent reacting ow modeling is the In Situ Adap-
tive Tabulation (ISAT) method (Pope 1997). This new algorithm allows for the
use of detailed chemistry schemes, like skeletal mechanisms with 15 or more
species, in PDF simulations of turbulent ames, keeping the computational
demands within reasonable limits.
The idea behind the ISAT algorithm is to integrate the reaction system
directly and to store the outcome of the integration for later use. Other points
in scalar space, with a small distance to points previously used, do not require
a new integration but obtain the reaction increment by an interpolation in
the table. Reaction increments that cannot be interpolated with sucient
accuracy require a new integration and are stored in the table. As a result,
only the accessed region in scalar space is stored in the table. Details on
the algorithm, its numerical implementation and performance can be found
in Pope (1997) and references therein. Applications have been reported in Yang
and Pope (1998) and Saxena and Pope (1998). A promising perspective is the
combination of ISAT and the use of augmented reduced mechanisms (Sung
et al. 1998, James et al. 1999). The computational advantage oered by the
ecient storage and lookup algorithm can be exploited by doing less reduction,
keeping suciently detailed chemistry at an aordable computing time.
7 Modeling of a blu-body-stabilized diusion ame
7.1 Introduction
To illustrate the theoretical developments described above, this section re-
ports on the study of a blu-body-stabilized methane-air diusion ame using
second-moment closures and multi-scalar chemistry. The blu-body congu-
ration serves as a model for industrial type low NO
x
burners. Here the blu-
body ame is chosen as a test case because the ow pattern exhibits complex
phenomena, like recirculation regions and stagnation points, and because the
coupling between turbulence and chemistry is strong. The ow is very sensitive
642 Roekaerts
to density uctuations and the ame shows eects of nite-rate chemistry. The
congurations studied here were subject of a 1994 ERCOFTAC-SIG workshop
(First ASCF Workshop 1994a,b). It was shown that the k- model is insuf-
cient to model the blu-body ow characteristics. Here two second-moment
closure models are used besides the standard k- model. Chemistry is modeled
with the conserved-scalar CE model and with the three-scalar ILDM scheme.
The objective is to make a comprehensive study of the relative importance
of the model used for convection, for micro-mixing, and for chemical kinetics.
Monte Carlo predictions are compared to available experimental data of mean
velocities, turbulent kinetic energy and temperature.
The Monte Carlo model was implemented in a computer code PDF2DS pro-
vided by S.B. Pope. (The Delft version of this code is called PDFD). The nite-
volume submodel calculations are performed with BIGMIX which is an in-house
nite-volume code of TU Delft (see Peeters 1995).
7.2 Test case specication
The ow congurations studied are an axisymmetric blu-body-stabilized
methane-air diusion ame and the corresponding inert ow which were the
subject of a 1994 ERCOFTAC-SIG workshop (First ASCF Workshop 1994a,b).
Also other authors have reported their results in the literature (e.g. Fallot et
al. 1997, Warnatz et al. 1996). The conguration consists of a fuel jet 5.4 mm
in diameter, surrounded by a 50-mm-diameter blu body. The coaxial air inlet
has an outer diameter of 100 mm and a low-velocity coow of air is present
around the air inlet. The thickness of the coaxial air pipe is 1 mm. Figure 4
depicts the blu-body conguration and sketches a typical ow pattern.
In the experiment, the burner is unconned but in the simulation, a solid
wall is assumed at a radius of 200 mm to prevent numerical problems with
outow of particles in the Monte Carlo method. Test calculations have shown
that the presence of this connement does not inuence the ow in the regions
of interest.
The inlet velocities for the fuel, air and coow are 21 m/s, 25 m/s and 1 m/s,
respectively. The fuel inlet Reynolds number based on the fuel inlet diame-
ter is Re = 7000 and for the air inlet, based on the blu-body diameter,
Re = 80 000. Initially boundary conditions were the same as prescribed at
the SIG workshop (First ASCF Workshop 1994a). However, the original inlet
conditions are modied to improve the predictions with respect to the length
of the recirculation zone, and to obtain a stable convergence of the second-
moment equations. Inlet prole calculations are set up to approximate the
experimental conguration.
The nite-volume calculations are performed on a rectangular nonuniform
mesh of 150 150 points. On this mesh grid independent solutions are ob-
tained and dierences with solutions on an 8080 mesh are small. The Monte
Carlo simulations are performed on a 80 80 mesh using 150 particles per
[21] Joint velocity-scalar PDF methods 643
5.4 mm
50 mm
100 mm 1 mm
U
1
U
0
S
S
D
F
F
A
B
Figure 4: Blu-body ow conguration and characteristic ow pattern. Mean
fuel jet velocity
U
0
= 25 m/s, uniform air inlet velocity
U
1
= 21 m/s. Bold
faced letters indicate: S stagnation regions; D diusion ame region; F main
reaction or ame region; A and B scatterplot positions of gures 8 and 9.
cell. To reduce stochastic uctuations, mean properties are averaged over 500
iterations. Additional Monte Carlo calculations with 150 particles per cell on
a grid of 150 150 and with 400 particles per cell on a grid of 80 80, are
performed to test the inuence of grid dependence and the statistical accuracy.
Monte Carlo simulations require about 200 Mb internal memory and 100 hours
of CPU time on a HP735 workstation or 10 hours on 8 PEs of a massively
parallel CRAYT3E supercomputer.
7.3 Choice of PDF model
This section summarizes the models used in this study and gives a motivation
for the specic model choices.
Hybrid turbulence model
Three dierent hybrid turbulence models have been used. The hybrid SLM/k-
model is used because it is a standard turbulence modeling approach using
the hybrid Monte Carlo method (see Correa et al. 1994, Nooren et al. 1997).
In this ow the inconsistency of this model is apparent (see Wouters 1998).
Two consistent hybrid models used are the SLM/Rotta and the LIPM/IPM
models. The SLM/Rotta model is used because it is seen as the simplest con-
sistent hybrid turbulence model. The consistent LIPM/IPM model is used
644 Roekaerts
because the second-moment closure model is seen as a basic Reynolds-stress
model and it is expected to perform reasonably well for the blu-body cong-
uration.
In the inert ow, several other second-moment models were evaluated, with-
out reaching signicant improvement of the predictions.
The approach of tuning the turbulence model to the experimental ow data
separately for every chemistry and mixing model is not used here. Rather, the
turbulence model constants are kept constant while investigating the relative
performance of dierent micromixing and chemistry models. In the reacting
case, ow and scalar elds are strongly coupled and changes in the scalar eld
model will directly aect the ow eld. If the tting procedure is employed,
then changes in the turbulence model constants will be large, provided an
acceptable choice for the constants even exists.
Chemistry modeling
Initially, chemical reaction is simplied using the conserved-scalar constrained-
equilibrium (CE) model. It is assumed that the reaction is fast such that it is
limited only by the mixing of fuel and oxidizer.
CE predictions however, show an overprediction of the mean temperature
and, moreover, the ame length is overpredicted severely. Because of the strong
recirculation, burned products ow back into the inlet region where the mix-
ing rate of species is large. There, eects of partial premixing and nite-rate
kinetics are important. To capture these eects, a three-scalar ILDM reduced
kinetics scheme is used. The ILDM scheme can successfully describe nite-
rate kinetics eects and it is able to describe the pure diusion limited case
correctly.
Here, for the case of non-premixed methane-air combustion the detailed
mechanism is simplied to a three-dimensional manifold, parameterized by
mixture fraction and the H
2
O and CO
2
mass fractions. The remaining thermo-
chemical variables like the other mass fractions, temperature and density, and
the reaction rates of H
2
O and CO
2
are tabulated as a function of these three
parameters. In order to avoid a CPU-intensive integration of the rate equations
for H
2
O and CO
2
mass fractions during the Monte Carlo simulations, the
reaction rates are pre-integrated for the specic time-step of the simulation.
For all simulations performed in this study, thermo-chemical variables are
tabulated as a function of the describing variables in a locally rened table as
described in Peeters et al. (1993b) and Peeters (1995).
Scalar micro-mixing
Using the conserved-scalar CE chemistry model, modeling is required for
micro-mixing of a single inert scalar. For this case, four mixing models are
available: IEM (also called LSME), coalescence-dispersion (C/D), binomial
Langevin (BL) and mapping closure (MCM), which were introduced in sec-
tions 4.1, 4.2, 4.4 and 4.5 of [20].
[21] Joint velocity-scalar PDF methods 645
With ILDM chemistry, scalar mixing involves mixture fraction and two ad-
ditional reacting scalars, Y
CO
2
and Y
H
2
O
. The C/D model performs very well
for inert single-scalar mixing in jet ows (see Wouters et al. 1998) and it has
been shown to perform well in reacting jet diusion ames with CE and with
ILDM chemistry (Nooren et al. 1997). For these reasons, the C/D model is
selected as a standard model for mixing of multiple reacting scalars. A second
model that is selected for scalar mixing in combination with ILDM chemistry
is the modied multi-scalar BL model which was described in Wouters (1998).
The model parameter K, which also enters the one-scalar case discussed in
[20] takes a standard value of 0.3 but here it is varied between 0.1 and 2.0 to
study its eect on the overall reaction. The constant C
= 4 are
performed.
7.4 Results for reacting ow
Results for the reacting ow are given in the following order: rst a summary
of ow eld predictions is given, next results for thermo-chemical elds of the
conserved-scalar constrained-equilibrium model are shown and the limitations
of this model are discussed. Then, improvements using a three-scalar ILDM
reduced kinetics scheme are presented. Finally results on the performance of
micromixing models are given. A study of the inert ow, including consistency
tests can be found in Wouters (1998).
Flow eld predictions
Experimentally, the ow exhibits one stagnation point, at x = 70 mm with a
peak in turbulent kinetic energy of k = 44 m
2
/s
2
. All three hybrid models in
combination with CE chemistry and C/D micromixing, predict a minimum
in the axial velocity for x 70 mm with the minimal value close to zero.
The peak value of turbulent kinetic energy is underpredicted by all models.
With the IPM, predictions are reasonable for x < 70 mm. Here, the initial
decay of the axial velocity is predicted correctly. At higher axial distances, the
performance of the model is unclear. In this region the mean temperature is
overpredicted severely (see gure 5) which clearly has its eect on the ow
eld. Here, the limitations of the CE chemistry model play a role.
Constrained-Equilibrium results
Figure 5 depicts the axial prole of the mean temperature. Moving along
the axis in the axial direction, the experimental data show a sharp rise in
temperature, up to a peak temperature of 1667 K at x = 80 mm. After the peak
temperature is reached, the temperature drops rapidly. CE predictions are in
reasonable agreement with the experiments up to x = 80 mm but the predicted
temperature rises further and reaches a maximum of 2020 K at x = 145 mm.
Moreover, the high temperature zone extends over a much larger region.
646 Roekaerts
Figure 5: Axial proles of temperature. Predictions use IPM turbulence model
and C/D micro-mixing model. Symbols: CE chemistry model (T
max
=
2020 K); ILDM chemistry model (T
max
= 1980 K); measurements
(T
max
= 1667 K).
Figure 6: Radial proles of mean temperature. Predictions with IPM turbu-
lence model and C/D micro-mixing model. (a) CE chemistry, (b) ILDM chem-
istry. Lines: predictions, x = 10mm, x = 30mm, x = 50mm.
Symbols: measurements, x = 10mm, x = 30mm, x = 50mm.
Radial proles of temperature, which are depicted in gure 6a, show that
predictions are reasonable for x 50 mm.
At x = 10 and 30 mm the temperature is overpredicted by 250 K but the
shape of the proles is good. At r = 25 mm a peak indicates the presence of
a reaction zone starting at the edge of the blu-body. There the assumption
of a diusion ame, that can be described by a conserved-scalar model, seems
[21] Joint velocity-scalar PDF methods 647
to be reasonable. At x = 50 mm, the temperature maximum is overpredicted
because the temperature at the centerline rises too early (see also gure 5).
At higher axial distances near the stagnation zone, mixing rates are high
and a well-mixed composition of fuel and air exists above the recirculation
zone. Here, reaction is not limited by mixing and eects of nite-rate kinetics
are important. The conserved-scalar CE model is not able to capture these
eects and therefore it overpredicts the temperature in this region.
ILDM results
Final results of the ERCOFTAC-SIG workshop (First ASCF Workshop 1994b)
show that models that do not employ the fast-chemistry assumption perform
much better downstream of the recirculation zone, even when using k- turbu-
lence modeling. To satisfactorily describe the eects of partial premixing and
nite-rate kinetics, a three-scalar ILDM reduced kinetics scheme is used.
In gure 5, the axial prole of mean temperature is depicted for CE and
ILDM chemistry models and measurements, and the implications of the chem-
istry model are clearly seen. ILDM chemistry yields a faster axial decay of
temperature after the peak temperature is reached. The CE model overpre-
dicts the temperature in this region even though the mixture fraction elds in
the CE and ILDM calculations are almost the same. At x 45 mm, the ILDM
prediction shows a sharp rise in temperature which is not seen in the measure-
ments or in the CE predictions. The faster ignition is caused by dierences in
the ow eld. Because of the strong coupling between chemistry and ow eld,
small dierences in the mean density result in a much stronger recirculation
with a minimal axial velocity on the axis of 3.9 m/s at x = 50 mm.
Grey-scale plots of the mean temperature for CE and ILDM predictions and
the measurements are shown in gure 7. Looking at the overall picture, the
ILDM ame is more compact and it is located closer to the blu body than seen
in the experiments. Compared to the CE predictions the ame length is much
shorter and the predictions are in better agreement with the experiments.
However, the temperature is still overpredicted by approximately 200 K in
the ame region. The reason for the improvements obtained with the ILDM
chemistry scheme over the CE model is that the fast-chemistry assumption,
made in the CE model, is not valid in this ame.
To illustrate the dierences between the two chemistry models, gure 8
shows scatter plots of mixture fraction and H
2
O mass fraction at two positions
in the ame.
Scatter plot positions are sketched in gure 4. Figure 8a shows a scatter
plot at x = 37.5 mm and r = 7.86 mm (position A) in the region where the
mixture fraction variance and mixing rate reach a maximum. A scatter plot
at x = 38 mm and r = 15 mm (position B) is shown in gure 8b. This posi-
tion is located in the shear layer above the edge of the blu body at an axial
location where the source terms of Y
CO
2
and Y
H
2
O
reach a local maximum. At
648 Roekaerts
(a) (b) (c)
300 K 1900 K
Figure 7: Greyscale plot of the mean temperature. (a) CE chemistry predic-
tions (T
max
= 2020 K), (b) measurements (T
max
= 1667 K), (c) ILDM chem-
istry predictions (T
max
= 1980 K). Contour values at 300, 500, . . . 1900 K. For
the calculations only part of the domain is shown.
position A, the mixture fraction variance is large and almost the entire mix-
ture fraction space is occupied. For > 0.25 points are scattered around the
equilibrium line but for lower mixture fraction values, many points are found
below the equilibrium limit. At this position, the ILDM model predicts a mean
temperature below the equilibrium value. Position B is located in the shear
layer above the edge of the blu body where the CE results showed that the
assumption of diusion limited chemistry gives reasonable predictions of the
mean temperature (see also gure 6a). The scatter plot shows that the ILDM
results are far from equilibrium. The gure clearly shows that mixing rates are
high which yields many points near stoichiometry but with low Y
H
2
O
values.
[21] Joint velocity-scalar PDF methods 649
Figure 8: Scatterplot of H
2
O mass fraction versus mixture fraction. ILDM
chemistry with C/D micro-mixing. (a) at x = 37.5 mm and r = 7.86 mm,
takes the standard value 2. To investigate the eects of the mixing rate, a
calculation with the ILDM chemistry model and the C/D mixing model was
performed using C
= 2. Summarizing, for
this ame, no increase of the mixing rate is needed to obtain ame stabi-
lization and the temperature overprediction is larger using a higher mixing
rate. The eects of the mixing rate on the predictions is not studied further
here.
Micro- and macromixing The remaining deciencies of the present PDF
model are attributed to modeling of momentum and scalar transport. Scalar
micro-mixing and the coupling between micro-mixing and reaction are not
crucial in this ow. Apparently, macro-mixing by turbulent convection is the
driving force in scalar mixing. The underprediction of the turbulent kinetic
energy at the stagnation point positions and the low Reynolds numbers of the
fuel and air inlets (Re = 7000 and Re = 48 000 respectively) indicate that
this ow exhibits unstable phenomena. A low-frequency oscillation in the ow
eld causes an overprediction of the measured turbulent kinetic energy. The
instationary convective scalar transport cannot be described by the present
stationary Favre-averaged PDF model.
7.5 Discussion
Although the ame predictions with the ILDM chemistry scheme show a large
improvement over the conserved-scalar CE predictions, results are still far from
perfect. The remaining deciencies of the present PDF model are attributed
to the modeling of transport of scalars and momentum. Our study of scalar
micro-mixing shows that this eect is not crucial in this ow. Apparently here
turbulent transport or turbulent diusion is the driving mechanism in scalar
mixing. The large underprediction of turbulent kinetic energy at the stagnation
point positions and the low Reynolds numbers of the fuel and air jets indicates
652 Roekaerts
that this ow exhibits unstable phenomena which cannot be described by our
stationary Favre-averaged PDF model.
Since the calculation of an instationary ow using Monte Carlo PDF meth-
ods is still limited by the available computer power this test case cannot
be described accurately by the velocity-scalar hybrid PDF model. A Eule-
rian scalar-PDF transport method does not have these limitations. In cases
where the turbulent scalar uxes follow simple gradient-diusion laws, Eule-
rian scalar-PDF transport methods can be used to perform instationary cal-
culations. Otherwise, the self-contained joint velocity-dissipation-scalar PDF
method, or Large Eddy Simulation and ltered-PDF techniques of Colucci et
al. (1998), are needed to perform instationary PDF simulations.
References
Anand, M.S. and Pope, S.B. (1985) Diusion behind a line source in grid turbulence.
In Turbulent Shear Flows 4, L.J.S. Bradbury et al (eds.), Springer-Verlag, 4661.
Bilger, R.W. and Starner, S.H. (1983) A simple model for carbon monoxide in
laminar and turbulent hydrocarbon diusion ames, Comb. Flame 51 155176.
Burke, S.P. and Schumann, T.E.W. (1928) Diusion ames, Indus. Engin. Chem.
20 9991004.
Colucci, P.J., Jaberi, F.A., Givi, P. and Pope, S.B. (1998) Filtered density function
for large eddy simulation of turbulent reacting ows, Phys. Fluids 10 499515.
Correa, S.M. and Pope, S.B. (1992) Comparison of a Monte Carlo PDF/nite-
volume mean ow model with blu-body Raman data, Proc. Combustion Institute
24 279285.
Correa, S.M., Gulati, A. and Pope, S.B. (1994) Raman measurements and joint PDF
modeling of a nonpremixed blu-body-stabilized methane ame, Proc. Combustion
Institute 25 11671173.
Daly, B.J. and Harlow, F.H. (1970) Transport equations in turbulence, Phys. Fluids
13(11), 26342649.
Dopazo, C. (1994) Recent developments in PDF methods. In Turbulent Reacting
Flows, P. Libby and F. Williams (eds.), Academic Press, 375474.
Fallot, L., Gonzalez, M., Elamraoui, R. and Obounou, M. (1997) Modelling nite-
rate chemistry eects in non-premixed turbulent combustion: test on the blu-body
stabilized ame, Comb. Flame 110 298318.
First ASCF Workshop (1994) Steady-state combustion chambers and furnaces. Test
case specications, ERCOFTAC-SIG.
First ASCF Workshop (1994) Steady-state combustion chambers and furnaces.
Final results, ERCOFTAC-SIG.
Fox, R.O. (1996) On velocity-conditioned scalar mixing in homogeneous turbulence,
[21] Joint velocity-scalar PDF methods 653
Phys. Fluids 8(10), 26782691.
Gardiner, W.C. (1984) Introduction to combustion modeling. In Combustion Chem-
istry, W. Gardiner (ed.), Springer-Verlag, 119.
Gardiner, C.W. (1990) Handbook of Stochastic Methods, second edition. Springer-
Verlag.
Girimaji, S.S. and Pope, S.B. (1990) A diusion model for velocity gradients in
turbulence, Phys. Fluids A 2(2), 242256.
Haworth, D.C. and El Tahry, S.H. (1990) Probability density function approach for
multidimensional turbulent ow calculations with application to in-cylinder ows
in reciprocating engines, AIAA Journal 29(2), 208218.
Haworth, D.C. and Pope, S.B. (1986) A generalized Langevin model for turbulent
ows, Phys. Fluids 29 387405.
James, S., Anand, M.S., Razdan, M.K. and Pope, S.B. (1999) In situ detailed
chemistry calculations in combustor ow analysis. In 44th ASME Gas Turbine
and Aeroengine Technical Congress. Indianapolis.
Jones, W.P. (1994) Turbulence modelling and numerical solution methods for vari-
able density and combusting ows. In Turbulent Reacting Flows, P. Libby and
F. Williams (eds.), Academic Press, 309374.
Maas, U. and Pope, S.B. (1992) Simplifying chemical kinetics: Intrinsic low-dimen-
sional manifolds in composition space, Comb. Flame 88(3), 239264.
Masri, A.R. and Pope, S.B. (1990) PDF calculations of piloted turbulent non-
premixed ames of methane, Comb. Flame 81(1), 1329.
Monin, A.S. and Yaglom, A.M. (1971) Statistical Fluid Mechanics. MIT Press.
Muradoglu, M., Jenny, P., Pope, S.B. and Caughey, D.A. (1999) A consistent hybrid
nite-volume/particle method for the PDF equations of turbulent reactive ows,
J. Comp. Phys. 154 342371.
Nooren, P.A., Wouters, H.A., Peeters, T.W.J., Roekaerts, D., Maas, U. and Schmidt,
D. (1997) Monte Carlo PDF modelling of a turbulent natural-gas diusion ame,
Comb. Theor. Modelling 1(1), 7996.
