North-Holland, Amsterdam
James M. HYMAN
Center for Nonlinear Studies, Theoretical Division, MS B284, Los Alamos National Laboratory, Los Alamos, NM 87545, USA
An overview is given of special numerical methods for tracking discontinuous fronts and interfaces. These methods include
surface tracking methods based on connected marker points along the interface, volume tracking methods that track the
volume occupied by the solution regions bounded by the interfaces, and moving-mesh methods where the underlying mesh is
aligned and moved with the interface. The pros and cons of the current methods are discussed and a new method is proposed
that overcomes some of the difficulties encountered in approximating equations with multiple interacting interfaces.
adaptive methods automatically adjust the mesh so location. The regions are identified by marker
that it is dense in regions with sharp transitions points, or alternatively, the fractional volume of
and sparse where the solution is smooth. The each solution region located in each computational
artificially smeared discontinuities are likely to cell is calculated and advanced during the compu-
have the large gradients causing the adaptive mesh tation by solving an auxiliary evolutionary PDE.
algorithm to introduce mesh points that resolve These fractional volumes can be used to recon-
the structure of the transition layer. This structure struct an approximate interface location at any
is often an artifact of the artificial dissipation and time. Unlike the surface tracking methods, very
does not significantly affect the behavior of the little subgrid scale structure is retained during the
solution. (If it did, then the artificial dissipation calculation.
method would not have been appropriate.) These Moving-mesh methods can be used to track the
extra mesh points can greatly increase the compu- location, account for changes in the interface
tational expense without significantly enhancing topology, and resolve small-scale structures in the
the accuracy of the calculation. interfaces. Here a multivalued solution is defined
The interface tracking methods described in this at mesh points located on the interface. The inter-
paper were developed to overcome the deficiencies face mesh points move with the interface in a
of the artificial dissipation approach. The tracking Lagrangian manner. To prevent the mesh from
methods have little or no artificial dissipation near tangling, a dynamic data structure is used so the
the interface since the singularity is directly com- moving mesh points can change their nearest
puted and treated explicitly as a discontinuity. neighbors during the calculation. Also, new points
These methods are more difficult to implement, are added when the mesh becomes sparse or a new
perform excellently for a large class of problems, interface appears, and mesh points are removed
but, like artificial dissipation methods, are not well when they are more dense than necessary for an
understood theoretically for the more difficult accurate calculation.
problems. In all the above methods, the location of the
Interface tracking methods can be divided into interface is advanced by solving a lower dimen-
three catagories. In order of increasing flexibility sional PDE derived from an appropriate constitu-
and computational complexity, these are surface tive jump condition. The effect of the interface
tracking methods, volume tracking methods, and movement is transferred to the solution to the
moving mesh methods. original PDEs on either side where the interface is
Surface tracking methods track the location of treated as a moving boundary.
the interface by interpolating between marker par-
ticles along the interface. Because this is a lower 2. Surface tracking methods
dimensional problem, the additional effort to accu-
rately resolve small-subgrid-scale structure in the In surface tracking methods the interface is
interface is usually small compared with the over- specified by an ordered set of marker points located
all solution time. The surface tracking methods are on the interface [ll, 12, 22, 251. Between these
the simplest to implement, until interactions occur points its position is approximated by an inter-
that change the topology of the interface during polant, usually a piecewise polynomial. These
the computation. time-dependent interfaces divide the problem do-
Volume tracking methods overcome the chang- main into connected regions. The solution defined
ing topology problems by dividing the domain at the marker points and along the interpolated
into a union of disjoint solution regions. The interface may be multivalued to account for dis-
boundary between these regions is the interface continuities.
398 J.M. Hyman/ Numerical methodr for tracking interfaces
2.1. Surface representation smooth solution algorithm. Both the smooth solu-
tion and the interface are treated as separate com-
The marker points may be represented by the
putational objects. The interface position is stored
distance from some reference surface such as the
and dynamically updated along with the smooth
chain-of-line segments defined by a height func-
solution away from the interface. The numerical
tion (fig. la) or by a parametric interpolant as
method may be implemented to do this simulta-
shown (fig. lb).
neously, or operator splitting may be used, first
The distance function is simpler to implement,
advancing one and then the other.
but the interface deformation is severely limited
because this representation breaks down if the
2.2. Surface evolution
curve becomes multivalued with respect to the
reference surface. The parameter representation The evolution equations for the interface are
does not have these limitations and is only slightly lower dimensional differential equations derived
more complicated to implement. from constitutive relationships, usually obtained
Both of these methods can provide the fine by applying the divergence theorem to an integral
resolution and detail needed to track small-sub- formulation of the physical model. Similar meth-
grid-scale structures in the interface in two and ods are used to define an accurate approximation
three space dimensions. This can be especially to the smooth solution near an interface. This
important when tracking an unstable interface. approach, sometimes called the finite volume
The underlying computational grid resolution is method, results in equations for the interface-not
usually chosen to resolve the structure of the for the marker points. In practice, however, it is
smooth solution away from the interface and is the marker points that are evolved. Because of
rarely sufficient to resolve the fine-scale interface this, extra care is needed to maintain the ap-
detail, such as the onset of a slip-line Helmholtz propriate relationships, such as the conservation
instability rollup. laws, near the interface.
The surface tracking methods are hybrid When the underlying PDE system is hyperbolic,
numerical schemes - splitting the solution process the equations for the interface can be derived by
into two parts: the interface tracking and the solving one-dimensional Riemann problems nor-
a b
Fig. 1. a) Multiple distance functions designate the locations of the interfaces. b) A parametric interpolant designates the locations of
the interfaces.
