nd
International Conference on Engineering Optimization
September 6-9, 2010, Lisbon, Portugal
Design of an Airfoil from a Given Pressure Distribution Using an
Approximate Inverse Operator
Jan
Simak, Jaroslav Pelant
Aeronautical Research and Test Institute, Beranov ych 130, 199 05 Prague - Let nany, Czech Republic
simak@vzlu.cz, pelant@vzlu.cz
Abstract
This contribution describes a numerical solution of an inverse problem for a ow around an airfoil. Its
aim is to design an airfoil shape based on a given pressure distribution on its surface. The pressure
distribution aects forces acting on the airfoil, so it is possible through its modication to obtain shapes
which satisfy requested parameters. The method is suitable to solve problems in subsonic regimes. The
main idea of the method is a combination of a direct and an approximate inverse operator. Using both
operators, the resulting shape is corrected in each iteration until the desired precision is achieved. The
direct operator means, in this notation, a solution of a 2D ow around the airfoil. There is some degree
of freedom when choosing the model. In this case, the Navier-Stokes equations completed with a k-omega
turbulence model are used. A stationary solution of this system is obtained by an implicit FVM. The
approximate inverse operator is based on the thin-airfoil theory for a potential ow. The airfoil is de-
scribed by a mean camber line and a thickness function, which ensures the capability to deal not only
with thin airfoils. Numerical examples are presented.
Keywords: inverse problem, CFD, turbulence, airfoil
1. Introduction
This paper concerns a numerical method for a solution of an airfoil design inverse problem. Using this
method, it is possible to get an airfoil shape that satises a prescribed pressure distribution on its surface
under given ow conditions. It can be used as a part of some optimization process.
The literature describes various approaches to the solution of this problem. The geometrical inverse
design methods use usually some auxiliary equations, based on the actual pressure distribution, to correct
the shape. The main idea of the presented method is to construct an inexact inversion, a mapping
between a pressure distribution and an airfoil shape. The inversion in its original form is almost exact
for a potential incompressible ow. Hence this mapping is used in an iterative process until the desired
airfoil is obtained. It is possible to couple this inexact inversion with an arbitrary ow solver.
The method is aimed to a subsonic ow, the angle of attack is one of the results of the method. The
method presented here is an extension of methods described in previous works [3], [6], [7]. The original
method deals with an inviscid ow, later with a laminar viscous ow. The current method deals with
a turbulent viscous ow described by the Navier-Stokes equations equipped with the k turbulence
model which improves its applicability. In the following text a short description of the method is given.
2. Inverse problem
As was said above, the goal of the method is to nd an airfoil shape corresponding to a prescribed
pressure distribution. The method utilizes the possibility of derivation of an approximate inversion.
Thus the problem could be written as:
Find a pressure pseudo-distribution p such that
PL(p) = f , (1)
where f is the given pressure distribution, P is an operator representing a solution of a ow problem and
nally L represents an approximate inversion to P.
This problem is solved by the method of successive iterations, where the solution of the Eq.(1) is a
limit of the sequence
{p
k
}
k=0
, p
k+1
= p
k
+(f PLp
k
) . (2)
1
The parameter is a positive real number chosen such that the sequence converges. According to
experiences from numerical results, the choice = 0.6 is sucient to achieve convergence in most cases.
Lower values ensures better convergence but also increases the number of iterations.
The approximate inverse operator L is derived using the thin airfoil theory. The details of this
can be found in [3]. In short, the airfoil is composed of a mean camber line and a thickness function,
mathematically written
1
(x) = x t(x)
s
(x)
_
1 +s
2
(x)
,
2
(x) = s(x) t(x)
1
_
1 +s
2
(x)
, x 0, 1 . (3)
In this notation the airfoil coordinates are denoted by (
1
,
2
), the upper sign is for the upper part of
the airfoil and the bottom sign for the bottom part.
The functions s(x) (=the mean camber line) and t(x) (=the thickness function) are derived on the
idea of articial distributions of vorticity and sources q along the chord. The velocities induced by this
distributions are described by
u
y
(x) =
1
2
_
1
0
()
x
d, (4)
u
x
(x) =
1
2
_
1
0
q()
x
d. (5)
Under assumption that the velocity is tangent to the airfoil, the following result can be obtained
s(x) =
x
2
_
1
0
(u
up
() u
lo
()) ln
1
2
_
1
0
(u
up
() u
lo
()) ln
d , (6)
t(x) =
1
_
1
0
_
u
up
() +u
lo
()
2
1
_
ln
1 +
_
( x)/(x x)
1
_
( x)/(x x)
d . (7)
The variables x and run along the normalized chord line, given by the interval 0, 1. The symbols
u
up
and u
lo
represent velocity distributions on the airfoil surface, normalized by the free stream velocity.
These functions need not represent a physically relevant distribution, they are related to the sequence
(2) instead. Hence, similar to the notation used in Eq.(1), they can be called pseudo-distributions. The
transformation between the pressure pseudo-distribution and the velocity pseudo-distribution, in the
formula denoted as u(x), is the following:
_
u(x)
u
_
2
=
2/M
2
+ 1
1
_
1
_
p(x)
p
0
_
(1)/
_
. (8)
The symbol p
0
is the pressure at zero velocity, M
is the free
stream velocity and is the Poisson adiabatic constant.
From the construction of the airfoil coordinates, it is clear that u
up
and u
lo
need to be functions.