Nooren, P.A. (1998) Stochastic modeling of turbulent natural-gas ames. PhD thesis,
Delft University of Technology.
Peeters, T.W.J., Roekaerts, D. and Hoogendoorn, C.J. (1993) Modelling of turbulent
non-premixed ames. Part 1. Conserved-scalar chemistry models. Technical Report
BCWT.93.1, Technische Universiteit Delft.
Peeters, T.W.J., Roekaerts, D. and Hoogendoorn, C.J. (1993) Modelling of turbulent
non-premixed ames. Part 4. Adaptive property tabulation. Technical Report
BCWT.93.4, Technische Universiteit Delft.
Peeters, T.W.J. (1995) Numerical modeling of turbulent natural-gas diusion ames.
PhD thesis, Delft Universitity of Technology.
654 Roekaerts
Peters, N. and Rogg, B. (1993) Reduced Kinetic Mechanisms for Applications in
Combustion Systems, Lecture Notes in Physics m15. Springer-Verlag.
Pope, S.B. and Chen, Y.L. (1990) The velocity-dissipation probability density func-
tion model for turbulent ows, Phys. Fluids A 2(8), 14371449.
Pope, S.B. (1983) A Lagrangian two-time probability density function equation for
inhomogeneous turbulent ows, Phys. Fluids 26 34483450.
Pope, S.B. (1985) PDF methods for turbulent reactive ows, Prog. Comb. Theory
Comb. Sci. 11 119192.
Pope, S.B. (1991) Application of the velocity-dissipation probability density function
model to inhomogeneous turbulent ows, Phys. Fluids A 3(8), 19471957.
Pope, S.B (1994) Lagrangian PDF methods for turbulent ows, Ann. Rev. Fluid
Mech. 26 2363.
Pope, S.B (1994) On the relationship between stochastic Lagrangian models of
turbulence and second-moment closures, Phys. Fluids 6(2), 973985.
Pope, S.B. (1997) Computationally ecient implementation of combustion chemistry
using in situ adaptive tabulation, Comb. Theor. Modelling 1(1), 4163.
Risken, H. (1984) The FokkerPlanck Equation. Springer-Verlag.
Saxena, V. and Pope, S.B. (1998) PDF calculations of major and minor species in
a turbulent piloted jet ame. In Proc. Combustion Institute 27 10811086.
Schmidt, D, Riedel, U., Segatz, J., Warnatz, J. and Maas, U. (1996) Simulation
of laminar methane-air ames using automatically simplied chemical kinetics,
Comb. Sci. Tech. 113-114 316.
Seshadri, K. and Williams, F.A. (1994) Reduced chemical systems and their ap-
plication in turbulent combustion. In Turbulent Reacting Flows, P. Libby and
F. Williams (eds.), Academic Press, 153210.
Smooke, M.D. (1991) Reduced Kinetic Mechanisms and Asymptotic Approximations
for Methane Air Flames, Lecture Notes in Physics 384. Springer-Verlag.
Sung, C.J., Law, C.K. and Chen, J.Y. (1998) An augmented reduced mechanism
for methane oxidation with comprehensive global parametric validation, Proc.
Combustion Institute 27 295304.
Van Slooten, P.R. and Pope, S.B. (1997) PDF modeling of inhomogeneous turbulence
with exact representation of rapid distortions, Phys. Fluids 9 10851105.
Warnatz, J., Maas, U. and Dibble, R.W. (1996) Combustion. Springer-Ferlag.
Williams, F.A. (1985) Combustion Theory, second edition. Benjamin/Cummings.
Wouters, H.A., Nooren, P.A., Peeters, T.W.J. and Roekaerts, D. (1996) Simulation
of a blu-body-stabilized diusion ame using second-moment closure and Monte
Carlo methods, Proc. Combustion Institute 26 177185.
Wouters, H.A., Nooren, P.A., Peeters, T.W.J. and Roekaerts, D. (1998) Eects of
[21] Joint velocity-scalar PDF methods 655
micro-mixing in gas-phase turbulent jets, Int. J. Heat Fluid Flow 19 201207.
Wouters, H.A. (1998) Lagrangian models for turbulent reacting ows. PhD thesis,
Delft University of Technology.
Yang, B. and Pope, S.B. (1998) Treating chemistry in combustion with detailed
mechanisms in situ Adaptive Tabulation in Principle Directions premixed com-
bustion, Comb. Flame 112 85112.
Part C.
Future Directions
22
Simulation of Coherent Eddy Structure in
Buoyancy-Driven Flows with Single-Point
Turbulence Closure Models
K. Hanjalic and S. Kenjeres
1 Introduction
A major deciency of conventional single-point closure models that are used
in conjunction with Reynolds-averaged NavierStokes methods (RANS) is in
their inability to account for a wide range of turbulence time and length scales,
which characterize various mechanisms in turbulence dynamics. By their na-
ture, RANS methods conceal any spectral and structural information, and
have been regarded as unsuitable for detecting any identiable eddy struc-
ture. This, in turn, prevents accounting for any spectral features, interactions
between eddies of dierent scales and energy transfer through the spectrum.
Yet, it is known that the spectral dynamics is dierent in every ow and inu-
ences the gross turbulence features. Single-point RANS closures compensate
for these deciencies partially by various additions to the basic model to ac-
count nominally for departure from the simple ows in which they were tuned
and thus, indirectly, for spectral non-equilibrium.
A way to partially overcome the problem is to introduce one or more ad-
ditional turbulence scales by which to represent distinct parts of the turbu-
lence spectrum. Such multi-scale models have been proposed e.g. by Hanjalic,
Launder and Schiestel (1980), Schiestel (1987) and others. Here the turbu-
lence energy spectrum is divided into two or more parts, and the transport
equations for turbulent kinetic energy (or the full stress tensor) and for the
energy dissipation rate are solved for each spectral slice, including also their
mutual interaction. Other multiple-scale models have also been proposed in
the literature, with moderate success. A major problem appears in the large
number of unavoidable empirical coecients, which are dicult to determine
because of the lack of information on spectral dynamics in various ows. More
advanced spectral models, that account fully for spectral dynamics have also
been proposed (e.g. Besnard et al. 1996, Touil et al. 2000). These models look
promising, but have not yet been widely tested and it is uncertain how suitable
and attractive they are for predicting real complex ows.
A particular problem arises when the ow is dominated by large coherent
vortical structures. These large-scale eddies can be discerned in many tur-
bulent ows whether in the laboratory, in nature or in industrial appliances.
659
660 Hanjalic and Kenjeres
They act as a major carrier of momentum, heat and species. In wall bounded
non-isothermal ows, where the walls are at dierent temperatures, these ed-
dies often provide the major communication link between the bounding walls
and thus control the exchange of heat between them. Often, as in the case
of vortex shedding, in ows driven by buoyancy, or subjected to a magnetic
eld or system rotation, the large structure has a deterministic character. Such
vortical structures also appear in laminar ows, but in turbulent regimes it is
not always clear whether these structures should be regarded as true turbu-
lence (smooth spectrum and probability density functions), or whether they
should be interpreted as a form of mean motion with inherent, but determin-
istic (organized) unsteadiness. While the form of secondary motion can be
captured with advanced (e.g. second-moment) closures, the convective trans-
port by large turbulent eddies, usually referred to as turbulent diusion, is
modelled by gradient hypotheses, which is unrealistic for representing such a
process. Because of their importance in governing the turbulence dynamics,
the only way to account fully for their eects is to resolve the large-scale eddies
in space and time.
Large-eddy structures can be fully resolved by Large Eddy Simulation (LES)
techniques and these have long been viewed as the future preferred computa-
tional method that should soon replace the RANS approach as an industrial
tool. While conventional LES has proven to be a very useful method in study-
ing turbulent ows, over the years it has shown some serious limitations. The
major problem is in the excessive computational costs which increase sharply
with the Reynolds and Rayleigh numbers, associated with the need to resolve
a larger and larger range of eddy scales. Handling complex geometries, the
treatment of wall boundaries and the resolution of near-wall ow regions in
wall-bounded ows are other challenges that have not yet been fully resolved.
The latter issue is especially critical if wall friction and heat transfer are in
focus. A proper resolving of the buer region and capturing the streaky struc-
tures and small-scale eddies occupying this region require a very ne computa-
tional mesh not only in the wall-normal direction, but also in the spanwise and
streamwise directions, imposing formidable requirements on the mesh density,
which becomes comparable with that for a complete direct numerical simu-
lation (DNS). In fact, most LES reported in the literature have reproduced
successfully the wall friction factor and heat transfer coecients by employing
a very ne mesh so that the subgrid-scale stresses are of the same order as
the viscous stresses. The use of wall-functions to bridge the near-wall region,
as is done in RANS methods, appears to be a possible compromise, but so far
without much success.
These features make the LES still inapplicable to solving technological ows
at very high Reynolds and Rayleigh numbers. To quote Speziale (1998) the
traditional LES has never really lived up to its promise.
[22] Simulation of coherent eddy structure 661
2 VLES and Hybrid RANS/LES methods
In order to overcome the cost problems and still resolve large-scale structures
at least in ows where they play a dominant role, several intermediate or hy-
brid methods have been proposed recently, aimed at utilizing the advantages
of both approaches: the simplicity and computational eciency of RANS and
the potential of LES to fully resolve a large-scale part of turbulence spectrum.
These methods can generally be classied as Very-Large Eddy Simulation
(VLES), the name implying essentially a form of LES with a cut-o lter at
much lower wave number. This means: resolve less and model more! Model-
ing a larger part of the spectrum requires a more sophisticated model than
the standard sub-grid-scale model, i.e. a form of RANS model that is not re-
lated to the size of the numerical mesh. The solution of the resolved part of
the spectrum can either follow the traditional LES practice using grid size as
a basis for dening the lter (hence the name hybrid RANS/LES), or solve
ensemble- or conditionally-averaged equations which implicitly involve time
ltering.
Ha Minh and Kourta (1993) (also Aubrun et al. 1999) proposed, what
they called a semi-deterministic method (SDM), based on the decompo-
sition of physical variables into a coherent part, obtained by an ensemble-
averaging, and an incoherent part which is the residue of the ensemble-average
operator, see Fig. 1. The coherent part is resolved by time integration of
the three-dimensional ensemble-averaged NavierStokes equations (having the
same form as the Reynolds-averaged equations), in which the ensemble-ave-
raged (incoherent) stresses are provided from a conventional single-point
low-Re-number k- model. However, in the ow which they considered (the
backward-facing step), such a model gave a too high turbulent viscosity that
suppressed instabilities and unsteadiness and failing to resolve large eddy
structure, resulted in a steady solution. For that reason, the eddy viscosity
was arbitrarily reduced by decreasing the value of the coecient C
from
the conventional value of 0.09 to 0.06. This problem may be associated with
the particular turbulence model used, which may be too dissipative: Tatsumi
et al. (1999) succeeded in obtaining unsteady solutions and captured large
eddy structures behind a step using a more advanced non-linear eddy-viscosity
model without any re-tuning of the model.
Unsteady computations of ows around blu bodies with eddy shedding,
even in cases when this shedding is not periodic but stationary over long-term
average (quasi-steady recirculation), using conventional single-point closure
with a suciently small time step, can also be classied as a form of SDM.
Such an approach essentially implies a separation of time scales: those of the
shedding eddies and those of the rest of turbulence. Such computations, with
a proper accounting for the contribution of the shedding eddies to the second-
moments (apparent stresses), which may even outweigh the modelled ones,
have produced much better agreement with experiments than steady compu-
662 Hanjalic and Kenjeres
Figure 1: Original sketch of spectrum decomposition from Ha Minh and Kourta
(1993)
tations with the same model (e.g. Johansson et al. 1993, Durbin 1995, Bosch
and Rodi 1998).
More recently, Spalart (1999) (also Travin et al. 1999, Nikitin et al. 2000)
proposed a detached-eddy simulation (DES) method as a hybrid RANS/LES
technique in which a single turbulence model (of any level) is used as a sub-
grid-scale model in regions away from solid boundaries, and as a RANS model
in the wall boundary layer. The transition from the RANS to LES regions is
continuous and smooth, activated automatically when the distance from the
wall becomes larger than the prescribed criterion here the largest dimen-
sion of the grid cell. In the LES region, the method requires grid spacing in
all directions to be ne enough to allow traditional large eddy simulations,
whereas in the RANS region (close to a wall) the grid renement is needed
only in the wall-normal direction, as required by the turbulence model ap-
plied. This technique was originally aimed at external aerodynamics at high
Reynolds numbers, with thin boundary layers and massive separation regions,
but seems applicable to other types of ow. Although it brings signicant cost
saving by allowing near-wall spanwise and streamwise grid spacing to be much
larger than in full LES (no resolution of streaks and near-wall eddy structure),
the costs may still be excessive for ows at very large Reynolds numbers. Den-
ing the distance from a wall of complex topology may also pose a problem and
brings in some arbitrariness.
3 The Time-Dependent RANS (T-RANS)
We consider here the Time-dependent RANS (T-RANS) approach as a conve-
nient method to compute ows with coherent vortical structure at very high
Reynolds and Rayleigh numbers, where DNS and LES are inapplicable, and
where some hybrid methods, such as DES may also have diculties. Specif-
ically, we focus here on RayleighBenard convection in which a distinct or-
[22] Simulation of coherent eddy structure 663
ganized large-scale structure is known to exist. The approach can also be
applied to other ows with a dominant large eddy structure (vortex shedding,
internal separation and recirculation, longitudinal vortices, natural convection
in enclosures, ows with rotation and in a magnetic eld). In addition to pre-
dicting ow features and heat transfer accurately we also demonstrate that the
T-RANS approach can serve to identify the organized motion and its reorgani-
zation due to imposed ow control methods, be it of a distributed type (extra
body force, such as magnetic), or a boundary control (non-plane wall congu-
ration). The approach belongs to the class of VLES, and is close to the SDM
method of Ha Minh and Kourta (1993)
1
by which the large structure is fully
resolved by solving three-dimensional ensemble-averaged NavierStokes equa-
tions in time and three-dimensional space, whereas the rest of the turbulence
is modelled by conventional turbulence closure models (subscale model). As
compared to LES, here both contributions to the turbulent uctuations and
long-term statistical averages are of equal order of magnitude. Close to a solid
wall, the unresolved (modelled) part is more dominant, and this imposes a
special requirement to model accurately the wall phenomena. On the other
hand, resolving the large-scale motion enables one to capture accurately the
large scale transport, which in conventional RANS methods is usually mod-
elled inadequately by gradient hypotheses.
For buoyancy-driven ows the unresolved random motion is modelled using
a low-Re-number k--
2
algebraic stress/ux closure model. The large-scale
deterministic motion, which is the major mode of heat and momentum transfer
in the central region, is fully resolved by the time solution.
The considered T-RANS closure also brings a substantial computational
advantage. The sub-scale model here RANS is less dependent on the
spatial grid. The time step can be larger, allowing implicit time marching,
the numerical mesh away from a solid boundary does not need to be very
ne and is not directly related to the subscale model. Because only very large
coherent eddies are resolved, good statistics can be obtained with a relatively
small number of realizations. The problem of dening inow conditions at
open boundaries is less restrictive than in LES. Because of the possibility
of treating very high Ra and Re numbers the approach can be used for the
computation of complex technological and environmental ows of practical
relevance.
It is worth mentioning that the T-RANS approach can also be used for
two-dimensional simulations of ows with one homogeneous direction. In this
case the simulation can still reproduce the coherent large-scale structure, but
ensemble-averaged in the homogeneous direction, which is still more than a
standard RANS can do. Such a simulation of RayleighBenard convection
1
At the time we started exploring this approach in early 1995, we were not aware of the
work of Ha Minhs group: we considered a very dierent type of ows, driven by buoyancy,
and the subscale model was also dierent, hence a dierent name, T-RANS
664 Hanjalic and Kenjeres
revealed the convective-roll-cell pattern and the wall Nusselt number in broad
accord with the ensemble-averaged full 3-dimensional T-RANS simulations
(Kenjeres and Hanjalic 2000).
3.1 Equations
The instantaneous motion of an incompressible uid ow driven by thermal
buoyancy is described by the continuity, momentum and energy equation (us-
ing the Boussinesq approximation for density variation):
U
i
x
i
= 0 (1)
U
i
t
=
x
j
_
U
i
x
j
U
i
U
j
_
P P
ref
_
x
i
+g
i
_
ref
_
+
F
i
(2)
t
=
x
j
_
Pr
x
j
U
j
_
. (3)
The term
F
i
stands for other body forces if they appear (Coriolis force if the
ow is subjected to system rotation, Lorentz force in the case of an electrically
conductive uid in a magnetic eld, etc.). Here we deal only with the thermal
buoyancy and this force has been expressed separately, but the same approach
has also been successfully used for ows subjected to combined buoyancy and
magnetic force (Hanjalic and Kenjeres 2000).
For the resolved (ltered) motion, the above equations can be written as
(essentially the same form as for the LES):
U
i
)
x
i
= 0 (4)
U
i
)
t
+U
j
)
U
i
)
x
j
=
x
j
_
U
i
)
x
j
ij
_
(P) P
ref
)
x
i
+g
i
()
ref
) +
F
i
) (5)
)
t
+U
j
)
)
x
j
=
x
j
_
Pr
)
x
j
j
_
, (6)
where the ) stands for resolved (implicitly ltered) quantities, and
ij
and
j
represent contributions due to unresolved scales to the momentum and
temperature equations respectively, which are both provided by the subscale
model.
[22] Simulation of coherent eddy structure 665
3.2 The subscale model
Because only the largest scales are fully resolved in time and space, care must
be taken to provide an adequate subscale model for the remaining, relatively
large, unresolved part of turbulence spectrum. In the near-wall region the di-
rect eect of the large-scale structure is small and the subscale model provides
a major contribution to the turbulent transport of momentum, heat and mat-
ter. Hence, the role of the subscale model is more important than, for example,
the subgrid-scale model in LES. Besides, the subscale model should not be de-
pendent on grid size, as it is in LES, simply because the unresolved motion
spans a larger range of scales than dened by the numerical mesh. It seems
natural to use a single-point closure as practiced in the RANS approach. In
some ows where the coherent structure is not as dominant and the spectrum
has a smooth shape, the single-point closure model may need modication as
mentioned earlier (e.g. Ha Minh and Kourta 1993). However, because in the
ows considered here the scale of the large structure is well separated from the
rest of turbulence, there is no need for such a modication. Furthermore, the
turbulent transport by large-scale motion is fully resolved, and the importance
of the subscale model is particularly signicant very close to the wall where the
large-scale convection is usually negligible. Hence, there seems to be no need
to solve dierential transport equations for second-moments (turbulent stress,
heat ux), and simple algebraic models may suce. The study reported here
was performed using a reduced algebraic expression for heat ux
i
u
i
),
derived by truncation of the modelled RANS dierential transport equation
for u
i
by assuming weak equilibrium, but retaining all major ux production
terms (all treated as time-dependent):
i
= C
k)
)
_
ij
)
x
j
+
j
U
i
)
x
j
+
_
g
i
2
) +
f
i
)
_
_
, (7)
where
f
i
represents other uctuating body forces, and = 1 C
2
= 0.45
and = 1 C
3
= 0.45 are empirical coecients (see Section 2.3 in [2]).
The turbulent stress tensor
ij
= u
i
u
j
) and the correlations involving the
other body forces should also be expressed in similar algebraic forms by trun-
cation of the full transport equations for these quantities. However, in the
present work we use simple eddy diusivity expressions for these moments
and account for the buoyancy by additional terms in the transport equations
for the turbulent kinetic energy k and its dissipation rate , e.g.
ij
=
t
_
U
i
)
x
j
+
U
j
)
x
i
_
+
2
3
k)
ij
. (8)
The closure of the expressions for subscale quantities is achieved by solving
the equations for turbulent kinetic energy k), its dissipation rate ) and tem-
perature variance
2
) (all modied for low-Re-number and near-wall eects),
666 Hanjalic and Kenjeres
where ) (omitted for clarity in the equations that follow) indicates that all
terms in these equations are computed from the instantaneous resolved quan-
tities, resulting in the three-equation model k--
2
:
Dk
Dt
= T
k
+P
k
+G
g
k
+G
f
k
(9)
D
2
Dt
= T
+P
(10)
D
Dt
= T
+P
1
+P
2
+G
g
+G
f
Y (11)
where T stands for diusion, P is production by mean eld gradients, G
g
is
production by thermal buoyancy and G
f
by other body forces, is turbulence
energy dissipation, and Y is destruction of . All these terms are dened by
conventional expressions using standard values for the coecients and the
thermal-to-mechanical turbulence time-scale ratio R =
2
/k
= 0.5.
For the implementation of Lorentz-force eects in an electrically conductive
uid in a magnetic eld, see Hanjalic and Kenjeres (2000a, b).