Since the chord line doesnt need to connect the leftmost point with the rightmost one, simply taking
the x-coordinate of a point on the surface doesnt satisfy this requirement, in general. From that reason
the distribution is assumed along the mean camber line, with the x-coordinate as the leading variable x.
The mean camber line is evaluated in each iteration, so the only additional expense is the inversion of
the mapping Eq.(3).
The integrals in Eqs.(6) and (7) are evaluated using a suitable numerical quadrature. Care must be
taken, since the integrands have singularities for = x, = 0 and = 1. From this reason a Chebyschev-
Gauss quadrature was chosen, where the interval is divided by the roots of a Chebyschev polynomial.
The integrating formula is
_
1
0
f(x
i
, ) d
2
N
N/2
k=1
f(x
i
, x
2k1
)
_
x
2k1
(1 x
2k1
), i = 0, 2, 4 . . . , N (9)
2
with x
i
= 1/2 (1 + cos (i/N)) , i = 0, 1, . . . , N.
3. Flow problem
The viscous compressible ow around an airfoil is described by the system of the Navier-Stokes equations.
Since most of the ow in real situations is turbulent, the laminar model seems insucient. To improve
the quality of the predicted ow and also the stability of the method, a k model of turbulence is
included (see [2], [8]).
3.1. Model of the ow
The system of the equations can be written in a vector form
w
t
+
2
j=1
F
j
( w)
x
j
=
2
j=1
R
j
( w, w)
x
j
+
S ( w, w) , (10)
where
w = (, v
1
, v
2
, E, k, )
T
, (11)
F
j
( w) = (v
j
, v
1
v
j
+
1j
p, v
2
v
j
+
2j
p, (E +p)v
j
, kv
j
, v
j
)
T
, (12)
R
j
( w, w) =
_
0,
j1
,
j2
,
j1
v
1
+
j2
v
2
+
_
P
r
+
T
P
rT
_
e
x
j
, ( +
k
T
)
k
x
j
,
( +
T
)
x
j
_
T
, (13)
S( w, w) =
_
0, 0, 0, 0, P
k
k, P
2
+ C
D
_
T
. (14)
Using the common notation, denotes a density, p is a pressure, v
1
, v
2
are velocity components, E is an
energy, k is a turbulent kinetic energy and nally is a specic turbulent dissipation. The symbol P
r
represents the Prandtl number (the subscript T denotes the turbulence). The viscosity coecient is
evaluated using the Sutherlands formula. The symbol
T
denotes an eddy viscosity coecient, which is
given by the formula
T
=
k
. (15)
The stress tensor in the N.-S. equations is given by relations
11
= ( +
T
)
_
4
3
v
1
x
1
2
3
v
2
x
2
_
2k
3
,
22
= ( +
T
)
_
2
3
v
1
x
1
+
4
3
v
2
x
2
_
2k
3
, (16)
12
=
21
= ( +
T
)
_
v
1
x
2
+
v
2
x
1
_
.
The production of the turbulence P
k
and the production of the dissipation P
are expressed as
P
k
=
11
v
1
x
1
+
12
_
v
1
x
2
+
v
2
x
1
_
+
22
v
2
x
2
, (17)
P
P
k
k
, (18)
where
ij
=
ij
for = 0. Finally, the cross-diusion term C
D
is given by the relation
C
D
=
D
max
_
k
x
1
x
1
+
k
x
2
x
2
, 0
_
. (19)
The turbulence model is closed by parameters
= 0.09, = 5
/6,
= /
2
/
(where
= 0.41 is the von Karman constant),
k
= 2/3,
= 0.5 and
D
= 0.5. This choice of parameters
resolves the dependence of the k model on the free stream values [2].
If the turbulent kinetic energy k is set to zero, the turbulence model has no inuence upon the N.-S.
equations and the laminar model is described.
3
3.2. Numerical treatment
The system of equations mentioned above is solved by the implicit nite volume method. The variables
are normalized using critical values of the density, velocity and pressure. The resulting dimensionless
system has the same form as the original one and thus no modication to the system is needed. The
computational domain is discretized by a structured quadrilateral C-type mesh.
Since the coupling between the equations describing the ow and the equations describing the turbu-
lence is only by the viscous terms, it is possible to solve the problem in two parts [8]. In the rst part
(continuity equation, momentum equations, energy equation), the variables k and are assumed time
independent. Similarly, in the second part (k equations), the variables p, , v
1
, v
2
are held constant
in time and the system of two equations is solved with respect to the unknowns k and . These systems
can be solved independently of each other. This approach reduces computational costs and allows to
easily modify a laminar solver into a turbulent one.
The linearized system of algebraic equations is solved by the GMRES method (using software SPARSKIT2
[4]). The convective terms
F
i
are evaluated using the Osher-Solomon numerical ux in the case of the
ow part and by the Vijayasundaram numerical ux in the turbulent part. A higher order reconstruction
based on the Van Albada limiter is also implemented. The numerical evaluation of a gradient on an edge
of two cells is based on the values in the centres of the six neighbouring cells.
Since a suitable angle of attack has to be found in order to satisfy the condition on the position of
the stagnation point on the leading edge, the airfoil is rotated round a chosen point.
4. Numerical examples
The rst example shows an asymmetric case. The method was examined on the NACA4412 airfoil. The
Reynolds and Mach numbers are Re = 6 10
6
and M
= 1.8
= 0.82