3.3 Evaluation of second moments
Comparison of T-RANS and DNS statistics, which serves as a rst check if
the T-RANS approach is meaningful, requires care in the interpretation of the
statistics in order to account for the contribution of the resolved motion. The
triple decomposition, by which the instantaneous eld is assumed to consist of
long-term average (time-mean), ensemble-averaged (quasi-periodic) large-scale
structure and random (stochastic) uctuations, provides a satisfactory tool,
indicating that the long-term statistics can be well reproduced by accounting
for both the random and ensemble-averaged contributions. For steady ows
with a distinct large-scale deterministic structure, any instantaneous uid ow
property at a point
(x
i
, t) can be decomposed into time-mean (x
i
), quasi-
periodic (deterministic)
(x
i
, t) and random (x
i
, t):
(x
i
, t) = (x
i
) +
(x
i
, t) + (x
i
, t) = ) (x
i
, t) + (x
i
, t) . (12)
By performing long-term time averaging at a point in space and assuming
that because of the spectral gap between them, the deterministic and random
motion are not directly interacting, the second moments are obtained as the
sum of deterministic and modelled contributions. For example, it follows from
applying long-term averaging, the second moment for two arbitrary variables
and is:
= +
+. (13)
In order to illustrate the implication of the approach proposed above, we
consider the long-term averaged energy equation, which for steady Rayleigh
[22] Simulation of coherent eddy structure 667
Benard convection reduces to:
x
j
_
Pr
)
x
j
)U
j
)
j
_
= 0, (14)
where
i
is provided from the subscale model presented above, i.e.:
i
= u
i
= C
k)
)
_
ij
)
x
j
+
j
U
i
)
x
j
+
_
g
i
2
) +
f
i
)
_
_
. (15)
Further averaging over homogeneous (horizontal) planes yields the expres-
sion for the total heat ux in the vertical (z) direction W = U
z
and U
j
= 0):
Pr
)
z
W
w
= const. (16)
4 T-RANS Simulation of RayleighBenard Convec-
tion
In order to demonstrate the capability of the T-RANS approach, we investi-
gate several cases of steady RayleighBenard convection over at and wavy
bottom walls. RayleighBenard convection is characterized by self-organized,
large-scale convective roll cells, which ll the vertical spacing between the two
horizontal walls. This structure originates from plumes which rise from the
outer edge of the boundary layer at the heated surface (updrafts) and sink
downward from the upper cold boundary (downdrafts). The rise of plumes
and their impingement on the opposite horizontal surface produce a horizon-
tal motion in the wall boundary layer which governs the wall heat transfer.
This in turn generates buoyancy which causes the rise of plumes. As the Ra
number increases, the regularity of the cell pattern disappears, the plumes
detach from the horizontal boundary layers and evolve into thermals. The
large convective roll cells become unsteady and more disorderly, uctuating in
amplitude and orientation with a frequency characterized by the Ra number.
However, the small scale uctuations in velocity and temperature, originating
primarily in the wall boundary layers, are transported by convective cells much
as a passive scalar (e.g. Cioni et al. 1996, 1997). Both experiments (e.g. Chu
and Goldstein 1973) and direct numerical simulations, DNS, (Gr otzbach 1982,
Cortese and Balachandar 1993) indicate that despite disorder, large coherent
structures can be identied even at very high Ra numbers. A recent DNS by
Kerr (1996) in the range of Ra numbers close to the fully turbulent regime
(Ra 2 10
7
) shows that the large structure governs the apparent chaotic
behaviour of turbulent RayleighBenard convection. This evidence indicates
the existence of two distinct scales of motion: large amplitudes associated with
668 Hanjalic and Kenjeres
thermals, plumes and convective cells, and the small-scale turbulence gener-
ated mainly in the wall boundary layer and carried away by the large scale
structure.
The separation of the scales of the coherent convective cellular motion from
the rest of turbulence in RayleighBenard convection (and other turbulent
ows with dominant large structures) makes these ows very suitable for the
VLES technique. By fully resolving the large-scale deterministic convective
structure and associated momentum and heat transport (regarded as particu-
larly dicult to model with single-point closures), a simple eddy-diusivity
or algebraic closure can be used to model the unresolved motion.
For the conventional at-wall RayleighBenard convection a series of com-
putations was performed covering Rayleigh numbers from 6.510
5
to 10
12
(currently being extended to Ra = 10
15
!). It is noted here that the present
limit on DNS is Ra 2 10
7
(Kerr 1996), and that the highest LES simu-
lated Ra number available in the literature is Ra = 10
8
(Eidson 1985, Peng
and Davidson 2000). However, already at this Ra number the well resolved
LES becomes very costly, at least when accurate heat transfer is required, be-
cause of the need to use a very dense grid in all three directions in the very
thin thermal boundary layers on the walls.
The T-RANS computations were rst performed for Ra = 6.5 10
5
which
is the same as in the DNS of W orner (1994), to enable a direct comparison.
We also performed LES for the same Ra number, using the same mesh as
for T-RANS, which enabled comparison of all three simulation approaches.
The larger values of Ra number were then considered in order to demonstrate
applicability of T-RANS to high Ra numbers where neither DNS nor LES can
be applied.
4.1 RayleighBenard convection over a at wall
Quantitative validation. We present rst the T-RANS computations of
the long-term averaged temperature proles and important second-moments:
turbulent heat ux and temperature variance in the vertical direction for classi-
cal RayleighBenard convection. The results were obtained by averaging over
horizontal homogeneous planes and are compared with the DNS results of
Gr ozbach (1990) and W orner (1994). The mean temperature proles, (denoted
in the gures as T rather than ) normalized with the wall temperature T
w
and the domain height D for a range of Rayleigh numbers from 3.8 10
5
to
10
9
, see Fig. 2, show the characteristic uniform temperature in the core re-
gion with steep gradients in the wall thermal boundary layers that become
progressively thinner as the Rayleigh number increases. Fig. 3 shows the pre-
dicted long-term averaged turbulent heat ux in the vertical direction (a) and
temperature variance (b) over the channel height for Ra = 10
7
.
The contributions of both, the resolved and modelled parts are of the same
order of magnitude, the resolved part dominating over most of the channel
[22] Simulation of coherent eddy structure 669
Figure 2: Long-term averaged vertical temperature proles in RayleighBenard
convection (Kenjeres and Hanjalic 1999a)
central region, and the modelled part dominating in the near-wall region. The
molecular contribution to the heat ux is also plotted, and the sum of all three
a. b.
Figure 3: The modelled and resolved parts of: (a) vertical turbulent heat ux;
(b) temperature variance (Kenjeres and Hanjalic 1999a)
670 Hanjalic and Kenjeres
a.
b.
Figure 4: Normalized mean temperature proles in RayleighBenard convec-
tion, (a.) near-wall blow-up, (b.) semi-logarithmic plot for T-RANS and DNS
over range of Ra numbers (Hanjalic and Kenjeres 2000a)
contributions, normalized with the wall heat ux, shows that the long-term
averaged total heat ux is constant over the entire channel height (here equal to
1). A near-wall blow-up of the same results, as well as semi-logarithmic plots
over the entire channel height, normalized with the buoyancy temperature
T
q
= Q/U
q
and length scale Z
q
= /U
q
, where Q is the wall heat ux and
U
q
= (
2
gQ/)
1/4
is the buoyancy velocity, are shown in Figs. 4 and 5.
[22] Simulation of coherent eddy structure 671
a.
b.
Figure 5: Distribution of normalized vertical heat ux (near-wall blow up) (a.),
and temperature variance (b.) in RayleighBenard convection: T-RANS and
DNS (Hanjalic and Kenjeres 2000a)
Excellent agreement is obtained despite signicant dierences in Rayleigh
numbers, providing a sucient proof that the T-RANS method is capable of
reproducing faithfully the long-term averaged temperature eld.
It is noted that although, in the long-term average over homogeneous hori-
zontal planes, the ow is steady and seems very simple with only one (vertical)
inhomogeneous direction, the dominating large-scale structure in transporting
672 Hanjalic and Kenjeres
momentum and heat is the major reason for the failure of a steady compu-
tations with conventional single-point closures to reproduce the mean ow
features and turbulence statistics in RayleighBenard convection. The main
deciency of eddy-viscosity/diusivity models, if considered in a steady mode,
is the gradient transport hypothesis for the momentum and heat ux. Second-
moment closure, in which the turbulent ux is determined from a dierential
equation, oers no better prospects because the gradient transport model of
triple moments and, especially, of pressure diusion, seems to be totally in-
adequate for this type of ows (e.g. W orner 1994). It will be interesting to
see whether the 3rd-moment treatments advocated in [14] and [15] are more
successful in reproducing such ows.
Qualitative validation: structure morphology. A second test of the ap-
plicability of the T-RANS approach is the comparison of structure morphology
in DNS, LES and T-RANS realizations. Various identication criteria have
been applied including numerical visualization of the coherent structure mor-
phology, critical point theory and vortex dynamics.
An illustration of the T-RANS performance and its relation to DNS and
LES is given in Fig. 6, showing the instantaneous structure, captured by the
kinematic vorticity number N
k
= (
2
k
/2S
ij
S
ij
)
1/2
for DNS, LES and T-RANS.
A very similar picture is obtained if the second invariant of the velocity gra-
dient tensor A
ij
= U
i
/x
j
is used instead of N
k
(not shown here). All three
computations were performed for the same Rayleigh number, Ra = 6.5 10
5
.
The captured structure depends on the adopted threshold value of N
k
, which,
in this case, was chosen to be 2. In the background is the numerical mesh used
in each computation. Note that the grid is uniform in the homogeneous direc-
tion and clustered close to the walls in the wall-normal direction. The gures
illustrate the eect of ltering: while the structure in all cases is irregular but
evenly distributed in the homogeneous planes, it is clear that both the LES
and the T-RANS lter out the small-scales, and that the implicit cut-o wave
number in the T-RANS is smaller than in the LES, the former displaying only
the very large structures.
A three-dimensional snap-shot of the plumes and thermals structure is given
in Fig. 7a, showing two isothermal surfaces of the instantaneous nondimen-
sional temperature
=
ref
0.1, where the + sign corresponds to plumes
rising from the bottom wall and the sign to those descending from the upper
wall. Note that the two surfaces in Fig. 7a for the selected values of
corre-
spond to temperature contours more in the plume cores, hence they exhibit a
deviation from the expected mushroom shapes. Such types of structure with
nger-like plumes and with horizontal (planform) structures in between have
been detected earlier both by experiments and by DNS. Temperature elds
in the vertical and horizontal midplanes, given by dierent intensities of grey
scale, for the same instant of realization as in Fig. 7a, are given in Figs. 7b
[22] Simulation of coherent eddy structure 673
DNS: 20120151, W orner (1994)
LES: 828232 T-RANS:
Figure 6: Qualitative comparison of captured vortical structures for DNS, LES
and T-RANS realization: Ra = 6.510
5
, Pr = 0.71, ^
k
=
_
[
i
[
2
/2S
ij
S
ij
_
1/2
=
2; top view, with underlying computational mesh (Hanjalic and Kenjeres
2000b)
and 7c. The network of polygonal cells with ngerlike plumes in between,
usually associated with laminar RayleighBenard convection, is clearly visible
in the horizontal plane despite the fact that the Ra number (10
7
) is well into
the turbulent regime. These plumes are actually the main carriers of heat,
and the plume locations correspond to those where the local Nusselt numbers
reach their maximum values, as seen in Fig. 12a (to be discussed below). The
plumes move randomly and interact with each other causing a strong vertical
motion. Again, very similar pictures were obtained from DNS (not shown here)
674 Hanjalic and Kenjeres
a.
b.
c.
Figure 7: Thermal plumes (a) and temperature elds in vertical (b) and hor-
izontal (c) midplanes in RayleighBenard convection, Ra = 10
7
, Pr = 0.71
(Hanjalic and Kenjeres 2000b)
[22] Simulation of coherent eddy structure 675
Figure 8: Instantaneous trajectories of massless particles in DNS at Ra =
6.5 10
5
(left) and T-RANS at Ra = 10
7
(right) indicating spiraling updrafts
(above) and evolution from three line sources (below) (Kenjeres and Hanjalic
1999b)
and T-RANS results, except for a dierence in scales. The identical scenario
of planform structures has been observed in experimental studies of Theertan
and Arakeri (1997).
Next we compare directly some results of visualization of the instantaneous
structure patterns (randomly selected realizations) in DNS and T-RANS com-
putations for the same or similar Rayleigh numbers. Of particular interest is
the identication of the nger-like regions with intense spiraling updrafts ex-
tending in the vertical direction. Cortese and Balachandar (1993) detected
such a structure in their DNS arguing that the origin of vorticity is in hori-
zontal ow, induced by uid motion towards irregularly spaced sites of plume
or thermal release (or away from irregularly spaced stagnation points). The
tilting and stretching of rising plumes by buoyancy in the vertical direction
produce spiral structures. Fig. 8a, obtained by releasing a number of massless
particles in one frozen realization of DNS of W orner (1994) and one T-RANS
realization, indeed supports the above nding: the trajectories show a spiraling
tendency.
A release of particles from line sources placed at the vertical midplane
(Fig. 8b) showed also the spiraling motion around randomly oriented hor-
izontal axes. A better demonstration of the existence of three-dimensional
vortical structures is obtained by plotting the projection of instantaneous tra-
jectories on the central horizontal plane, z
= 200 (Hanjalic
and Kenjeres 2000a). Left gures 2D wavy wall; right gures 3D wavy wall
such as the proles of turbulence variance and uxes were not very sensitive to
the variations of the bottom-wall topology. On the other hand, the motions
structure persisted considerably longer over the wavy terrain than over at
surfaces.
An insight into the eect of wall topology on the spatial organization of
the large coherent structures can be gained from Fig. 10. Here we show the
instantaneous structures over the 2D and 3D bottom wavy walls, with topology
dened above. The structures are identied again by the kinematic vorticity
number ^
k
and show the situation in the early stage of ow development
(
= 200)
are presented in Figs. 10b, 10c and 10d, where the eect of the bottom wall
conguration is shown on the mean properties (W, ) and second moments
(
2
)) respectively for each of the two wall topologies. In the initial stage of
heating,
= 300), particularly
for 2D waviness. It is also interesting to note that the 3D waviness tends to
smooth the Nusselt number distribution in the later stage more eciently than
the 2D topology.
[22] Simulation of coherent eddy structure 679
Figure 11: Inuence of bottom wall topology on the instantaneous local Nusselt
number: (a) at bottom wall, (b) 2D waviness, S
B
= 0.1 cos (x), (c) 3D
waviness S
B
= 0.1 cos (x) cos (y) (Hanjalic and Kenjeres 2000a)
680 Hanjalic and Kenjeres
Figure 12: NuRa correlation obtained by 2D and 3D T-RANS compared with the
available DNS and experiments (Kenjeres and Hanjalic 2000a)
4.3 The NusseltRayleigh Number Correlation
We now turn back to classical RayleighBenard convection over a at wall and
discuss the T-RANS prediction of integral heat transfer. A number of experi-
mental, DNS and LES results are available in the literature. The DNS results
can serve as a good reference for T-RANS validation at low Ra numbers (up
to 210
7
), where they show a general trend Nu Ra
2/70.01
in accord with
some earlier experiments. For moderate Ra numbers, 10
7
< Ra < 10
10
, several
experimental correlations seem to follow the same NuRa correlation. How-
ever, the high Ra number range is more uncertain: several recent experiments
for Ra > 10
10
disagree considerably over the Ra-number exponent. While
Chavanne et al. (1997) observed a continuous increase in the Ra number ex-
ponent, approaching 0.4 at Ra = 5 10
12
, recent experiments by Niemela et
al. (2000) extending to Ra numbers of 10
17
found no such trend. The latter
authors reported that their measurements are correlated remarkably well by a
single power law Nu = 0.124Ra
0.309
over eleven decades of Ra number, from
10
6
to 10
17
!
It should be noted that early experiments at low Ra numbers were per-
formed in geometries with a low aspect ratio, hence they could be contam-
inated by the side-wall eects (friction, blockage, heat losses). More recent
experiments at very high Ra numbers have been performed with liquid he-
lium at very low temperatures, which can yield only the long-term averaged
[22] Simulation of coherent eddy structure 681
integral heat transfer coecients, and hardly any information about the ow
structure.
This brief outline of current uncertainties in establishing the NuRa cor-
relation for very high Ra number, even for a very simple geometry such as
RayleighBenard convection, is included here to emphasize the potential of
the T-RANS and other VLES techniques, which at present oer the only
viable method both to predict heat transfer and to gain insight into the con-
vective roll-cell pattern and other coherent structures and their role in ows
driven by buoyancy or other body forces. Fig. 12 shows the NuRa correlation
obtained by both the 2-dimensional and 3-dimensional T-RANS for a range
of Ra numbers from 10
5
to 10
12
. As mentioned earlier, we are extending cur-
rently these simulations to still higher Ra numbers. The T-RANS results are
in excellent agreement with the available DNS results for lower Ra numbers,
as well as with most experimental data, including the most recent correlation
of Niemela et al. (2000) for high Ra numbers.
5 Conclusions
A time-dependent Reynolds-averaged NavierStokes (T-RANS) approach has
been presented and its potential as a simulation technique for solving turbulent
ows at very high Reynolds and Rayleigh numbers discussed. The particular
strategy adopted is restricted to ows with distinct coherent large-eddy struc-
tures. The approach belongs to the class of Very Large Eddy Simulations,
and is similar to the semi-deterministic model of Ha Minh and Kourta (1993).
The ensemble-averaged NavierStokes equations are solved in time and space,
with a single-point closure playing the role of the subscale model for the
unresolved motion. In comparison with conventional LES, the model of the
unresolved motion covers a much larger part of the turbulence spectrum, in
fact, almost the complete stochastic turbulence, whereas the large determin-
istic structure is fully resolved. Unlike LES, both the resolved and unresolved
contributions are of the same order of magnitude, and in the near-wall regions
the contribution of the unresolved part is dominant. This places especial im-
portance on the subscale model, which needs to provide a good reproduction
of near-wall turbulence transport.
The method was applied to analyze the RayleighBenard convection over
both at and wavy walls, over a range of Rayleigh numbers. The large-scale
convective roll cells are numerically fully resolved in time and space, whereas
the second-moments (heat ux, stress, scalar variance) associated with the
remaining unresolved turbulence spectrum for velocity and scalar variables
are provided by a subscale model for which a single-point, three-equation k--
2
algebraic stress/ux closure was used.
The T-RANS simulations of the classical RayleighBenard convection repro-
duce distributions of the mean ow properties, wall heat transfer and second-
682 Hanjalic and Kenjeres
moment turbulence statistics in agreement with the DNS results for the same
Rayleigh numbers. Moreover, when scaled with the buoyancy velocity dened
in terms of the wall heat ux, the proles of the rst and second moments in
the near-wall regions collapse almost onto one curve over a range of Rayleigh
numbers spanning four decades. The computations also captured the large-
scale deterministic structure in accord with the DNS, LES and experimental
ndings, as demonstrated by comparing qualitatively the structure morphol-
ogy identied by several criteria, primarily the kinematic vorticity number
and second invariant of the velocity gradient tensor. Comparisons of velocity
vectors, instantaneous trajectories of massless particles, planform and plume
structures, (Kenjeres and Hanjalic 1999a, 1999b) provided additional illustra-
tion of a striking similarity between selected T-RANS and DNS realizations,
thus supporting the claim that the T-RANS approach can be used to study co-
herent structures and their organization in various types of RayleighBenard
and similar ows.
The method has also been shown to be useful in predicting the eect of
various means to control the ow and turbulence. As an example of boundary
control, two types of bottom-wall non-planar topologies, 2D and 3D wavy
walls, were considered, showing the expected structural reorganization and its
eect on mean ow properties, heat transfer and turbulence statistics.
The method has also been extended to simulate ows of electrically conduc-
tive uid subjected simultaneously to thermal buoyancy and a magnetic eld
(magnetic RayleighBenard convection), (Hanjalic and Kenjeres 2000b), the
latter being regarded as an example of distributed (body force) ow control.
The method demonstrated its ability to predict the eect of the magnetic
eld on the reorganziation of the large eddy structure, as well as the resultant
modication of the heat transfer.
Finally, the method was shown to be useful for predicting the ow structure
and heat transfer at very high Rayleigh numbers (at present Ra = 10
12
), which
are inaccessible to any other numerical simulation technique, and are also still
challenging for experiments. The method can also be used for the computation
of technological and environmental ows with realistic complex geometries.
References
Aubrun, S., Kao, P.L., Ha Minh, H. and Boisson, H. (1999). The semi-deterministic
approach as way to study coherent structures: case of a turbulent ow behind
a backward-facing step. In Engineering Turbulence Modelling and Experiments,
edited by W. Rodi and D. Laurence, Elsevier, 491499.
Balachandar, S. (1992). Structure in turbulent thermal convection, Phys. Fluids A,
4 27152726.
Besnard, D., Harlow, F., Rauenzahn, R and Zemach, C. (1996). Spectral transport
model of turbulence, Theor. Comput. Fluid Dyn. 8 1.
[22] Simulation of coherent eddy structure 683
Bosch, G. and Rodi, W. (1998). Simulation of vortex shedding past a square cylinder
with dierent turbulence models, Int. J. Num. Meth. Fluids 28 601611.
Chavanne, X., Chill a, B., Castaign, B., Hebral, B., Chabaud, B. and Chaussy, J.
(1997). Observation of ultimate regime in RayleighBenard convection, Phys. Rev.
Lett. 79 36483651.
Chu, T. and Goldstein, R.J. (1973). Turbulent natural convection in a horizontal
layer of water, J. Fluid Mech. 60 141159.
Cortese, T. and Balachandar, S. (1993). Vortical nature of thermal plumes in turbu-
lent convection, Phys. Fluids A 5 (12) 32263232.
Durbin, P.A. (1995). Separated ow computations with the k--v
2
model, AIAA
Journal 33 (4) 659664.
Gr otzbach, G. (1982). Direct numerical simulation of laminar and turbulent Benard
convection, J. Fluid Mech. 119 2753.
Gr otzbach, G. (1983). Spatial resolution requirement for direct numerical simulation
of RayleighBenard convection, J. Comput. Phys. 9 241264.
Ha Minh, H. and Kourta, A. (1993). Semi-deterministic turbulence modelling for
ows dominated by strong organized structures. In Proc. 9th Int. Symp. on Tur-
bulent Shear Flows, Kyoto, Japan 10.5-110.5-6.
Hanjalic, K. and Kenjeres, S. (2000a). T-RANS simulation of deterministic eddy
structure in ows driven by thermal buoyancy and Lorentz force. Submitted for
publication in Flow Turbulence and Combustion.
Hanjalic, K. and Kenjeres, S. (2000b). Reorganization of turbulence structure in
magnetic RayleighBenard convection: a T-RANS study, J. Turbulence 1 (8) 1
22.
Hanjalic, K., Launder, B.E. and Schiestel, R. (1980). Multiple-time-scale concepts
in turbulent transport modelling. In Turbulent Shear Flows 2, edited by L.J.S.
Bradbury et al., Springer, 3649.
Johansson, S. H., Davidson, L. and Olsson, E. (1993). Numerical simulation of vortex
shedding past triangular cylinders at high Reynolds number using a k- turbulence
model, Int. J. Num. Meth. in Fluids 16 859.
Kenjeres, S. and Hanjalic, K. (1995). Prediction of turbulent thermal convection in
concentric and eccentric annuli, Int. J. Heat and Fluid Flow 16 (5) 428439.
Kenjeres, S. and Hanjalic, K. (1999a). Transient analysis of RayleighBenard con-
vection with a RANS Model, Int. J. Heat and Fluid Flow 20 329340.
Kenjeres, S. and Hanjalic, K. (1999b). Identication and visualization of coherent
structures in RayleighBenard convection with a time-dependent RANS, J. Visu-
alization 2 (2) 169176.
Kenjeres, S. and Hanjalic, K. (2000). Convective rolls and heat transfer in nite-
length RayleighBenard convection: a two-dimensional study, Phys. Rev. E 62
(6A), 79877998.
Kerr, M.R. (1996). Rayleigh number scaling in numerical convection, J. Fluid Mech.
310 139179.
684 Hanjalic and Kenjeres
Kretenauer, K. and Schumann, U. (1992). Numerical simulation of turbulent convec-
tion over wavy terrain, J. Fluid Mech., 237 261299.
Niemela, J.J., Skrbek, L., Sreenivasan, K.R. and Donnelly, R.J. (2000). Turbulent
convection at very high Rayleigh numbers, Nature 404 837840.
Nikitin, N.V., Nicoud, F., Wasistho, B., Squires, K.D. and Spalart, P.R. (2000). An
approach to wall modeling in large-eddy simulations, Phys. Fluids 12 (7) 1629
1632.
Peng, S.-H. and Davidson, L. (private communication)
Schiestel, R. (1987). Multiple time scale modeling of turbulent ows in one-point
closures, Phys. Fluids 30 722731.
Spalart, P.R. (1999). Strategies for turbulence modelling and simulations. In En-
gineering Turbulence Modelling and Experiments 4, edited by W. Rodi and D.
Laurence, Elsevier, 317.
Speziale, C.G. (1998). Turbulence modeling for time-dependent RANS and VLES: a
review, AIAA Journal 36 (2) 173184.
Tatsumi, K., Iwai, E., Neo, E.C., Inaoka, K. and Suzuki, K. (1999). Prediction of
time-mean characteristics and periodical uctuation of velocity and thermal elds
of a backward-facing step. In Turbulence and Shear Flow Phenomena 1, edited by
S. Banerjee and J.K. Eaton, Begell House, Inc., New York, 11671172.
Theerthan, S.A. and Arakeri, J.H. (1997). Planform structure of turbulent free con-
vection on horizontal surfaces. In 2nd Int. Symposium on Turbulence, Heat and
Mass Transfer 573580.
Touil, H., Bertoglio, J.P. and Parpais, D. (2000). A spectral closure applied to aniso-
tropic inhomogeneous turbulence. In Advances in Turbulence VIII, edited by C.
Dopazo et al., CIMNE Barcelona, 689692.
Travin, A., Shur, M., Strelets, M. and Spalart, P. (1999). Detached-eddy simulations
past a circular cylinder, Flow, Turbulence and Combustion 63 293313.
Worner, M. (1994). Direkte Simulation turbulenter RayleighBenard Konvektion in
ussigem Natrium. Dissertation, Univ. of Karlsruhe, KfK 5228, Kernforschungszen-
trum Karlsruhe.
23
Use of Higher Moments to Construct PDFs
in Stratied Flows
B.B. Ilyushin
1 Introduction
The distributions of moments calculated using turbulence models of dierent
closure levels contain information about the statistical structure of turbulent
uctuations. This information is sucient for a wide range of practical prob-
lems. However, as was noted in [15], for a complete statistical description of
the ow-eld characteristics of a turbulent ow one should dene all the multi-
dimensional joint probability distributions for values of these characteristics
at every possible ensemble of points in space and time. The active use of this
approach (the use of PDFs in the study of turbulence) began with the work
of Monin (1967) and Lundgren (1967). These papers include the set of equa-
tions for the PDF of a random velocity eld for developed turbulence. The
advantage of using the PDF method for studying turbulence begins with the
fact that every PDF from the family of nite-dimensional PDFs of the veloc-
ity eld contains information about the whole associated system of statistical
moments. However, the set of equations for the PDFs of the turbulence eld
is not closed. The complexity of solving this problem can be judged by noting
that a closed form equation for the one-point joint PDF for developed inhomo-
geneous turbulence can only be obtained by applying additional hypotheses
and the one-point joint PDF is itself multi-dimensional in character. In the
papers of Lundgren (1967) and Onufriev (1970, 1977b) the model equations for
the one-point PDF of the velocity and scalar elds have been derived on the
basis of a number of phenomenological assumptions. Use of these equations
opens the possibility of extending the semi-empirical methods for turbulent
transfer, Onufriev (1977a). One can nd an example of the numerical simu-
lation of turbulent diusion processes through the application of the model
of Onufriev (1977b) for the one-point PDF in the paper by Belotserkovskii
(1987) and the use of the PDF method for describing turbulent diusion of a
scalar in Pope (1980, 1983). Numerical integration of the model equation for
the PDF usually requires powerful computer facilities and is rather expensive.
However, the inherent advantages of the PDF method lead the writer to fore-
see the extensive use of this approach for studying the structure of turbulent
ows in the near future. In this chapter another useful aspect of the use of the
PDF for representing the turbulent transport structure will be considered.
As is known, when large-scale eddy structures with long lifetimes (coher-
ent structures) are formed in a turbulent ow, the turbulent transfer mecha-
685
686 Ilyushin
nism exhibits non-local properties which cannot be described adequately with
gradient-diusion models. For a satisfactory description it is necessary to apply
higher-order closure models. Chapter [15] is devoted to the use of such models
for calculating the turbulence structure of the velocity and temperature elds
in stably stratied ows. Application of the higher-order closure models to
describe the turbulent transfer of a passive scalar requires the use of model
equations (or approximations) for mixed covariances (for example, the con-
centration and temperature-uctuation covariances). Information about their
behaviour is absent so a direct check of the adequacy of such models is dicult.
However there is an alternative way to solve this problem.
The long lifetime of coherent structures (compared with the typical time
scale of turbulence) presuppose their weak statistical dependence on eddies
in the inertial range of the spectrum. This allows us to consider the transfer
by eddies in the inertial range of the turbulent uctuation spectrum (back-
ground turbulence) to be represented by gradient models of turbulent diusion
and the transfer under the action of coherent structures as statistically inde-
pendent processes. The PDF of the background turbulence has a Gaussian
distribution and can be easily distinguished from the PDF of the complete
velocity uctuations. The remaining part of the PDF corresponds to eddies of
the energy-containing range. From a representation of this part one can recon-
struct the velocity eld of the coherent structures u
i
which can then be used
to take into account their contribution to turbulent transfer as a correction
to the mean velocity eld U
i
. This approach thus allows for mass transfer by
coherent structures directly in the convection terms of transport equation for
species concentration as well as for turbulent transfer through the eect of
background turbulence in the diusion term using a standard gradient type
model:
C
t
+ (U
i
+ u
i
..
coherent
structures
)
C
x
i
=
x
i
_
C
u
i
u
j
..
background
turbulence
)
C
x
j
_
To adopt this strategy one needs to reconstruct the velocity-uctuations PDF
with known (calculated) moments of the PDF distribution.
2 Use of higher moments to construct PDFs in
stratied ows
The theory allows us to reconstruct the single and exact PDF from a knowledge
of the distribution of the moments. However, it requires a knowledge of all the
moments of the considered PDF distribution. In practice only some lower-order
moments are usually known. Thus the innite series is clipped after a small
number of terms. As a consequence the PDF thus obtained does not always
satisfy the necessary requirement of positive values within the whole domain
[23] Use of higher moments to construct PDFs in stratied ows 687
of denition. Indeed, reconstructing the PDF for example as a GramCharlier
series
P(w) =
1
2
exp
_
2
2
_
_
1 +
1
3!
S
w
_
3
3
_
+
1
4!
w
_
4
6
2
+ 3
_
+
_
,
(1)
(where = w/, is the dispersion, S
w
is the skewness of the PDF distribution
of velocity w and
w
is its excess) using rst-, second- and third-order moments,
we do not satisfy the positivity condition. The PDFs calculated from (1) are
shown in Figure 1 as well as a graph of the polynomial in square brackets in
(1). It is seen that using the assumption of zero cumulants of order higher than
third order (corresponding to curve 2 in Figure 1), the third degree polynomial
remaining in square brackets (and also the PDF, as a consequence) has a region
of negative values at any value of skewness (see Figure 1b). This approximation
corresponds to the quasi-normal hypothesis of Millionshtchikov (1941). When
clipping the series (1) at the fourth term (that is taking into account non-
zero values of the fourth-order cumulant), the distribution obtained becomes
positive everywhere for some values of S
w
and
w
, in particular for the stably
stratied planetary boundary layer (PBL). For example, for S
w
= 0.2 (Chiba
1978) and with
w
= 2.1 (see Figure 3 of [15]) the PDF reconstructed from
the rst four moments with the help of the series (1) is everywhere positive.
Thus, the use of the quasi-normal hypothesis of Millionshtchikov (1941) does
not allow one to reconstruct a positive PDF by means of (1), whereas in
some cases the assumption of zero cumulants of order higher than fourth does
permit the use of the GramCharlier series (1) for reconstructing the PDF.
However, the PDF created in this a way is not necessarily positive over the
whole domain in ows where there is a signicant asymmetry of turbulent
uctuations (for example, for values S
w
and
w
corresponding to the convective
PBL). This fact restricts the use of the GramCharlier series (as well as other
exact formulations of the PDF as an innite series based on moments or its
compositions) for PDF-reconstruction to the rst three or four moments of the
distribution. This restriction is connected closely with the inadmissibility of
arbitrary clipping of the Taylor series of the characteristic function logarithm.
Thus, in this chapter another way of reconstructing the PDF of the vertical
velocity is applied and considered for the case of a horizontally homogeneous
convective PBL.
3 Constructing PDFs in the convective PBL
The convective boundary layer is usually taken to mean the atmospheric
layer where there is a direct inuence from the underlying surface heated by
the suns rays (Figure 2). For conditions where the Rayleigh number in the
lower regions of the PBL becomes larger than some critical value, large-eddy
convective motions, similar to Benard cells, appear and these play the main
role in the vertical transfer of momentum, heat and matter. Circulation in such
688 Ilyushin
5 0 5
5 0 5
0 . 3
4 8
w
w
1
2
3
2
3
1
2
3
( a )
( b )
P ( w )
GramCharlier polynomial
Figure 1: (a) PDF reconstructed with use of the GramCharlier series: (1):
Gaussian function (S
w=0
,
w
= 0); (2): S
w=0.2
,
w
= 0; (3): S
w=0.2
,
w
=
2.1; and GramCharlier polynomial (b).
cells is shown schematically in Figure 3b on the x-y plane. Non-symmetric
boundary conditions (heating below, stable stratication at the upper bound-
ary) cause an asymmetric distribution of the PBL vertical velocity amplitudes:
steep rising motions and gentler descending ones (see Figure 3a). Such a char-
acter of the motion has been conrmed by eld experiments (Byzova et al.
1991). The curve in Figure 3c represents a typical realization of the vertical
component of velocity for unstable stratication obtained at an altitude of
120m and smoothed over a period of about one minute.
This curve is not exactly periodic because it represents a cross-section of a
quasi-regular system of cells. However, noting the character of this curve, one
can recognize that such a realization corresponds to the velocity eld along
a line passing through the cell centers. The horizontal size of the cells is
between three and ve kilometers. The cells occupy, in the vertical direction,
practically the whole PBL. The distributions of statistical characteristics of the
velocity and temperature elds (second- and third-order moments) reecting
such a structure of the convective PBL have been calculated by Ilyushin (1998)
[23] Use of higher moments to construct PDFs in stratied ows 689
Figure 2: Structure of the convective atmospheric boundary layer.
Figure 3: Vertical motion shape in the convective large scale eddies (a), (b);
and the measured realization of vertical velocity smoothed with the period
1min. (c).
and are shown in Figure 4 together with the data of measurements in the
convective PBL taken from Caughey (1982) and Lenschow et al. (1980). One
can see that the calculated proles are in good agreement with the data of
observations and can be used to reconstruct the PDF of the vertical velocity
uctuation in the convective PBL.
The PDF of the turbulent uctuations of the vertical velocity is represented
as a superposition of two independent distributions: the inertial range of the
turbulent spectrum (background turbulence) P
b
(u) and a large-wave region
PDF spectrum P
c
(v) (u and v are the vertical velocity uctuations of back-
690 Ilyushin
<w
3
>/w
*
3
<Ew>/w
*
3
<w Q
2
>/
Q
o
w
*
<w
2
>w
*
/Q
o
2
z /z
i
2
o
*
0
1
0 . 0 0 . 2
1
0 . 0 0 . 2 0 . 0 0 . 8 0 4
1 0 1
z / z
i
2
*
2
h
*
2
i
*
3
0
2
< w w
> /
e
/ w
z
/ w
< w> / Q
< > / (
Q
/
w
)
1 1 1
Figure 4: Results of simulation of the convective ABL evolution: solid lines
are calculated proles of second- and third-order correlations (dierent lines
correspond to times from 10 am to 5 pm); symbols are the experimental data.
ground turbulence and the vertical velocity eld of coherent structures, re-
spectively)
P(u, v) = P
b
(u)P
c
(v) =
1
2
2
b
exp
_
u
2
2
2
b
_
. .
background
turbulence PDF
_
_
a
+
+
c
exp
_
(m
+
v)
2
2(
+
c
)
2
_
. .
upow
+
a
c
exp
_
(m
v)
2
2(
c
)
2
_
. .
downow
_
_
. .
Large Scale Eddy Formation PDF
, (2)
where
b
is the dispersion of background turbulence, a
+
and a
are weighting
[23] Use of higher moments to construct PDFs in stratied ows 691
coecients,
+
c
and
c
are the dispersions, and m
+
and m
(m
+
w)
2
2
2
+
_
+
a
exp
_
(m
w)
2
2
2
_
, (3)
where
2
+
= (
+
c
)
2
+
2
b
and
2
= (
c
)
2
+
2
b
. From the conditions
_
R
P(w) dw = 1,
_
R
wP(w) dw = 0,
_
R
w
2
P(w) dw = w
2
),
_
R
w
3
P(w) dw = S
w
3
(4)
(here = w
2
)
1/2
is the dispersion and S
w
= w
3
)/w
2
)
3/2
is the skewness
factor), the connections among a
+
, a
,
2
+
,
2
, m
+
and m
are:
a
+
+a
= 1,
a
+
m
+
+a
= 0,
a
+
_
(m
+
)
2
+
2
+
+a
+
_
(m
+
)
2
+
2
+
2
a
+
_
(m
+
)
3
+ 3m
+
2
+
+a
_
(m
)
3
+ 3m
= S
w
3
.
(5)
The conditions for
2
+
and
2
should be equal to
the square of the average velocities of upow and downow (m
+
)
2
and (m
)
2
correspondingly:
2
+
= (m
+
)
2
,
2
= (m
)
2
. (6)
This assumption means that the ux directed positively (negatively) remains
mainly positive (negative) taking into account possible scatter. The necessary
conditions for closure of the equation set (for
b
) is found from the wavelet
model (Tennekes and Lumley 1972). Here we suppose that the main part
of the energy-containing range of the uctuation spectrum is dened by a
single main wavelet with a typical wave number corresponding to the spectrum
maximum (see Figure 5). The simplied eddy with a typical wave number
v
is considered in the form of a localized perturbation of energy in wave number
space (wavelet) with energy E
v
= E(
v
)
v
. This consideration ensures the
cascade transfer of turbulence energy from large-scale eddies to dissipative
eddies. In the present work the coherent structure in the convective PBL is
assumed to be the wavelet containing the energy E
c
= a
2/3
2/3
max
(where
a = 1.6 0.02 is the Kolmogoro constant, Andreas (1983), and
max
is the
maximum of the spectrum of the turbulent energy see Figure 5). The total
turbulence energy is equal to
E =
max
min
E
v
(
v
) = E
v
(
max
)
i=0
_
3
2/3
_
i
2E
c
. (7)
692 Ilyushin
Figure 5: Presentation of the long-wave part of the turbulent uctuations
spectrum as the wavelet.
A similar result can be obtained by integration of the inertial range (without
accounting for the departure from the Kolmogoro spectrum in the dissipation
range) from
max
to innity and taking into account also the energy of half of
the wavelet (which is outside of the inertial range, see Figure 5):
E
_
max
a
2/3
5/3
d +
1
2
E
c
= 2E
c
, (8)
or
2
b
= E (recall (
2
b
+
2
c
)/2 = E). Taking account of this result,
max
is
expressed as
max
= (2a/E)
3/2
and then
max
= 2/
max
. The calculated
proles of
max
and the ratio
2
b
/w
2
) in the convective atmospheric boundary
layer are shown in Figure 6. Here one can seen that in the mixing layer the cal-
culated distribution of
max
is in agreement with the observed data (Caughey
1982). The linear function
max
= 5.88z/z
i
corresponds to conditions of free
convection in the near-ground layer. The calculated ratio
2
b
/w
2
) is equal to
1/3 in the mixing layer of the PBL. Analysis of the experimental spectrum
of w
2
) in the atmosphere (Monin and Yaglom 1967) gives the same result
2
b
/w
2
) 1/3. The necessary condition for the dispersion of the background
turbulence
b
is
2
b
=
1
3
w
2
). (9)
Taking account of the variation in sign of m
+
> 0 and m
/ <
w
> b
2
2
1 / 3
z / z
i
Figure 6: The calculated and measured (Caughey 1982) proles of
max
=
2/(E/2a)
3/2
and the ratio
2
b
/w
2
).
sponds to upow and downow), equations (7), (9) have unique solutions:
m
+
=
4
_
S +
S
2
+ 8
_
; m
=
4
_
S
S
2
+ 8
_
;
a
+
=
S
S
2
+ 8
2
S
2
+ 8
; a
=
S
S
2
+ 8
2
S
2
+ 8
;
(
+
c
)
2
=
2
16
_
S +
S
2
+ 8
_
2
1
3
w
2
);
(
c
)
2
=
2
16
_
S
S
2
+ 8
_
2
1
3
w
2
).
(10)
The results of reconstructing the PDF using the calculated distributions of
the second- and third-order moments are shown in Figure 7. The picture shows
that the results of reconstruction correspond to the observed data: an upow
within a coherent structure with larger turbulent energy (
+
c
) and larger ve-
locity (m
+
) occupying a smaller region (of size P
c
(m
+
) < P
c
(m
)) and a
downow with smaller energy (
c
<
+
c
) and smaller velocity (m
< m
+
).
Such a structure for the convective PBL arises from the vertical asymmetry in
the generation mechanism for the turbulent uctuations: in the near-ground
layer a growth of turbulence energy is caused by the mean velocity shear
whereas, in the upper (stably stratied) part of the PBL, turbulent uctua-
tions are strongly suppressed. Therefore the intensity of the vertical velocity
uctuations in upows appears to be larger than that in downows. Figure
8 shows the PDF proles obtained in the measurements (Byzova et al. 1991)
in the atmosphere compared with calculations using the model of Deardor.
As is obvious the reconstructed PDF in the present work P(w) (see Figure 7)
694 Ilyushin
2
2
2
0
Figure 7: The reconstructed prole of PDF of the vertical velocity uctuations
for z/z
i
= 0.5.
is in qualitative agreement with those shown in Figure 8: it has the maximum
in the region of negative velocities and the more convex part in that of positive
velocities.
To obtain the distribution of vertical velocity uctuations from individ-
ual realizations, it is necessary to apply a procedure of ensemble averaging.
For horizontally homogeneous conditions, ensemble averaging and averaging
over horizontal coordinates can be regarded as equivalent procedures. And
nally, taking into account Taylors hypothesis of frozen turbulence, averag-
ing over horizontal coordinates can be replaced by time averaging for xed
horizontal coordinates if the time of this averaging is larger than the drift
(under the action of averaged ow) time D/U of the largest turbulent eddies.
This allows us to decrease signicantly the time required to obtain the mea-
surements. Proles of PDF shown in Figure 8 have been obtained using this
method. The following assumptions have been made. For the horizontally ho-
mogeneous PBL, the probability, P(w) dw, of the velocity falling within the
range [w; w+dw] is proportional to the horizontal size of the region with such
velocity, P(w) dw dx. The continuity condition for velocity applied to a
region between the large-scale eddies requires that the adjoining eddies have
opposite directions of swirling (see Figure 9). A pair of such eddies we denote
as a simple cell. For a statistical analysis of the sequence of such identical cells
carried by the wind it suces to study just one cell. This cell is represented
as a coherent structure which (within the adopted assumption) is a wavelet
centered at the location
max
and of horizontal size
max
. It is evident that the
iterative character of the structures in the PBL makes the averaging within
[23] Use of higher moments to construct PDFs in stratied ows 695
Figure 8: Distributions of vertical velocity variance in the convective PBL: (a)
observed data and (b) calculation using Deardors model (from Byzova et al.
1991).
the time interval
max
/(2U) and the averaging over an innite time equivalent.
Thus, we may consider that the part, P
c
(w), of the reconstructed PDF that
has to do with coherent structures, characterizes the local vertical uxes up
and down, that is one eddy (half of the cell). This allows us to reconstruct
a single eld of the vertical velocity (and then of the horizontal one) for this
cell.
As mentioned above, the size of the horizontal region for the coherent struc-
tures (where the vertical velocity is equal to w) is taken to be proportional to
the probability P
c
( w):
P
c
( w) d w =
dx
(
max
/2)
.
The vertical velocity eld of the coherent structures w(x, z) can be found
from this dierential equation. The horizontal velocity eld u(x, z) is deter-
696 Ilyushin
z/z
i
1 0
1
x/
max
Figure 9: The reconstructed velocity eld of the coherent structure in the
convective PBL.
Figure 10: Conditionally averaged (a) updraughts and (b) downdraughts in
terms of (i) u and (ii) w (from Schmidt and Schumann (1989) calculated by
the LES-method.
mined from the equation of mass conservation. The result of reconstructing
the coherent-structure velocity eld ( u, w) in the convective PBL is shown in
Figure 9. The conditionally averaged (a) updraughts and (b) downdraughts
in terms of (i) u and (ii) w (from Schmidt and Schumann 1989) are shown
in Figure 10 (solid lines are positive values and dashed lines are negative
ones). One can see qualitative agreement for the reconstructed eld of the
[23] Use of higher moments to construct PDFs in stratied ows 697
coherent-structure velocities with that calculated in this latter paper by an
LES simulation.
4 Modeling the turbulent transport of substance
To describe the process of species dispersion in the convective PBL, we use
a model taking directly into account the eect of mass transfer by coherent
structures (in the advection terms of the equation for the crosswind inte-
grated concentration, see Ilyushin and Kurbatskii (1996)). To account for the
turbulent diusion of matter by background turbulence, a standard gradient-
diusion model is applied:
C
y
t
+ (U + u) =
C
y
x
+ w
C
y
z
z
_
C
s
w
2
)
C
y
z
_
, (11)
where C
s
is the model coecient of the model whose value is determined from
the requirement that the coecient of turbulent diusion near the surface
should equal u
0
4.5
2
2
3
C
s
u
z C
s
0.07. (12)
The results of a simulation of a pollutant jet spreading from sources placed
both near ground level and in the middle of the mixed layer are presented. The
crosswind integrated concentration elds averaged oven one period,
max
/U
max
,
are shown in Figures 11 and 12. A near-ground source was realized at z
s
/z
i
=
0.07. The maximum concentration centerline (i.e the locus of maximum con-
centration) rst moves parallel to the surface and then starts to rise rapidly
at a downwind distance x
= 0.5 (here x
= xw
/(z
i
U
x
)), creating a local
maximum of concentration near the inversion layer. This maximum is located
at a height of z = 1.1z
i
at about x
= 670 target (Spalart 1988). Over 9 million grid points are needed to re-
solve this ow, requiring 13 Mwords of central memory. A uniform grid spacing
is applied in the streamwise and spanwise directions, with points clustered in
the vertical direction near the horizontal walls upstream and downstream of
the step. (The spanwise resolution is slightly too coarse very near the wall,
but this does not lead to a statistically signicant error.) Results were gen-
erated on a Cray C90, requiring on the order of 1000 CPU hours to obtain
well-converged time-averaged statistics.
An indication of the spatial resolution of the Le et al. simulation is given
by the instantaneous vorticity contours shown in gure 1; the overall delity
of their results is reected in the good agreement with Jovic and Drivers
(1994) experimental mean pressure and skin friction distributions (the sym-
[24] Direct numerical simulations of separation bubbles 705
0 5 10 15 20
x/h
0.004
0.002
0
0.002
0.004
C
f
Figure 3: Comparison of skin friction coecient from the simulation of Le et al.
(1997) (solid line) with the experiment of Jovic and Driver (1995) (symbols).
(From Le et al. (1997).)
bols in gures 2 and 3). This agreement was especially signicant, since at
this Reynolds number the magnitude of the negative mean skin friction in the
recirculation region is much larger than that found in experiments done at
larger Re. After demonstrating close correspondence between the DNS results
and other statistics measured by Jovic and Driver, including the location of
the mean re-attachment point (which agreed within 3%) and the downstream
evolution of the mean velocity and Reynolds-stress proles, Le and Moin turn
their attention to quantities only available from the simulation. For example,
they investigate the instantaneous structure of the detached shear layer (and
the associated temporal and spanwise variations of the re-attachment region)
and its connection to mean and instantaneous pressure uctuations through-
out the ow. Turbulent kinetic energy budgets are also examined, which reveal
the strong similarity of the region above the recirculation with two-dimensional
plane mixing layers, as well as the slow development of a conventional (nom-
inally) zero-pressure-gradient boundary layer downstream of re-attachment.
Even twenty step-heights downstream of the step, evidence of the sudden ex-
pansion and re-attachment can still be seen in the mean velocity prole. By
documenting the unexpectedly slow recovery, especially in terms of deviations
from the standard zero-pressure-gradient log law, Le and Moins ndings in-
directly exposed an inaccuracy in many of the previous backward-facing-step
skin-friction measurements. Since they were based on Clauser chart readings
(and thus implicitly assumed the log law was valid where it was not), the exper-
iments tended to report skin friction coecients that were signicantly smaller
than the actual value. Le et al. and Jovic and Driver illustrate the manner in
which DNS and laboratory experiments can complement each other, especially
706 Coleman and Sandham
when they are done simultaneously.
Data from the Le et al. simulation have already been widely used for tur-
bulence closure. The dierential a priori analysis by Parneix et al. (1998)
evaluated a second-moment model and showed that the main problem was
closure of the u
in
= 3238, (b) in 3350, (c) 3463, and (d) 3575. (From Na and Moin
(1998).)
et al. 1994; Maucher et al. 1994, 1999) focus primarily upon the growth of dis-
crete two- and three-dimensional instability waves introduced upstream of the
separation (which may or may not become turbulent and close into a bubble,
depending upon the APG and disturbances imposed); the latter are concerned
with separated laminar boundary layers, and their subsequent turbulent re-
attachment.
712 Coleman and Sandham
Figure 8: The mean ow structure of a laminar separation bubble. (Horton
1968).
This second category has been the subject of two independent investigations,
Alam and Sandham (2000) and Spalart and Strelets (2000). Both study separa-
tion of a laminar at-plate boundary layer, induced by aspiration through the
opposite boundary. In both cases the detached shear layer quickly becomes
turbulent, and its rapid spreading leads to re-attachment a short distance
downstream of the separation line. As a result, a closed bubble again forms,
but unlike the turbulent-detachment, here the re-attachment is not forced by
a favourable pressure gradient caused by blowing at the opposite wall, but
rather by the turbulence that develops in a detached free shear layer. After
the re-attachment, the ow relaxes towards a normal (undistorted) turbulent
boundary layer. Another dierence between this and the turbulent-separation
case is that the Reynolds numbers (based on bubble length) at which practical
laminar separation bubbles occur (e.g. over aerofoils) can be accommodated
by DNS. This rare and important advantage implies that all of these results
apply as is (without, for example, needing to appeal to Reynolds-number
similarity) to realistic and relevant engineering ows.
Both studies use spectral methods with a fringe zone near the downstream
boundary to allow the use of periodic boundary conditions in what is in real-
ity a spatially evolving ow. In both cases it is recognised that some low level
disturbances from the outow do recycle to the inow boundary. The eect
of this is swamped by applied forcing in Alam and Sandham (2000), while
in Spalart and Strelets (2000) these disturbances are not implicated in the
observed transition mechanism (visualised in gure 9). Alam and Sandhams
simulation is of a shorter separation bubble, and uses a computational grid of
up to 384 128 160 compared with Spalart and Strelets 1000 120 120.
Dierences in the suction prole applied to the upper boundary give rise to
two subtly dierent transition processes. In Spalart and Strelets the bubble
is longer, with signicant reverse ow within the bubble. This leads in their
case to a transition by contact where the recirculating turbulent ow from
[24] Direct numerical simulations of separation bubbles 713
Figure 9: Visualisation of the vorticity magnitude and temperature. Mean
separation is at x = 2.25 and mean reattachment at x = 4.25. (From Spalart
and Strelets (2000).)
714 Coleman and Sandham
Figure 10: Iso-surfaces of (a) wall-normal vorticity, showing (locations A and
B) -shaped structures at dierent stages of evolution, and (b) spanwise vor-
ticity, showing (locations A and B) shear layers forming above the -structures
shown in (a). (From Alam and Sandham (2000).)
the back of the bubble triggers transition in the shear layer above the front
of the bubble. As such, no upstream disturbances are required and the tran-
sition is not due to a convective instability. By contrast Alam and Sandham
have a shorter bubble with smaller peak reverse ow and a favourable pressure
gradient downstream of reattachment (which most likely accounts for their ob-
servation of 100% forward ow towards the end of the computational domain,
in contrast to Spalart and Strelets who always observed some reversed ow re-
gions). In the Alam and Sandham simulation upstream disturbances, applied
via a transition strip on the wall upstream of the separation point, are required
for transition. When these were turned o (Alam, private communication) the
transition process moved downstream. Alam and Sandham also observe the
[24] Direct numerical simulations of separation bubbles 715
Figure 11: Turbulence kinetic energy budgets at three locations (a) inside the
bubble, (b) just downstream of reattachment and (c) near end of computa-
tional box. The terms in the budgets are production (P), dissipation (), triple
moment transport (J
u
i,i
), pressure transport (J
p
i,i
), viscous transport (J
i,i
) and
convection (R
u
). (From Alam and Sandham (2000).)
-shaped vortices (gure 10) characteristic of convectively unstable boundary
layer transition. The change in transition behaviour may be linked to a change
in local stability characteristics from convective to absolute, which occurs for
maximum backows of the order of 15% (Alam and Sandham). This analy-
sis may be relevant to the phenomenon of transition by contact described by
Spalart and Strelets. Connections to large scale vortex shedding and eventual
failure of the separated ow to re-attach (bubble bursting) are at present
unclear.
Results from both simulations have already begun to be analysed for their
implications on closure methods. Budgets of turbulence kinetic energy at dif-
ferent downstream locations are shown in gure 11. Near reattachment these
resemble one half of a mixing layer (cf. Le and Moin, Na and Moin), with
signicant contributions from production, advection and the triple-moment
transport term, and smaller but still important contributions from pressure
transport. Further downstream the budgets illustrate the emergence of char-
acteristic near-wall turbulence. The proles of mean velocity in the recovery
region shown in gure 12 are typical of the simulation results. The recovery is
very gradual, occupying a distance up to 25 times the boundary layer thickness
at reattachment.
716 Coleman and Sandham
Figure 12: Streamwise proles in the relaxing boundary layer downstream of
reattachment. The locations are expressed in the legend according to number of
reattachment boundary layer (99.5%) thicknesses
R
downstream of reattach-
ment and (in brackets) from simulation inow in units of inow displacement
thickness. (From Alam and Sandham (2000).)
Spalart and Strelets included in their paper a direct comparison with results
from the one-equation SpalartAllmaras model (see Chapter [1]). The model
successfully captured the transition by contact process without any special
modication, and boundary layer thickness variations along the bubble were
well predicted. Skin friction distributions in the bubble and in the recovery
region were generally poorly predicted, along with heat transfer. The latter
is believed to be due to a violation of Reynolds analogy, as was also seen in
Spalart and Coleman (1997) described above. Another comparison of RANS
with this DNS has been carried out by Hadzic and Hanjalic (2000). They ap-
plied low Reynolds number second-moment closure and also correctly captured
the transition process without special treatments. The peak reverse ow skin
friction was a factor of two dierent from the DNS, but they obtained improved
behaviour in the recovery region downstream of reattachment compared with
simpler eddy viscosity models.
The Alam and Sandham (2000) simulation contains an additional challenge
for models as the transition process is dependent on small amplitude upstream
disturbances. Several two-equation models were tested for this ow by Howard
et al. (2000). In this study the model calculation replicated the DNS in the
use of the same fringe zone and periodic streamwise boundary conditions.
[24] Direct numerical simulations of separation bubbles 717
Thus there was no need to specify inow proles. The transition strip was
represented by a non-zero k distribution upstream of separation, with the k
distribution taken from the DNS. With standard models it was nevertheless
found to be necessary to x the transition by turning on the production of
turbulence term at some prescribed streamwise location. The only case where
this wasnt required was with the Wilcox (1994) k model for transition, in
which model constants are sensitised to the turbulence This model gave the
best overall results, but still signicantly underpredicted the peak negative
skin friction in the bubble. The two-equation models all gave very poor pre-
dictions of skin friction and integral boundary layer properties in the recovery
region downstream of reattachment.
5 Conclusions and future directions
DNS of separating and reattaching ows, possibly including transition to tur-
bulence are now clearly feasible. Transitional separation bubble simulations
are already at realistic Reynolds numbers, comparable to laboratory studies
and practical applications. Turbulent separation simulations are still at rela-
tively low Reynolds numbers, but are nevertheless beginning to provide data
for model development. In the future it is hoped that databases covering a
range of Reynolds numbers can be produced so that Reynolds number trends
can be unambiguously known and predicted. The backward-facing step DNS
highlighted problems with experimental measurements which have been reme-
died by new experiments. Current models generally perform only adequately
for these ows. A common error is under-prediction of the magnitude of the re-
verse ow inside the separation bubbles. From the limited studies so far it does
seem that second-moment approaches do better than standard two-equation
models. Nonlinear eddy viscosity models remain to be tested. There has also
been little published work to date on validation of subgrid models, wall func-
tions and other fundamental aspects of LES against the separated ow DNS.
Now that reliable DNS databases are appearing, one can expect over the next
few years to see a signicant expansion of such work.
References
Alam, M. and Sandham, N.D. (2000) Direct numerical simulation of short laminar
separation bubbles with turbulent reattachment J. Fluid Mech. 410, 128.
Coleman, G.N. and Spalart, P.R. (1993) Direct numerical simulation of a small sepa-
ration bubble. Near-Wall Turbulent Flows (ed. C.G. Speziale and B.E. Launder),
pp. 277286. Elsevier.
Dovgal, A.V., Kozlov, V.V. and Michalke, A. (1994) Laminar boundary layer separa-
tion: instability and associated phenomena Prog. Aerospace Sci. 30, 6194.
Hadzic, I. and Hanjalic, K. (2000) Separation-induced transition to turbulence: second
moment closure modelling. J. of Flow, Turb. and Comb. 63, 153173.
718 Coleman and Sandham
Hanjalic, K., Hadzic, I. and Jakirlic, S. (1999) Modeling turbulent wall ows subjected
to strong pressure variations. J. Fluids Engineering 121, 5764.
Horton, H.P. (1968) Laminar separation in two- and three-dimensional incompressible
ow. PhD Dissertation. University of London.
Howard, R.J.A., Alam, M.A. and Sandham, N.D. (2000) Two-equation turbulence
modelling of a transitional separation bubble. J. of Flow, Turb. and Comb. 63,
175191.
Jimenez, J. et al. (1998) A selection of test cases for the validation of large-eddy
simulations of turbulent ows. AGARD-AR-345.
Jovic, S. and Driver, D.M. (1994) Backward-facing step measurement at low Reynolds
number, Re
h
= 5000 NASA Tech. Mem. 10887.
Jovic, S. and Driver, D.M. (1995) Reynolds number eects on the skin friction in
separated ow behind a backward facing step. Exps. Fluids 18, 464467.
Karniadakis, G.E. and Triantafyllou, G.S. (1992) Three-dimensional dynamics and
transition to turbulence in the wake of blu objects. J. Fluid Mech. 238, 130.
Kim, J., Moin, P. and Moser, R.M. (1987) Turbulence statistics in fully-developed
channel ow at low Reynolds number. J. Fluid Mech. 177, 133166.
Le, H. and Moin, P. (1991) An improvement of fractional step methods for the in-
compressible NavierStokes equations. J. Comput. Phys. 92, 369379.
Le, H., Moin, P. and Kim, J. (1997) Direct numerical simulation of turbulent ow
over a backward-facing step. J. Fluid Mech. 330, 349374.
Maucher, U., Rist, U. and Wagner, S. (1994) Direct numerical simulaiton of airfoil
separation bubbles. Proc. Second European Computational Fluid Dynamics Con-
ference, Stuttgart, 58 Sept., 1994. (ed. S. Wagner et al.), 471477. Wiley.
Maucher, U., Rist, U. and Wagner, S. (1999) Transitional structures in a laminar
separation bubble. Proc. STAB-Symposium, Berlin, 1012 Nov., 1998. Notes on
Num. Fluid Mech., Vieweg.
Mittal, R. and Balachandar, S. (1995) Eect of three-dimensionality on the lift and
drag of nominally two-dimensional cylinders. Phys. Fluids 7, 18411865.
Na, Y. and Moin, P. (1998) Direct numerical simulation of a separated turbulent
boundary layer. J. Fluid Mech. 370, 175201.
Najjar, F.M. and Balachandar, S. (1998) Low-frequency unsteadiness in the wake of
a normal at plate J. Fluid Mech. 370, 101147.
Parneix, S., Laurence, D. and Durbin, P.A. (1998) A procedure for using DNS databases.
J. Fluids Eng. 120, 4047.
Pauley, L.L., Moin, P. and Reynolds, W.C. (1990) The structure of two-dimensional
steady separation. J. Fluid Mech. 220, 397411.
Peng, S-H. and Davidson, L. (2000) New two-equation eddy viscosity transport model
for turbulent ow computation. AIAA Journal 38, 11961205.
Perot, B. (1999) Turbulence modelling using body force potentials. Phys. Fluids 11,
26452656.
[24] Direct numerical simulations of separation bubbles 719
River, R.B., Johnston, J.P. and Eaton, J.K. (1994) Heat transfer on a at surface
under a region of turbulent separation. J. Turbomachinery 116, 5762.
Spalart, P. (1988) Direct simulation of a turbulent boundary-layer up to R
= 1410
J. Fluid Mech. 187, 6198.
Spalart, P.R. and Coleman, G.N. (1997) Numerical study of a separation bubble with
heat transfer. Eur. J. Mech. B Fluids 16, 169189.
Spalart, P.R. and Strelets, M.Kh. (2000) Mechanisms of transition and heat transfer
in a separation bubble. J. Fluid Mech. 403, 329349.
Spalart, P.R. and Watmu, J.H. (1993) Experimental and numerical study of a tur-
bulent boundary layer with pressure gradients. J. Fluid Mech. 249, 337371.
Zhang, K-Q., Fey, U., Noack, B.R., K onig, M. and Eckelmann (1995) On the transition
of the cylinder wake Phys. Fluids 7, 779794.
25
Is LES Ready for Complex Flows?
Bernard J. Geurts and Anthony Leonard
Abstract
Recent developments in LES modelling, combined with further advances in
capabilities of computers and numerical methods, provide a strong incentive
to apply LES to complex ows. This brings to focus a number of issues in the
development of LES that are not yet fully resolved. These include modelling
and numerical elements, their respective errors, and the potential for interac-
tion between these two sources of error. We discuss the relative importance of
some of the errors that can arise in simple as well as complex ows and give
global criteria and guidelines that can be helpful in order to arrive at a form
of LES that is robust and accurate.
1 Introduction
The intricate nature of high Reynolds number turbulent ow has to date deed
detailed rigorous or direct numerical analysis and, consequently, has given rise
to a number of modelling strategies. Such strategies are aimed at reducing the
complexity of the underlying system of equations while retaining sucient in-
formation to reliably predict the ow phenomena of interest in an application.
These two conicting requirements are prominent in large-eddy simulation.
In recent years there has been a signicant interest and associated devel-
opment in large-eddy simulation (LES) which raises the question whether the
LES approach can already be applied to ows of engineering interest and
consequently move away from academic problems. To address this question
objectively in this generality is quite dicult. Instead, we will focus on some
outstanding open problems that need to be confronted in order to develop
LES, in particular for complex ows. While there has undeniably been con-
siderable progress in LES concerning modelling and numerical issues which
has resulted in a number of large-scale complex ow predictions by a large
number of dierent research groups using LES, the signicance of some of the
not yet resolved problems suggests that both a yes and a no answer are still
possible for the question raised. In particular, as we discuss below, the answer
strongly depends on the type of information one wants to predict.
One can choose between the implicit and the explicit ltering approach
to LES (Ghosal 1999). We adopt here the explicit ltering approach because
720
[25] Is LES ready for complex ows? 721
it is more amenable to analysis and allows for a separation of issues related
to modelling and numerical treatment. In this approach a spatial lter is ap-
plied to the NavierStokes equations. The reduction of the ow complexity
and information content that is achieved in this way depends strongly on the
type and the width of the adopted lter. At one extreme, the width of the
applied lter may be so large that virtually all information contained in the
solution is removed while, at the other extreme, a very small lter-width may
be adopted which does not reduce the complexity at all. Subsequently, the
ltered equations need to be closed by the introduction of a model for the
sub-grid stresses and nally the resulting system of equations is treated nu-
merically. These elements of LES, i.e. the information retained in the ltered
eld, the suitability of and the need for accurate subgrid modelling, and the
role of and contamination arising from the numerical treatment will be fo-
cused upon. In view of the desired strong reduction in computational eort
compared to direct numerical simulation (DNS), these various sources of error
can be quite signicant (Ghosal 1996; Geurts 1999) and can lead to intricate
interactions with some unexpected consequences.
The large-eddy modelling of incompressible ow includes ltering of the
convective terms which leads to the turbulent stress tensor
ij
= u
i
u
j
u
i
u
j
; i, j = 1, 2, 3 (1)
as the major so-called subgrid term. This expression contains contributions
from the ltered (u
i
) and the unltered (u
i
) velocity elds and cannot, in
practice, be expressed in terms of the ltered solution alone. Various modelling
strategies have been proposed, some of which nd their origin in physical
arguments while others start from rigorous information about properties of
the stress tensor. We consider some recent modelling strategies which aim at
optimising the use of the scales that are available in an LES in order to arrive at
improvements in the subgrid model. This inverse modelling (Geurts 1997) or
approximate deconvolution (Domaradzki and Saiki 1997; Stolz and Adams
1999) can give rise to models that combine high correlation with suitable
transfer of energy from the resolved to the unresolved scales.
For geometrically complicated ows the use of a convolution lter, i.e. the
use of a constant lter-width, may not be desirable. In such ows one may
observe regions of high turbulence intensity with many small-scale contribu-
tions next to regions of weakly turbulent ow with predominantly large-scale
components in the solution or next to regions of little interest such as the far
wake of a blu-body ow. An ecient LES of such situations calls for spatial
lters with nonuniform width. This, however, gives rise to additional terms
that contribute to inter-scale energy-transfer in a specic way depending di-
rectly on the non-uniformity of the lter-width. This type of contribution has
largely been left unstudied in literature even though its order of magnitude
may be comparable to contributions from the turbulent stress tensor.
722 Geurts and Leonard
In order to arrive at a slightly more quantitative approach to the question
of adequacy of LES it is important to incorporate the type of information one
wants to extract from the simulation. Since this diers widely in the various
application areas the capabilities of present-day LES can already be of some
or even sucient use in certain areas while the same LES capabilities provide
insucient accuracy in other applications, e.g. due to restricted computational
resources and numerical capabilities. To quantify this somewhat, we consider
the prediction of a quantity q, which has an exact value q
e
. If we want to
predict q and allow an error then accurate predictions of only a certain range
of scales in the ow are required. We denote q(n) the value of q evaluated with
the rst n modes of the largest scales. The requirement [q(n)q
e
[ < species
the necessary number of modes n
q
() for the quantity q. If in an LES one can
provide n
q
() modes with sucient accuracy then the quantity q is adequately
predicted. Clearly, a decrease in will lead to a (strong) increase in n
q
() and,
moreover, dierent quantities q will require possibly very dierent values of
n
q
() in order to yield the desired accuracy.
However, the number of modes given by n
q
() is ideal assuming modelling
errors and numerical errors are negligible. Thus a central question is how
LES should be designed to arrive at an accurate prediction of at least these
rst n
q
() modes. Because of these two types of errors any LES aiming at an
adequate prediction of the quantity q should involve N
LES
> n
q
() modes.
The success of LES clearly depends on how much larger N
LES
needs to be,
compared to the absolute minimum n
q
(). One might expect good subgrid
models as well as good numerics to lead to a (slight) decrease in N
LES
. The
latter will be illustrated in this chapter. Other factors may, however, lead to
an increase in the required N
LES
. For example, a somewhat higher numerical
value for the minimal resolution may exist, e.g., in order to provide a stable
treatment of ow near walls or interfaces or to provide enough resolution to
generate a stable solution in case inow and outow boundaries are introduced
such as in spatially developing ows. It is clear that the lower limit n
q
() needs
to be respected in any LES and this may well explain some of the reported
failures of LES in predicting certain quantities in complex ows. A number of
test-cases have been studied with LES by a large number of groups in recent
years (see e.g. Rodi et al. 1997 and [12]) and several appear very illustrative
in relation to the failure to reach the lower value n
q
().
For a strict DNS the number of modes N
DNS
that needs to be incorporated
is related to the reciprocal Kolmogorov length. Of course the main virtue of
LES is that N
LES
is related to the reciprocal lter-width, i.e. 1/ so that
N
LES
can be much lower than N
DNS
. At the same time the desire to predict
q with an accuracy implies that N
LES
> n
q
() which therefore provides a
strict upper-bound for . Stated dierently, there is an obvious limit to the
amount of information that can be ltered away if one insists in maintaining
a minimal accuracy for the prediction of some ow property. In order to have
[25] Is LES ready for complex ows? 723
some independent control over the interaction between numerical and mod-
elling errors in the subsequent LES the resolution should be ne enough. In
particular this implies that the mesh-size h has to be suciently smaller than
. Appropriate values for the ratio /h depend on the spatial discretization
scheme that is used but should be at least larger than 2 or 4 for fourth- or
second-order methods, respectively. We will illustrate this below. In total, the
desired accuracy with which q needs to be predicted fully controls an upper-
bound for the mesh-size. Actual LES is awed and constrained in many ways
compared to the optimal eciency LES which would need only n
q
() modes.
Thus the actual number of modes used, N
LES
, must be larger than n
q
() and
the corresponding lter-width and resolution, h, need to be smaller. Better
numerics and modelling can help to increase acceptable values of and h and
vice versa, but never beyond bounds set by n
q
().
The organisation of this chapter is as follows. In section 2 we briey formu-
late the ltering approach to LES and identify some basic properties of the
LES modelling problem that deserve to be incorporated into any modelling
attempt. Recent developments in subgrid modelling which aim at incorporat-
ing information from the scales that are directly available in an LES will be
considered. Section 3 describes additional complications to the LES equations
that arise from extension to complex ows. This involves the use of nonuni-
form lters and gives rise to additional terms in the ltered equations which
have a specic contribution to the inter-scale energy transfer that will be in-
terpreted and estimated. The numerical treatment of the LES equations is
another source of unavoidable and sometimes surprising errors. Section 4 is
devoted to some unexpected consequences and paradoxes associated with the
interaction of numerical and modelling errors. All sources of error in LES can
be controlled to some degree at the expense of adding to the computational
eort. Some guidelines for LES which aim at keeping the mixture of errors
within reasonable bounds will be suggested in section 5 where we also collect
some concluding remarks.
2 The ltering and inversion approach to LES
In this section we briey introduce the ltering of the NavierStokes equations
to derive the governing equations for large-eddy simulation. Some properties
of the ltered equations will be mentioned which all have consequences for the
underlying modelling problem. As an illustration, algebraic properties of the
turbulent stress tensor will be considered in more detail as well as the use of
approximate inversion techniques and dynamic modelling.
The starting point in the ltering approach is the introduction of the l-
ter operator L which is used to lter the NavierStokes equations. The lter
considered in this section is a convolution lter and in one dimension this is
724 Geurts and Leonard
dened by
u(x, t) = L(u) =
_
L
ij
= u
i
u
j
u
i
u
j
= L(S(u
i
, u
j
)) S(L(u
i
), L(u
j
))
= [L, S](u
i
, u
j
). (3)
Here we introduced the commutator of the lter L with the product operator
S for later convenience. This relation shows that the basic modelling problem
in LES is completely identied with properties of the commutator [L, S]. The
ltered NavierStokes equations take on the same form as the unltered equa-
tions with the exception that the divergence of the turbulent stress tensor ap-
pears as an extra term in the ltered equations. With the unltered equations
written as NS(u) = 0 with NS a symbolic notation for the NavierStokes
operator, the ltered equations can be written as NS(u) =
j
L
ij
. In this
way the divergence of the turbulent stress tensor appears as a source-term
for the evolution of the ltered solution. In practice it cannot be expressed in
terms of the ltered solution alone and ideally would require full knowledge
of the unltered solution. Thus the closure problem in LES is to nd suitable
expressions for
L
in terms of u alone. This modelling process is central in LES
and can be guided by incorporating any sound physical properties of small-
scale phenomena in turbulent ow or by taking into account the mathematical
structure of the ltered equations. We illustrate some of the latter possibilities
next.
The ltered equations have a number of rigorous properties which can
be used to assist in the modelling process. There are several symmetries of
the ltered equations known, such as translational and rotational symmetry,
[25] Is LES ready for complex ows? 725
Gallilean invariance and scale invariance (Ghosal 1999) involving an equiva-
lence transformation. Similarly the realisability requirements for the turbulent
stress tensor (Vreman et al. 1994) can be used to restrict the multitude of
possibilities for modelling
L
. We will not consider these properties here but
instead turn to algebraic properties of
L
and their use in subgrid modelling.
The commutator dening the turbulent stress tensor
L
shares a number of
properties with the Poisson-bracket in classical mechanics. An important prop-
erty of Poisson-brackets is in the context of LES known as Germanos identity
(Germano 1992)
[L
1
L
2
, S] = [L
1
, S]L
2
+L
1
[L
2
, S]
i.e.
L
1
L
2
=
L
1
L
2
+L
1
L
2
(4)
where L
1
and L
2
denote any two lter operators and
K
= [K, S] is the
turbulent stress tensor associated with a lter K. Similarly Jacobis identity
holds for S, L
1
and L
2
:
[L
1
, [L
2
, S]] + [L
2
, [S, L
1
]] = [S, [L
1
, L
2
]]
i.e.
[L
1
,
L
2
] [L
2
,
L
1
] =
[L
1
,L
2
]
. (5)
This formulation of the Jacobi identity holds for general lters. In case convo-
lution lters are considered the right hand side in (5) is zero. The expressions
in (4) and (5) provide relations between the turbulent stress tensor correspond-
ing to dierent lters and can be used to dynamically model
L
. The success
of models incorporating (4) is by now well established and applied in many
dierent ows. In the traditional formulation one selects L
1
= H and L
2
= L
where H is the so called test-lter. In this case one can specify Germanos
identity as
HL
(u) =
H
(L(u)) +H
_
L
(u)
_
. (6)
The rst term on the right hand side involves the operator
H
acting on the
resolved LES eld L(u) and during an LES this is known explicitly. The re-
maining terms need to be replaced by a model. In the dynamic modelling
approach the next step is to assume a base-model m
K
corresponding to the
lter-level K and optimise any coecients in it in accordance with e.g. an
optimal compliance with the Germano identity in a least squares sense (Lilly
1992). Several choices for the base model have been used varying from the
Smagorinsky eddy-viscosity model to mixed versions consisting of similarity
models, e.g. Bardinas model or the tensor-diusivity model, combined with
an eddy-viscosity term. The rst base model gives rise to the dynamic model
and the second option to what is known as the dynamic mixed models. In
actual simulations this approach has proven to be very successful, mainly be-
cause these models avoid excessive dissipation in relatively quiescent regions
726 Geurts and Leonard
of the ow whereas appropriately high values of eddy-viscosity arise in regions
with large turbulent intensity. In implementations of the dynamic procedure
shortcomings of the assumed base model can require some technical adjust-
ments. As an example, the use of dynamic eddy viscosity is not guaranteed to
yield positive and relatively smoothly varying dynamic coecients. This could
lead to numerical instabilities and for that reason clipping and averaging over
suitable parts of the ow domain are introduced. The self-adjusting property
of the model-parameters proceeds dynamically in accordance with the local in-
stantaneous ow properties and does not require ad hoc parameters other than
in specifying the test-lter H. In addition, the dynamic modelling is appealing
in many applications because it displays a self-restoring feedback mechanism.
In fact, an under-prediction of the dynamic eddy viscosity typically tends to
lead to a slight increase of small scale components of sizes comparable to the
lter-width which in turn will increase the eddy viscosity and thus remove
some of the newly arisen small scale components. This feedback has several
appealing consequences for applications of LES. A quite complete comparison
of a large number of subgrid models, combining the ideas of energy-dissipation,
similarity and Germanos identity for turbulent ow in a temporal mixing layer
can be found in Vreman et al. (1997). Recently, the use of inverse modelling
approaches has been developed which gives rise to a further development of
dynamic mixed subgrid models. We give a brief illustration of this next.
The operator formulation allows to readily identify generalised similarity
models which involve approximate inversion dened by L
1
(L(x
k
)) = x
k
for
0 k N (Geurts 1997). With this operation it is possible to partially re-
construct the unltered solution u from the ltered solution u and use this
information in the denition of a subgrid model. Without any inversion the
original similarity model by Bardina et al. (1983) can be written in a shorthand
notation, as m
B
= [L, S](L(u)), i.e. applying the denition of the turbulent
stress tensor directly to the available ltered eld. A direct generalisation of
this arises from m
GB
= [L, S]
_
L
1
(L(u))
_
using the approximate inversion.
This model was analysed in a kinematic simulation as well as for single Fourier
modes. Compared to the original Bardina similarity model the generalised
model showed to combine high correlation with improvements in dissipative
properties while retaining the possibility to represent backscatter of energy.
More recently the approximate inversion was combined with dynamic mod-
elling (Kuerten and Geurts 1999). This was based on the choice L
1
= H and
L
2
= H
1
L for which Germanos identity can be specied to
L
(u) =
H
_
H
1
L(u)
_
+H
_
H
1
L
(u)
_
. (7)
Compared to the traditional formulation which involves the modelling of terms
which correspond to length scales
L
and
HL
this extension which incorpo-
rates the (approximate) inverse of the test-lter H
1
requires modelling of
terms on the scale of
L
as before and
H
1
L
. Since
H
1
L
L
HL
the
[25] Is LES ready for complex ows? 727
terms that require modelling are smaller and at the same time it is easier to
maintain modelling assumptions, e.g. involving properties of an inertial range.
Dynamic models based on the above have been applied successfully. A further
extension involving repeated application of H
1
is also possible formally. How-
ever, since (approximate) inversion is not a very well behaved operation for
the smaller scales, in actual applications one faces the risk of reconstructing
small scale contributions which have been contaminated with possible numeri-
cal artifacts. Therefore there is a clear practical upper-bound to the number of
times H
1
can be used benecially and from recent experience 3 or 4 appears
a denite upper-bound.
Another way of optimising the use of the information contained in an LES
which is more implicit gives rise to the tensor-diusivity or gradient model.
Again, the basis for this model is, essentially, approximate deconvolution. Con-
sider, for example, the Gaussian lter
G(z) =
exp(z
2
/
2
)
. (8)
For such a lter we nd that the commutator
L
is given by
uv u v =
k=1
_
2
2
_
k
1
k!
k
u
x
k
k
v
x
k
. (9)
The full innite series above is equivalent to deconvolving u and v (clearly a
singular operation) then forming the product uv and then applying the lter
operation. Taking only the rst term in the series above as a subgrid model
we have in d dimensions
uv u v
2
2
u
x
v
x
(10)
where repeated indicies are summed and = 1, 2, . . . , d. Use of this approxi-
mation on the ltered, constant density, incompressible momentum equation
for u
i
yields the following subgrid force on the RHS
2
2
S
jk
2
u
i
x
j
x
k
(11)
where S
jk
is the strain-rate tensor of the ltered velocity eld. Hence the term
tensor diusivity model which dates back to Leonard (1974). Similarly one
could arrive at this model using a Taylor expansion on the Bardina model.
A direct application of this model in LES can lead to an ill-behaved system
of equations as was analysed in Vreman et al. (1996b). However, the appeal-
ing property of being able to represent backscatter, without the need of an
additional ltering, can be retained if this base model is combined with a dy-
namic eddy-viscosity. In that case a computationally ecient and competitive
subgrid model is arrived at (Vreman et al. 1997; Leonard and Winckelmans
1999).
728 Geurts and Leonard
3 Non-uniform lters and LES of complex ows
The desire to extend LES to complex ows in an ecient way implies that one
typically encounters situations in which the turbulence intensities vary consid-
erably within the ow domain. In certain regions of the ow a nearly laminar,
smoothly evolving ow may arise while a lively, ne scale turbulent ow can
be present in another region. This calls for a ltering approach involving a
lter with a non-uniform lter-width. Here we will study some consequences
of applying such lters and in particular identify and estimate the additional
terms that arise from using a variable lter-width.
We consider the eect of applying a general, compact support lter which
is dened by u u:
u(x, t) L(u) =
_
x+
+
(x)
x
(x)
H(x, )
(x)
u(, t)d (12)
where
+
,
j
u
j
= 0 ;
t
u
i
+
j
(u
i
u
j
) +
i
p
1
Re
jj
u
i
= 0 ; i = 1, 2, 3 (13)
where p denotes the pressure, Re the Reynolds number and the summation
convention is adopted. If we apply the lter to the system of equations in (13)
commutators of L with partial derivatives and multiplication arise. After some
manipulation we nd:
j
u
j
= [L,
j
](u
j
). (14)
This shows that application of a non-convolution lter to the continuity equa-
tion gives rise to terms which in general violate the local conservation form.
Filtering the NavierStokes equations yields
t
u
i
+
j
(u
i
u
j
) +
i
p
1
Re
jj
u
i
=
_
[L,
i
](p)
1
Re
[L,
jj
](u
i
) + [L,
j
](S(u
i
, u
j
)) +
j
([L, S](u
i
, u
j
))
_
. (15)
in which the NavierStokes operator applied to the ltered eld is identied
on the left hand side. The rst three terms on the right hand side are related to
[25] Is LES ready for complex ows? 729
commutators of L and partial derivatives
j
which as before implies violation of
the conservation property in general. The turbulent stress tensor arises in the
last term on the right hand side. It is the only lter-term in case convolution
lters are adopted while more general lters yield the full system of equations
(14) and (15). The central modelling problem for the continuous formulation
and general lters is now extended to approximations modelling commutators
like [L,
x
], [L,
xx
] and [L, S] in terms of operations on u. A mathematically
consistent modelling can be arrived at using approximate inversion (Geurts
1997) which can be extended to non-uniform lters in a consistent way with
the help of symbolic manipulation software such as maple. The new terms that
have arisen can be shown to obey the same algebraic identities as put forward
in relation to [L, S]. These identities can be used e.g. in a dynamic modelling
of the commutators of non-uniform ltering and partial derivatives which need
to be taken into account for complex ows.
In order to establish the importance of these new commutators in relation
to the regular lter terms [L, S] and to interpret in what way these terms
contribute to the inter-scale energy transfer we analyse the commutators for
general high order lters acting on suciently smooth signals (Geurts et al.
1997). Such Nth-order lters are dened by requiring the rst N moments to
be invariant, i.e. L(x
k
) = x
k
for k = 0, 1, . . . , N 1. In the following we will
apply such lters and retain only the leading order terms assuming suciently
smooth signals for the moment. In one dimension one may readily show that
u(x) = u(x) +
_
N
(x)M
N
(x)
_
u
(N)
(x) + (16)
for Nth-order lters. Here M
N
L(x
N
) is the Nth-order moment. If we turn
to the decomposition of a typical term:
x
(u
2
) =
x
(u
2
) +
x
([L, S](u)) + [L,
x
](S(u)) (17)
one observes the commutators [L, S] and [L,
x
] to arise naturally. As derived
in detail e.g. in Geurts et al. (1997) and Geurts (1999) one nds corresponding
expressions for these commutators given by
[L,
x
](S(u)) = A(x)
_
N1
M
N
_
+B(x)
_
N
M
N
_
+
x
([L, S](u)) = a(x)
_
N1
M
N
_
+b(x)
_
N
M
N
_
+c(x)
_
N
M
N
_
+
where A, B, a, b and c are smooth functions containing combinations of deriva-
tives of the solution u. From this we infer that a constant lter-width implies
[L,
x
] = 0 and the leading order term of the turbulent stress tensor equals
[L, S]
N
for Nth-order lters. However, non-uniform lters clearly give
rise to contributions to the commutators which are a priori of equal order
of magnitude. From this we infer, unlike ndings in Van der Ven (1995) and
730 Geurts and Leonard
Vasilyev et al. (1998) that it is not possible to remove the commutators [L,
x
]
by a careful selection of the lter. In fact, all lters that would reduce this
commutator are of higher order and consequently will also reduce the usual
term [L, S]. The only possibility to control [L,
x
] independently is by reducing
non-uniformity of the lter-width, e.g. by keeping grid-nonuniformity, which
usually denes local lter-widths, small. We will quantify this to some extent
next and illustrate the dynamic eect associated with the new commutators.
A detailed analysis of the new commutators can be obtained e.g. in a
single-wave analysis in which we assume a solution u = sin(x). For illus-
tration purposes we consider a symmetric top-hat lter for which one has
[L,
x
](sin(x)) = /
=
_
cos(/2) 2 sin(/2)
(/2)
_
=
1
24
(()
2
)
+
1
1920
(()
4
)
+ .
(18)
In order to appreciate the magnitude of this commutator compared to [L, S]
in a dynamic context we recall that [L,
x
](S(u)) needs to be compared to
x
([L, S](u)). After some manipulation we nd
[L,
x
](S(u)) +
x
([L, S](u)) = C((x))sin(2x) +
(x)
(x)
(cos(2x) 1)
(19)
where the characteristic ux function C for this lter is given by
C(z) =
z sin(z) 2 + 2 cos(z)
z
2
=
1
12
z
2
+
1
180
z
4
+O(z
6
). (20)
The two contributions to the ux have a weight and (
/) respectively
from which we infer that if variations in are suciently slow, i.e. [
[ [[
then lter-width non-uniformity can be disregarded. We infer that the dynamic
eect of the new commutator is related to the sign of
j
(u
i
u
j
) +
i
p = [
j
(u
i
u
j
) +
i
p +T
i
] +
j
ij
= [
j
(u
i
u
j
) +
i
p +T
i
] + [
j
m
ij
+
i
]
=
j
(u
i
u
j
) +
i
p +
j
m
ij
+ [T
i
+T
(m)
i
+
i
] (21)
where T
i
denotes the discretization error arising from application of a spa-
tial discretization method
j
to the convective terms, T
(m)
i
is the error when
implementing the model m
ij
, e.g. ltering as well as discretization errors and
i
=
j
(
ij
m
ij
) is the total model-residue associated with m
ij
. This term
can only be determined in a priori evaluations and is of course unknown dur-
ing an actual LES. So, whereas formally
j
(u
i
u
j
) +
i
p is needed in an LES
strictly speaking only
j
(u
i
u
j
) +
i
p is directly available and two main sources
of discrepancy can be identied. Whereas the subgrid-term
j
ij
is usually
modelled with a subgrid-model, the discretization error T
i
is not taken into
account. The rst question is whether this is justied and for this purpose an a
priori comparison of dierent spatial discretization methods and lter-widths
was made for turbulent ow in a mixing layer (Vreman et al. 1994, 1996a).
The magnitude and ratio of the discretization error T
i
and the ux due to the
turbulent ux
j
ij
determines in large part the reliability of LES predictions.
We evaluated these terms for a well developed ow and evaluated the errors
732 Geurts and Leonard
associated with a second- and a fourth-order nite volume discretization oper-
ator
j
. It was shown that if = h, the discretization error T
i
is larger than
the subgrid term for both methods and in this case LES predictions would not
be reliable even with a perfect subgrid-model for . If is suciently larger
than h, i.e. smoother elds are represented on the same grid, the contribution
of
j
ij
is considerably larger than T
i
. In this regime the second-order method
shows only a relatively small decrease of T
i
for increasing , whereas the
fourth-order method shows a rapid decrease of T
i
. This was observed for both
ne and very coarse LES-grids and can serve as a guidance in selecting for
a specic discretization method on a given grid. An interesting related study
can be found in Mason and Callen (1986) in which the Smagorinsky constant
was varied at xed grid resolution. When is large the ltered elds become
smoother which reduces the discretization error at the expense of containing
only little information about the smaller scales. A good compromise appears
the choice = 2h for the fourth-order method while a second-order method
requires a higher value of /h. In that case fourth-order discretizations are
more ecient than second-order ones.
The second question we address is how the dierent errors T
i
, T
(m)
i
and
i
interact dynamically. We consider the dynamic mixed model in combina-
tion with the discretization schemes used above as well as a pseudo-spectral
method. Discrepancies between LES and ltered DNS results arise mainly
from shortcomings of the model and from numerical discretization on a rel-
atively coarse grid. In LES these sources of error interact which complicates
testing, since separation of subgrid-modelling and numerical eects is dicult
(Meneveau 1994). We propose approximate separation of the eects of mod-
elling and discretization error by incorporating LES at higher resolution. We
consider the evolution of the total kinetic energy E:
E =
_
1
2
u
i
u
i
dx (22)
where is the ow domain. As a function of time E displays a gradual de-
crease. Variations in spatial discretization method show variations in the pre-
dictions for E whose magnitude is of the same order as would arise when chang-
ing from a dynamic subgrid model to e.g. the Bardina similarity model. These
eects of the errors increase considerably if we use = h instead of = 2h.
We can approximately separate the modelling and discretization eects and
focus on their interaction by incorporating a ne-grid LES. The discretization
error in LES will become smaller if the resolution is increased at constant .
The discretization error in such a ne-grid LES will be considerably smaller
and we can obtain LES predictions with negligible discretization error eects.
The dierence between these two large-eddy simulations can then give an in-
dication of the eect of the discretization error:
d
= E
LES
E
ne-grid LES
whereas the dierence between the ne-grid LES and the ltered DNS mea-
sures the eect of the modelling error:
m
= E
ne-grid LES
E
ltered DNS
.
[25] Is LES ready for complex ows? 733
0 20 40 60 80 100
4000
3000
2000
1000
0
1000
2000
time
e
r
r
o
r
s
i
n
t
h
e
t
o
t
a
l
k
i
n
e
t
i
c
e
n
e
r
g
y
0 20 40 60 80 100
2000
1000
0
1000
2000
3000
4000
time
e
r
r
o
r
s
i
n
t
h
e
t
o
t
a
l
k
i
n
e
t
i
c
e
n
e
r
g
y
Figure 1: Error decomposition with modelling error (solid) and discretizations
error eect for the three discretization methods: dashed (second-order), dash-
dotted (spectral) and dotted (fourth-order). Left gure = h and right gure
= 2h.
We calculated these errors for the turbulent ow in a mixing layer on a
representative grid which is about six times coarser in each direction than
required for DNS. The corresponding ne-grid LES has been performed with
the ratio between and h in the ne-grid LES suciently large, i.e. the elds
are quite smooth on grid-scale and discretization errors will be considerably
reduced. The quantities
d
and
m
are shown in gure 1. The discretization
error eects are smaller than the modelling error only if 2h whereas with
implicit ltering
d
is even larger than
m
. The second-order scheme is ob-
served to give the smallest discretization error eect for this quantity. This
does not imply that the discretization error itself is small, but only its eect
on the evolution of the total kinetic energy. For the fourth-order and pseudo-
spectral methods the discretization error and modelling error eect have oppo-
site sign, which implies that the discretization error assists the subgrid-model
in the representation of this quantity: the total error is considerably smaller
than the modelling error. These observations suggest that e.g. for the spectral
scheme, improvement of the subgrid-model (decrease of the modelling error)
is expected to give worse results, since the total error will increase. Likewise
one can infer that an increase in the resolution may result in worse predic-
tions. All these errors and their interactions can be extremely disturbing for
more complex ows for which no proper separation is available and one has
to rely on intuition and previous experience in order to judge and justify the
outcome of a particular simulation. Since in the past this has proven to be
a very underdeveloped area, we list some (hopefully) useful guidelines in the
next section.
734 Geurts and Leonard
5 Some guidelines for predictable LES
In this section we will formulate some general guidelines which can enhance
the credibility of a ow simulation within the LES approach. Since there has
been a large development in the capabilities of computers and numerical meth-
ods, it has become possible to use some of these capabilities to systematically
vary certain numerical elements within a reference LES and monitor the sen-
sitivity of the predictions. Since LES is in many respects close to a direct
numerical simulation, several of the guidelines are of relevance for numerical
reasons only whereas certain suggestions are more specic to the LES context.
The following list which compiles the guidelines is necessarily incomplete and
somewhat biased given the fact that LES is still a lively and rapidly developing
eld of research. Moreover, we have put forward some guidelines which may
add too much to the computational cost; however, we have taken the liberty to
formulate an LES approach focusing more on reliability than on strict saving
of computational eort. Eventually, LES can oer an expensive and reliable
answer irrespective of the quality of the subgrid model and in part also quite
independent of the quality of the numerical methods involved, provided the
subgrid contributions are suciently reduced. This escape-route, however,
would be virtually identical to a well resolved DNS and be not very practical
in most cases. However, the fact that LES has this limit built into it can be
used also to infer about the reliability of any given reference LES.
The list of guidelines presented below has been split into mainly numerical,
mainly modelling, and interaction issues. There can be a strong interaction
and interdependence between modelling and numerics and for that reason
some points are described under more than one topic. As a whole, an LES is
as strong as its weakest element (as are many other approaches).
Numerical guidelines:
Use smooth grids with low stretching and skewness and equal resolu-
tion in each direction. This is of importance since the formal as well as
the attained accuracy of spatial discretization schemes can be consider-
ably aected by either shortcoming of the grid. Moreover, the resolution
should be comparable in each of the coordinate directions in order
to avoid contamination of the solution in an under-resolved direction
through folding back of energy contained in modes which are well re-
solved in another coordinate direction. This obviously is ow-dependent
and is associated with the degree of nonhomogeneity in the ow locally.
Avoid numerical dissipation. In particular for turbulent ow the presence
of some numerical dissipation can cover shortcomings e.g. in resolution,
grid properties or modelling used in the approach. Although this may
appear helpful in cases in which the resolution is too low anyway, it adds
to the unreliability of LES and can seriously aect the predictions, in
[25] Is LES ready for complex ows? 735
particular if the same approach would be used for other ows, at other
resolutions or for other ow conditions.
Validate your code. In order to eliminate as much as possible numer-
ical artifacts and remaining uncertainties regarding resolution, inow
and outow conditions, geometrical description of the ow domain etc.
validation is essential. This could incorporate comparison with simpler
theories, e.g. using linear stability theory, checking whether basic sym-
metries of the equations are also contained in the numerical formulation,
using experimental data if available and comparison with available l-
tered DNS data in case validation for simple ows is included. It is
sensible to have some discipline of version management of the software.
Vary numerical parameters. In any numerical study a certain number of
relevant numerical parameters appear and basically no physically rele-
vant prediction should depend on any of these parameters. Variations in
resolution, denition of geometry, inow/outow boundaries, numerical
method and method of evaluation of the simulation results should be
considered for LES as well as DNS.
Incorporate LES predictions at dierent lter-width to mesh-size ratios
into a ow analysis. As was illustrated in the previous section a carefully
selected set of LES predictions can be used to appreciate the inuence of
some of the errors involved and to some extent one could estimate these
errors from the combined LES predictions. With present day comput-
ers it is now feasible to conduct several LES investigations of one ow
and from it extract a quantitative appreciation of the reliability of the
predictions.
Modelling guidelines:
Use dynamic modelling. Dynamic modelling is appealing since it does not
add any ad hoc parameters to the modelling other than properties of the
test-lter. The approach has proven to be quite robust and possesses a
self-restoring property if the resolution is suciently high. Moreover, at
suitable resolution it avoids the introduction of special wall treatment.
Both properties give rise to a prediction of the large-scale ow which
is quite robust, suitable for transitional and turbulent ow and can be
used in quite general inhomogeneous ows.
Use explicit ltering approach. In contrast to the implicit ltering setting
of LES the explicit ltering oers independent control and treatment
of the various steps relevant within LES. Using all available resolution
aimed only at predicting small scale properties of the ow can lead to
considerable unreliability and contamination of much of the predictions
736 Geurts and Leonard
of all the other scales. In the explicit ltering some of the resolution
can be used to increase the reliability with which the smaller scales are
predicted.
Incorporate similarity and dissipation into modelling. Both these prop-
erties arise naturally from spectral considerations of LES; similarity is
an inertial range property of the turbulent stress tensor itself and the
required energy transfer to smaller scales is eciently represented by dis-
sipation although more rened ways of energy transfer may be required
if other elements in the LES approach become more rened.
Use smooth grids with low stretching and skewness and equal resolu-
tion in each direction. For modelling this point is relevant since high
stretching and skewness give rise to signicant additional terms in the
equations which require to be modelled. Similarly, unevenness in the grid
can lead to sizeable eects in the implementation and evaluation of the
actual model and obscure much of the models potential.
Incorporate LES predictions at dierent lter-width to mesh-size ratios
into a ow analysis. From a carefully designed set of LES predictions
one could infer errors arising from the subgrid modelling.
Vary numerical parameters. The inuence of the subgrid model can be
controlled to some extent by a suitable change in the numerical param-
eters. Moreover, several models require additional ltering and dieren-
tiation and the number of points available to do this is usually quite
restricted and hence can make the implemented model appear to have
dierent properties compared to the continuum formulation.
Optimise use of scales available in LES. The options for LES modelling
oered by inverse modelling, approximate deconvolution and/or subgrid
estimation have not been fully exploited and can be benecial to LES.
Restricted interaction guidelines:
Choose /h appropriately. The ratio of the lter-width to (local)
mesh-size h is an important parameter in LES. If it is large then the
LES prediction will appear smooth on the grid-scale and the quality of
the prediction will be mainly restricted by the quality of the subgrid
model. Conversely, if this ratio is small then the eects of numerics will
be large. In practice a ratio /h 2 appears adequate when fourth-
order methods are employed but this ratio should be increased to 4 if
second-order methods are used.
Incorporate LES predictions at dierent lter-width to mesh-size ratios
into a ow analysis.
[25] Is LES ready for complex ows? 737
In order to develop LES for general complex ows the treatment of the near
wall region is crucial and not yet well developed in LES. Similarly, if shocks
or detailed capturing of chemistry is required also including multi-phase ows
then a proper modelling of this near interface region is vital. Finally, an ap-
preciation and possibly an estimate of the error in the LES predictions should
be aimed at. For this purpose the use of approximate inversion of the ltering,
the compliance with algebraic properties and other rigorous characteristics of
the subgrid terms and a systematic variation of e.g. the resolution, indepen-
dent of the lter width, should be developed. The lter-width (with suitable
h) and a certain subgrid model imply a certain number of modes N
LES
to be
involved which should at any rate be suciently larger than the number of
modes n
q
() needed to predict a quantity q with a desired level of accuracy.
It appears relevant to quantify suitable numbers n
q
and corresponding N
LES
for several geometrically simple ows which are well documented and allow
for a fully resolved DNS as well in order to provide a well controlled point of
reference. From an estimate of N
LES
for these ows and a number of subgrid
models and numerical methods it would be possible to formulate more general
selection and design rules for reliable LES in the future. For the modelling pro-
cess it is advisable to respect rigorous guidelines (e.g. symmetries, realisability,
algebraic properties, inequalities) and to formulate some error monitoring and
control aiming at the prediction of an error-bar during a simulation.
Acknowledgements
Stimulating and inspiring discussion with several participants in the turbulence
programme of the Isaac Newton Institute has been much appreciated and
helpful. One of us (BJG) would like to thank the European Science Foundation
for supporting part of this work.
References
Bardina, J., Ferziger, J.H., Reynolds, W.C. (1983). Improved turbulence models
based on large eddy simulations of homogeneous incompressible turbulence, Stan-
ford University, Report TF-19.
Domaradzki, J.A., Saiki, E.M. (1997). A subgrid-scale model based on the estimation
of unresolved scales of turbulence Phys. Fluids 9, 1.
Germano, M. (1992). Turbulence: the ltering approach, J. Fluid Mech. 238, 325.
Germano, M., Piomelli U., Moin P., Cabot W.H. (1991). A dynamic subgrid-scale
eddy viscosity model, Phys. Fluids 3, 1760.
Geurts, B.J. (1997). Inverse modeling for large-eddy simulation, Phys. Fluids 9, 3585.
Geurts B.J., Vreman A.W., Kuerten J.G.M., Van Buuren R. (1997). Non-commuting
lters and dynamic modelling for LES of turbulent compressible ow in 3D shear
738 Geurts and Leonard
layers. In Direct and Large-Eddy Simulation II: Grenoble, P.R. Voke, L. Kleiser,
J.P. Chollet (eds.), Elsevier, 47.
Geurts B.J., Vreman A.W., Kuerten J.G.M. (1994). Comparison of DNS and LES
of transitional and turbulent compressible ow: at plate and mixing layer. In
74th Fluid Dynamics Panel and Symposium on Application of DNS and LES to
transition and turbulence, Crete, AGARD Conf. Proceedings 551:51.
Geurts B.J. (1999). Balancing errors in LES. In Direct and Large-Eddy Simulation
III, Proceedings of the Isaac Newton Institute Symposium/ERCOFTAC Workshop,
Cambridge UK, 1214 May, 1999, N.D. Sandham, P.R. Voke, L. Kleiser (eds.),
Elsevier, 112.
Ghosal, S. (1996). An analysis of numerical errors in large-eddy simulations of tur-
bulence, J. Comp. Phys. 125, 187.
Ghosal, S. (1999). Mathematical and physical constraints on large-eddy simulation
of turbulence, AIAA J. 37, 425.
Kuerten J.G.M., Geurts B.J. (1999). Dynamic inverse modelling in les of the temporal
mixing layer, Proceedings Second AFOSR Conference on DNS and LES.
Leonard, A. (1974). Energy cascade in large-eddy simulations of turbulent uid ows,
Adv. Geophys., 18, 237.
Leonard, A., Winckelmans, G.S. (1999). A tensor-diusivity model for Large-eddy
Simulation. In Direct and Large-Eddy simulation III: Cambridge, Proceedings of
the Isaac Newton Institute Symposium/ERCOFTAC Workshop, Cambridge UK,
1214 May, 1999, N.D. Sandham, P.R. Voke, L. Kleiser (eds.), Elsevier, 147162.
Lilly, D.K. (1992). A proposed modication of the Germano subgrid-scale closure
method, Phys. Fluids A 4, 633.
Mason, P.J., Callen, N.S. (1986). On the magnitude of the subgrid-scale eddy co-
ecient in large-eddy simulations of turbulent channel ow, J. Fluid Mech. 162,
439.
Meneveau, C. (1994). Statistics of turbulence subgrid-scale stresses: Necessary con-
ditions and experimental tests, Phys. Fluids 6, 815.
Piomelli, C., Moin P., Ferziger J.H. (1998). Model consistency in large eddy simula-
tion of turbulent channel ows, Phys. Fluids 31, 1884.
Rodi, W., Ferziger J.H., Breuer M., Pourquie M. (1997). Status of large-eddy simu-
lation: results of a workshop, Trans. ASME 119, 284.
Stolz, S., Adams, N.A. (1999). An approximate deconvolution procedure for large-
eddy simulation, Phys. Fluids, 11, 1699.
Vasilyev, O.V., Lund, T.S., Moin, P. (1998). A general class of commutative lters
for LES in complex geometries, J. Comput. Phys. 146, 82.
Van der Ven, H. (1995). A family of large eddy simulation lters with nonuniform
lter widths, Phys. of Fluids 7, 1171.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1994). Realisability conditions for the
turbulent stress tensor in large eddy simulation, J. Fluid Mech. 278, 351.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1995). A priori tests of Large Eddy
[25] Is LES ready for complex ows? 739
Simulation of the compressible plane mixing layer, J. Eng. Math. 29, 299.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1994). Discretization error dominance
over subgrid-terms in large eddy simulations of compressible shear layers, Comm.
Num. Meth. Eng. Math. 10, 785.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1996a). Comparison of numerical
schemes in Large Eddy Simulation of the temporal mixing layer, Int. J. Num.
Meth. in Fluids 22, 297.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1996b). Large eddy simulation of the
temporal mixing layer using the Clark model, Theor. Comp. Fluid Dyn. 8, 309.
Vreman A.W., Geurts B.J., Kuerten J.G.M. (1997). Large-eddy simulation of the
turbulent mixing layer, J. Fluid Mech. 339, 357
26
Recent Developments in Two-Point Closures
Claude Cambon
Abstract
Extensions of the frontiers of rapid distortion theory (RDT) and multi-point
closures are discussed, especially developments leading towards inhomogeneous
turbulence. Recent works related to zonal RDT and stability analyses for
wavepacket disturbances to non-parallel rotational base ows are presented.
Application of linear theories to compressible ows are touched upon. Ho-
mogeneous turbulence is revisited in the presence of dispersive waves, taking
advantage of the close relationship between recent theories of weakly nonlin-
ear interactions, or wave-turbulence, and classical two-point closure theories.
Among various approaches to multi-point description and modelling, a review
is given of multi-scale or multi-tensor transport models, which use, more or
less explicitly, a spectral formulation.
1 Inhomogeneous turbulence
Multi-point formulations are not nearly as well-developed for inhomogeneous
as for homogeneous turbulence. An assumption of weak inhomogeneity, in
which variations of the ow statistics take place over distances greater than
O(), the size of the large turbulent eddies, allows some progress to be made,
as, to a lesser extent, does the RDT limit of weak turbulence.
1.1 Linear theories
The solution of RDT with a known mean ow U
i
, arbitrarily varying in space,
is a dicult problem in general, but becomes somewhat simpler if the mean
ow is irrotational, as in the classical case of high Reynolds number ow past
a body, outside the wake and boundary layer. Consider a particle convected
by the ow, having position x
at time t
) =
x
i
x
j
(1)
and the evolution of vorticity for inviscid incompressible ow is then described
by the Cauchy solution
i
(x, t) = F
ij
(x, t, t
)
j
(x
, t
) (2)
740
[26] Recent developments in two-point closures 741
The above formulation is exact and expresses the classical theory of invis-
cid vortical dynamics. However, in the context of RDT for irrotational mean
ows, one can neglect the uctuating part of the velocity compared with the
mean part, x and x
i
. The solution for the latter
can be shown to be
u
i
(x, t) = F
1
ji
(x, t, t
)u
j
(x
, t
) +
x
i
(3)
where the velocity potential can, in principle, be obtained using the incom-
pressibility requirement that .u
i
(x, t) on the entire velocity
eld at time t
. Given U
i
, the above formulation provides
a means of calculating the RDT solution for u
i
, with its Greens function in
physical space (
ij
(x, x
, t, t
F
ij
=
U
i
x
k
F
kj
(5)
with initial conditions F
ij
(x, t
, t
) =
ij
.
In addition to irrotational mean ows, only the case of pure plane shear
attracted signicant interest in classical RDT (Townsend 1976). The latter case
yields solutions for which it is no longer possible to separate pure distortion
from the potential term, as in (3), but advanced semi-analytical solutions exist
in the presence of simple boundary conditions (see Hunt and Carlotti 2000,
for promising new developments).
Assuming weak inhomogeneity, considerably more progress can be made
without the need for irrotationality of the mean ow, although simplications
occur in the latter case. As discussed earlier in [9], turbulence which is ne-
scale compared with the overall dimensions of the ow can be treated under
RDT by following a notional particle moving with the mean velocity. Thus, the
results obtained for strictly homogeneous turbulence can be extended to the
742 Cambon
weakly inhomogeneous case, but with a mean velocity gradient matrix
ij
(t)
which reects the U
i
/x
j
seen by the moving particle (Hunt 1973, Durbin
and Hunt 1980).
This idea has been formalised in the context of ow stability (see the ge-
ometric optics of Lifschitz and Hameiri 1991) using an asymptotic approach
based on the classical WKB method, which is traditionally used to analyse
the theoretical ray limit (i.e. short waves) in wave problems (see e.g. Lighthill
1978). The solution is written as
u
i
(x, t) = a
i
(x, t) exp[(x, t)/] (6)
with a similar expression for the uctuating pressure, where is a real phase
function and is a small parameter expressing the small scale of the waves
represented by (6), while a
i
(x, t) is a complex amplitude which is expanded in
powers of according to the WKB technique. Over distances of O(), one can
use a spatial Taylors series representation for , up to the linear term, and
approximate a
i
as constant. It is then apparent that (6) is locally a plane-wave
Fourier component of wavenumber
i
(x, t) =
1
x
i
(7)
The amplitude a
i
(x, t) in (6) and the corresponding equation for the uc-
tuating pressure are expanded as an asymptotic series in powers of and the
result inserted into the linearised equations without viscosity. At leading order,
one nds that
=
t
+ U
j
x
j
= 0, i.e. the wave crests of (6) are convected
by the mean ow, with its trajectories given by (4). The spatial derivatives of
= 0 yield
i
=
ji
(t)
j
(8)
where, as before,
ij
= U
i
/x
j
and the dot represents the mean-ow material
derivative /t +U
i
/x
i
. At the next order, one obtains
a
(0)
i
= M
ij
(t)a
(0)
j
(9)
after elimination of the pressure using the leading-order incompressibility con-
dition
i
a
(0)
i
= 0, where a
(0)
i
is the leading-order term in the expansion of
a
i
.
Equations (8) and (9) have exactly the same form as the basic equations of
homogeneous RDT (Townsends equations) and therefore, together with (4),
describe the weakly inhomogeneous case at leading order. The only dierence
is that, rather than being simple time derivatives, the dots represent mean-ow
material derivatives, implying that one should follow mean ow trajectories
which dier from one to another.
The stability context is not discussed here for the sake of brevity. One should
just point out that the geometric optics for short wave disturbances can
[26] Recent developments in two-point closures 743
provide real insight into the nature of instabilities e.g. elliptical, hyperbolic,
and centrifugal that occur in non-parallel ow, with and without system
rotation. Classical massive eigenvalue problems provide little or no such insight
(see Cambon 2001; Godeferd, Cambon and Leblanc 2001).
Continuing with pure linear theories, promising applications to compressible
turbulence follow from RDT and geometric optics, as well as from Ribners
(1954) theory for shock-wave turbulence interaction. It seems interesting to
explore the common background of these approaches, dierent though their
applications are. In both RDT and stability analysis, distortion of the distur-
bance eld is mediated by a large-scale base ow, whereas it is induced by
linearised jump relationship across an innitely thin shock wave in the Rib-
ners theory. The latter Linear Interaction Approximation (see Lele 1994 and
[19]), was recently revisited by Fabre et al. (2001) in order to complement an-
alytical calculation of the linear transfer matrix, which connects any upstream
disturbance mode to its downstream counterpart, when the disturbance eld
is passing through a shock wave. Any kind of upstream disturbance (input)
can be constructed, and its output calculated, with an arbitrary combination
of the three modes, solenoidal, acoustic and entropic, which are uncoupled in
the absence of shock waves or mean distortion in the linear limit (e.g. Ko-
vasznay 1953). Previous LIA applications have dealt with homogeneous quasi-
isotropic and incompressible upstream turbulence. Recent work (Fabre et al.
2001) shows the behaviour of an upstream spot of temperature, which turns
into a pair of vortices, as shown in gure 1 by results from LIA and DNS.
This shows how powerful the method is, and illustrates the fact that Fourier
synthesis for disturbances with a linear transfer function (as in homogeneous
RDT!) is not an impediment to address strong inhomogeneity. Homogeneous
RDT was generalized by Simone et al. (1997) towards isentropic compressible
turbulence, using Fourier modes for disturbances to a mean ow with spatially-
uniform velocity gradients, with emphasis on the stabilising-destabilising ef-
fects of compressibility on shear ows. This technique gave access to linear
coupling of pressure, solenoidal and dilatational disturbance modes by strong
mean velocity gradients, possibly including an acoustic mode. On the other
hand, geometric optics can also be applied to weakly compressible turbu-
lence, for wavepackets that are convected along mean streamlines (Eckho and
Storesletten 1978), but the short-wave WKB development discards the acous-
tic mode (long waves). An entropic disturbance mode, however, can be called
into play in a nontrivial way by the latter technique. Hence, homogeneous
RDT and geometric optics are no longer governed by similar disturbance
equations in the weakly compressible case. Simple explanations for dierent
types of compressible modes and their coupling, in the three theories, can be
proposed as follows. In both LIA and homogeneous RDT, there is no physical
cuto due to the external distortion, because the mean velocity gradients are
spatially-uniform in the former, and the shock-wave is innitely thin in the
latter. Hence, disturbances can be broadband or not, and long waves, such as
744 Cambon
Y
/
a
-6
-4
-2
0
2
4
6
Figure 1: Isolines of pressure downstream of a shock wave, for the case where
a Gaussian entropy spot s
/C
p
= exp(r
2
/a
2
) ( is a small amplitude) has
passed through the shock wave. The upstream Mach number is 4 and the
elapsed time corresponds to t = 4a/C
2
, with C
2
the upstream sound speed.
The linear solution for p
/(p
2
) with 0.02 step between isovalues is shown, with
its counterpart from DNS superimposed ( = 0.01 in DNS, and has not to be
specied in linear theory) (courtesy Fabre et al. 2000).
acoustic waves, are not a priori forbidden. This is no longer valid for geometric
optics, in which long waves are not allowed by the WKB development, even if
the small parameter in (6) does not need to be related to the physically rele-
vant length scales of the problem. In turn, the homogeneous assumption is not
consistent with the presence of the entropic disturbance mode in RDT, since
it actually consists of thermal inhomogeneities. It is not necessary to delve
further into this area to be convinced that exciting issues remain for investi-
gation. A more general theory has to be built in order to account for all the
linear interactions between the three (solenoidal, acoustic, entropic) or four
(solenoidal, dilatational, pressure and entropic) modes, for various external
distortions.
1.2 Transport models for statistical quantities
Only incompressible turbulence is considered in what follows. In turbulent
ows, the uctuating eld is not the single component (6), but instead con-
[26] Recent developments in two-point closures 745
sists of a random superposition of such components. The two-point velocity
correlations can be expressed as u
i
(x, t)u
j
(x +r, t)), but the spectral tensor
ij
(k, x, t) now varies (slowly) with position. As one might expect, given the
behaviour of the underlying local Fourier components described above, it can
be shown that, at leading order, weakly inhomogeneous turbulence evolves
according to
ij
+M
ik
kj
+M
jk
ik
= 0 (10)
where the dot now represents the operator
t
+U
i
x
i
ji
i
(11)
and expresses both convection by the mean ow and evolution of the wavenum-
ber of individual Fourier components according to (9). The spectral evolution
equation (10) corresponds to the RDT limit of its homogeneous equivalent,
Crayas equation, provided the dot operator is interpreted appropriately. Thus,
following the mean ow, the leading-order, local spectral tensor
ij
(k, x, t) be-
haves as in homogeneous RDT, being given in terms of its initial values and
the RDT Greens function. The obvious way to incorporate nonlinearity and
viscosity into this description is by employing
ij
+M
ik
kj
+M
jk
ik
= T
ij
D
ij
2
2
ij
(12)
rather than (10) to describe spectral evolution, where T
ij
could be modelled
using a homogeneous spectral closure, as in [9]. For the sake of completeness,
the tensor D
ij
would typically represent inhomogeneous diusion across the
mean streamline.
In practice, problems with the above approach include the assumption that
the turbulence is only weakly inhomogeneous, which may be a reasonable ap-
proximation for ne-scale turbulence incident on a body, outside the boundary
layer and wake, but is not appropriate in general. A second problem is the con-
siderable computer resources required for anisotropic spectral calculations, of
which one needs to conduct one per mean streamline even in the case of a
steady mean ow.
Even for homogeneous turbulence, going beyond the isotropic case entails
a high computational cost for two-point simulations using classical closures,
a cost which is not insignicant compared with that of direct numerical sim-
ulation. Thus, it is currently unattractive to solve the full set of equations
resulting from closures such as DIA, TFM or EDQNM in the inhomogeneous
case without simplications. In this context, two types of approach seem par-
ticularly promising.
The rst type of approach takes inhomogeneity into account via the basis
set of modes used to express the uctuations, while, as far as possible, main-
taining the structure of equations of the correlation matrix similar to that of
746 Cambon
0.6
0.6
0.85
Figure 2: Application of the SCIT model to the ow over an F2A airfoil.
The contours show the ratio of the ux dissipation (from the largest scale) to
the dissipation rate
f
/, as an indicator of the unsteadiness of some regions
(courtesy Hatem Touil, PhD Thesis).
the homogeneous case. The modes which substitute for Fourier components
may, for instance, be chosen to satisfy the boundary and incompressibility
conditions. Accordingly, strong inhomogeneity due to solid boundaries can
be accommodated by the very denition of the uctuation modes (see also
Cambon 1982, chapter V). This approach is illustrated by the recent work
of Turner (1999), who considered the problem of channel ow using suitably
chosen modes whose amplitude equations are analogous to those of Fourier
modes in the homogeneous case and which were closed via a random phase
approximation. The normal modes of the linear problem might well be good
candidates in this type of approach.
The second type of approach uses a dual physical/spectral space representa-
tion of the two-point correlations u
i
(x, t)u
j
(x
i
(x, t) = a
i
(t) exp[( x
t)] (13)
with a known analytical dispersion law for
= [[
in
(13) are almost exclusively treated in WT for deriving Kolmogorov spectra,
with the key hypothesis of constant and isotropic energy uxes across dier-
ent scales associated with a wavenumber [[ (Zakharov et al. 1992). By con-
trast, in geophysical ows, dispersion laws are anisotropic, with for instance
=
x
/
2
in the case of Rossby waves, = 2
and
. This
approach would lead to replacing U
i
by U
i
c
g
in the equation for characteris-
tic curves (4). Such applications of ray methods to turbulence transport have
received preliminary applications in stably stratied (Riley, Godeferd, private
communication) and rotating (Le Penven et al. 1993) weak turbulence. An-
other application would be to try to reconcile the geometric optics mentioned
above (Eckho and Storesletten 1978) to the geometric optics extensively used
in aeroacoustics, which deals with actual acoustic rays. Looking at nonlinear
TPC, a last example, of limited but rational application to inhomogeneous
turbulence, is the generalization of EDQNM with a related model for unidirec-
tional uxes (gradient in physical space) of triple velocity and pressure-velocity
correlations by Laporta (1995).
Ultimately, one wishes to derive engineering models using elements from
TPC. Unfortunately, a full angular dependence in Fourier space is presently too
complicated to account for in a weakly inhomogeneous transport model for the
[26] Recent developments in two-point closures 751
spectral Reynolds stress tensor components
ij
(, x, t). At most, spherically-
averaged spectra
ij
(, x, t) are considered, such as for the kinetic energy spec-
trum E(, x, t) =
ii
/2. Accordingly, one has to forget the idea of recovering
exactly the asymptotic RDT limit, and there is a need to model rapid terms
connected to wave-distortion, as in (11) and (10), as well as pressure-strain
correlations. The latter closure problem arises from spherically averaging lin-
ear algebraic terms in (10), (12), because of the angular-dependent factor in
M
ij
, which reects contributions from pressure uctuation in the RDT dy-
namics. A model for the angular dependence of
ij
, was proposed by Cambon
et al. (1981), in order to derive a consistent closure of a spherically averaged
equation (12) in homogeneous turbulence, and some interesting proposals were
made by Laporta (1995) to extend this towards some inhomogeneous cases.
For the sake of simplicity, such closure problems are treated in a similar man-
ner as in single-point closure modelling, both in the LWN approximation (Los
Alamos group) and in the SCIT models (Touil et al. 2000).
In the same way, only heuristic corrections have been proposed in the vicin-
ity of walls. Accordingly, the main advantage with respect to classic single-
point closure models is the access to radial energy transfer and the related
classic cascade, with a better prediction of the dissipation rate. For instance,
the imbalance between radial energy ux from the largest scales
f
and the
dissipation rate can be exploited, as also illustrated in gure 2.
Models from the proper orthogonal decomposition (POD) of Lumley (1967)
are good candidates for extending some two-point stochastic models in terms
of amplitudes, obtained by Galerkin projection of background dynamical equa-
tions, towards strongly inhomogeneous ows. Since a small number of modes
are sucient to represent the energetic domain, low dimension dynamical mod-
els can be derived. These procedures cannot, of course, model the smallest
scales which include high vorticity and are involved in the cascade process
near the dissipative range. There remains work to do matching low dimen-
sion, very large scale dynamic models, for instance using POD modes, with
quasi-homogeneous TPC and subgrid-scale models for LES.
Another domain of interest is Lagrangian statistics and dispersion of parti-
cles (Godeferd et al. 1997; Nicolleau and Vassilicos 2000; Kaneda and Ishida
1999), with important previous work with TPC including the Lagrangian
renormalised approximation (LRA) of Kaneda (1992). Even if the starting
point of kinematic simulation (Nicolleau and Vassilicos 1999) is similar to
that used for initialising pseudo-spectral DNS, the computation of trajectories
from frozen velocity eld yields interesting results for one- and two-particle dis-
persion. Linear wave dynamics, as in RDT, can be incorporated very easily in
any realization of the velocity eld (Godeferd et al. 1997), including random
phases and random orientation of wavevectors, yielding realistic anisotropic
dispersion in the presence of stratication, rotation, and probably other orga-
nizing eects such as mean shear. It should be noted that the RDT procedure
752 Cambon
cannot be used directly to provide Lagrangian correlations, due to the intrin-
sic nonlinearity of the uctuating trajectory equations, even if the velocity
eld is governed by a linear equation. However, it can still be used as an es-
sential ingredient to derive Lagrangian nonlocal models (see also Kaneda and
Ishida 2000), which include much more of the real dynamics than classic local
Lagrangian stochastic models.
References
Babin, A., Mahalov, A. and Nicolaenko, B. (1998). On the nonlinear baroclinic waves
and adjustment of pancake dynamics, Theor. Comput. Fluid Dynamics 11, 215
235.
Benney, D.J., Saman, P.G. (1966). Nonlinear interactions of random waves in a
dispersive medium, Proc. R. Soc. London, A. 289, 301320.
Besnard, D.C., Harlow, F.H., Rauenzahn, R.M., Zemach, C. (1996). Spectral trans-
port model for turbulence, Theor. Comput. Fluid Dyn. 8(1), 135.
Caillol, P., Zeitlin, W. (2000). Kinetic equations and stationary energy spectra of
weakly nonlinear internal gravity waves, Dyn. Atm. Oceans 32, 81112.
Cambon, C. (1982). Etude spectrale dun champ turbulent incompressible soumis `a
des eets couples de deformation et rotation imposes exterieurement. Th`ese de
Doctorat dEtat, Universite Lyon I, France.
Cambon, C. (2001). Turbulence and vortex structures in rotating and stratied ows,
Eur. J. Mech. B (uids), to appear.
Cambon C., Jacquin L., Lubrano J.-L. (1992). Towards a new Reynolds stress model
for rotating turbulent ows, Phys. Fluids A 4, 812824.
Cambon C., Jeandel D., Mathieu, J. (1981). Spectral modelling of homogeneous non-
isotropic turbulence, J. Fluid Mech. 104, 247262.
Cambon C., Mansour N.N., Godeferd F.S. (1997). Energy transfer in rotating tur-
bulence, J. Fluid Mech. 337, 303332.
Clark, T.T., Zemach, C. (1995). A spectral model applied to homogeneous turbu-
lence, Phys. Fluids 7(7), 16741694.
Craft, T.J., Launder, B.E. (1990). The decay of stably-stratied grid turbulence,
Colloquium on CFD, UMIST, Manchester.
Craya, A. (1958). Contribution ` a lanalyse de la turbulence associee `a des vitesses
moyennes, P.S.T. 345 Minist`ere de lair. France
Durbin P.A., Hunt J.C.R. (1980). On surface pressure uctuations beneath turbulent
ow round blu bodies, J. Fluid Mech. 100, 161164.
Eckho, K.S., and Storesletten, L. (1978). A note on the stability of steady inviscid
helical gas ows, J. Fluid Mech. 89, 401.
Fabre, D., Jacquin, L., Sesterhenn, J. (2001). Linear interaction of a cylindrical en-
tropy spot with a shock wave, Phys. Fluids, 13(8), 120.
Galtier, S., Nazarenko, S., Newell, A.C., Pouquet, A. (2000). A weak turbulence
theory for incompressible MHD, J. Plasma Physics 63, 447488.
[26] Recent developments in two-point closures 753
Godeferd F.S., Cambon C. (1994). Detailed investigation of energy transfers in ho-
mogeneous stratied turbulence, Phys. Fluids 6, 20842100.
Godeferd, F.S., Cambon, C., Leblanc, S. (2001). Zonal approach to centrifugal, elliptic
and hyperbolic instabilities in Suart vortices, J. Fluid Mech., to appear.
Godeferd, F.S., Lollini, L. (1999). Direct numerical simulations of turbulence with
connment and rotation, J. Fluid Mech. 393, 257308.
Godeferd F.S., Malik N.A., Cambon C., Nicolleau F. (1997). Eulerian and Lagrangian
statistics in homogeneous stratied ows, Appl. Sci. Res. 57, 319335.
Godeferd, F.S., Scott, J.F., Cambon, C. (2000). Asymptotic models for slow dynamics
and anisotropic structure of turbulence in a rotating uid, Advances in Turbulence
VIII, Proceedings of 8th European Conference on Turbulence, C. Dopazo (ed.), 981.
Goldstein M.E. (1978). Unsteady vortical and entropic distortions of potential ows
round arbitrary obstacles, J. Fluid Mech. 89, 431.
Hunt, J.C.R. (1973). A theory of turbulent ows around two-dimensional blu bod-
ies, J. Fluid Mech. 61, 625706.
Hunt, J.C.R., Carlotti, P. (2000). Statistical structure of the high Reynolds number
turbulent boundary layer, J. Fluid Mech., submitted.
Kaneda, Y. (1992). Application of a Monte-Carlo method to the Lagrangian renor-
malized approximation, Chaotic Dynamics and Transport in Fluid and Plasma, I.
Prigogine (ed.), American Institute of Physics, 4656.
Kaneda, Y. Ishida, T. (2000). Suppression of vertical diusion in strongly stratied
turbulence. J. Fluid Mech. 402, 311327.
Kovasznay, L.S.G. (1953). Turbulence in supersonic ow, J. Aeronaut. Sci. 20, 657
682.
Laporta, A. (1995). Etude spectrale et modelisation de la turbulence inhomog`ene. PhD
Thesis, Ecole Centrale de Lyon, France.
Le Penven, L., Bertoglio, J.-P., Shao, L. (1993). Dispersion of linear inertial waves in
a rotating shearless mixing layer, 9th Symp. Turb. Shear Flows, Kyoto, Japan.
Lele, S.K. (1994). Compressibility eects on turbulence, Ann. Rev. Fluid Mech. 26,
211254.
Lifschitz, A., Hameiri, E. (1991). Local stability conditions in uid dynamics, Phys.
Fluids A 3, 26442641.
Lighthill, M.J. (1978). Waves in Fluids. Cambridge University Press.
Lumley J.L. (1967). The structure of inhomogeneous turbulence ows. In Atmo-
spheric Turbulence and Radio Wave Propagation, A.M. Yaglom, V.I. Tatarsky,
(eds.) 16667. Nauka.
McEwan, A.D. (1970). Inertial oscillations in a rotating uid cylinder, J. Fluid Mech.
40, 603639.
Mowbray, D.E., Rarity, B.S.H. (1967). A theoretical and experimental investigation of
the phase conguration of internal waves of small amplitude in a density stratied
liquid, J. Fluid Mech. 28, 116.
754 Cambon
Nazarenko, S., Kevlahan, N.N., Dubrulle, B. (1999). A WKB theory for rapid dis-
tortion of inhomogeneous turbulence, J. Fluid Mech. 390, 325.
Nicolleau, F.C.G.A., Vassilicos, J.C. (2000). Turbulent diusion in stably stratied
non-decaying turbulence, J. Fluid Mech. 410, 123146.
Parpais, S., Bertoglio, J.P. (1996). A spectral closure for inhomogeneous turbulence
applied to turbulent conned ow. In Proc. 6th European Turbulence Conference.
Lausanne, Switzerland, July 1996.
Reynolds, W.C., Kassinos, S.C. (1995). One-point modeling for rapidly deformed
homogeneous turbulence, Proc. Roy. Soc. Lond. A 451, 87.
Ribner, H.S. (1954). Convection of a pattern of vorticity through a shock wave,
NACA Report 1164.
Sanderson, R.C., Hill, J.C., Herring, J.R. (1986). Transient behavior of a stably
stratied homogeneous turbulent ow. In Advances in Turbulence, G. Comte-Bellot
and J. Mathieu (eds.), Springer-Verlag, 184190.
Simone A., Coleman G.N., Cambon C. (1997). The eect of compressibility on tur-
bulent shear ow: a RDT and DNS study, J. Fluid Mech. 330, 307338
Teiss`edre, C. (1984). Application des harmoniques spheriques ` a letude des proprietes
anisotropes de la turbulence homog`ene. PhD Thesis, Ecole Centrale de Lyon, France.
Touil, H., Bertoglio, J.-P., Parpais, S. (2000). A spectral closure applied to anisotropic
inhomogeneous turbulence, Advances in Turbulence VIII, Proceedings of 8th Eu-
ropean Conference on Turbulence, C. Dopazo (ed.), 689.
Townsend A.A. (1956). The Structure of Turbulent Shear Flow. Revised edition 1976.
Cambridge University Press.
Turner, L. (1999). Macroscopic structures of inhomogeneous, NavierStokes turbu-
lence, Phys. Fluids 11, 23672380.
Zakharov, V.E., Lvov, V.S., Falkowich, G. (1992). Kolmogoro Spectra of Turbulence
I. Wave turbulence. Springer-Verlag.