J. M. Hyman / Numerical methods for tracking interfaces 399
ma1 to the interface [ll-131. A Riemann problem The above procedure is simplified if a local
is a type of a nonlinear normal mode expansion curvilinear coordinate system orthogonal to the
for a one-dimensional Cauchy problem with initial interface is employed. The orthogonal coordinate
data that is constant everywhere except for a single system also simplifies the interface boundary con-
jump discontinuity. The solution evolves into non- ditions that connect the smooth solution to the
linear waves that propagate coherently in time. interface. Additionally, when approximating the
For example, a material interface discontinuity spatial derivatives, we can smoothly map this coor-
moves with the underlying fluid velocity, and a dinate system onto a fixed regular computational
shock wave moves with the speed given by the grid [19, 241.
Rankine-Hugoniot jump conditions. When used Sometimes the evolution of the interface is sen-
to advance a 2-D interface in the normal direction, sitive to small amounts of noise in the computed
curvature effects can be incorporated as source solution on either side, and it may be necessary to
terms. regularize or smooth the solution states in the
When multiple waves are emitted in the tangential direction. Glimm, Marchesin, and
Riemann problem, then a single wave, such as the McBryan [ll, 121 average in a circle about each
contact discontinuity corresponding to a material marker point on the interface to filter out short
interface, must be selected and tracked at all the wavelength fluctuations. Another way to reduce
points along the same interface. Alternatively, ad- the effects of small errors on the motion of the
ditional interfaces can be inserted to track the new interface is to modify the interface evolution equa-
discontinuities that arise. tions slightly by adding artificial surface tension
The constant states used for the Riemann prob- along the interface.
lem are the bounding states on either side of the Instead of solving Riemann problems, the vortex
interface. A slightly improved version could be methods [3, 261 introduce small lines on surfaces
implemented by using a linear or quadratic varia- of vorticity along the interface. The vortex motion
tion in the states normal to the front, extending is defined by a Hamiltonian system of ordinary
the ideas of van Leer [28]. The velocities for the differential equations with a Coulomb-type inter-
marker points are taken to be the tracked wave action term. For viscous PDEs, these equations
velocities normal to the interface. The tangential also contain a diffusive term.
velocities of the underlying solution also contrib- The contour dynamics method of Overman and
ute to the interface movement, but are less Zabusky [25] follows the motion of point vortices
important, since when the marker points are dis- along a closed contour line bounding a region of
placed tangentially, they remain on the interface. constant density or vorticity. The contour velocity
In addition to the Riemann problem equations, is obtained from boundary integral equations de-
the evolution equations may contain additional rived by applying Greens theorem to the area
terms to approximate the effects of surface tension integral in the Greens function solution of a Pois-
[15, 211, flame propagation [l, 3, 261 or phase son equation.
changes. These all influence the boundary condi-
tions imposed on the smooth solution at the inter-
face. For example, if the interface were a flame
2.3. Interpolants
front that converts the fuel ahead of the interface
to burned material behind it, the interface The accuracy of the surface tracking methods
boundary conditions would be conservative out- depends strongly on stability and accuracy of the
flow or inflow conditions, respectively, with the interpolation method approximating the interface
appropriate heat source term to account for the location between the marker points. The shape-
energy released at the interface. preserving Hermite piecewise-polynomial inter-
400 J. M. Hyman / Numerical methods for tracking interfaces
polants [16] remain consistently well behaved and interaction points where three or more interfaces
smooth by retaining the convexity and monotonic- meet are common occurrences and also require
ity properties of the original data. Therefore, ex- special data structures.
traneous bumps or wiggles are not introduced Fortunately, the interactions are usually a lower
between the data points. This is particularly im- dimensional event; in 2-D calculations the 1-D
portant for unstable interfaces such as a interfaces intersect at points; in 3-D, the 2-D
Rayleigh-Taylor or Buckely-Leverett interface. surfaces interact along 1-D lines. Therefore, it
The piecewise-linear interpolant is the simplest takes little computer time to identify and track the
shape- and area-preserving interpolant, but in- interactions. Unfortunately, the data structure and
troduces fictitious comers in the interface location. algorithms needed by surface tracking methods to
In fluid flows, unless the solution is artificially account for interactions greatly increase the pro-
smoothed, these corners can create unrealistic gram complexity. Also, near the interaction many
velocities in the smooth solution that eventually of the lower dimensional interface evolution equa-
destroy the global accuracy of the calculation. The tions based on 1-D Riemann problems are no
higher order Hermite piecewise-polynomial inter- longer valid. That is, near an interaction point, the
polants are smoother but can still easily accom- solution often is intrinsically 2-D or 3-D and the
modate necessary corners in the interface by interface motion cannot be well approximated by
retaining left and right derivatives at the marker a 1-D Riemann approximation. The reason for
points. Usually, cubic piecewise polynomials are this is that near the interface the components of
sufficient. the solution normal to the interface interact
As the interface grows or shrinks, the distribu- strongly with the interface through the interface
tion of the interface marker points must change to boundary conditions, whereas the tangential com-
continually resolve the interface. The static rezone ponents have a weak, if any, interaction with the
methods [18] use a mesh function or performance interface. Therefore, away from interactions,
index based on the arc length and curvature of the applying the 1-D jump conditions normal to the
interface as a guide when removing extraneous interface is a valid approximation. Near an inter-
mesh points or adding new ones when the existing action, however, the solution cannot be split into
points are too dense or sparse. Alternatively, a components that are normal and tangential to
fixed number of marker points can be continu- both interfaces simultaneously and the splitting
ously redistributed so as to best resolve the inter- algorithm is no longer valid.
face. The first approach is usually more efficient in Simple material interfaces which move with the
2-D calculations where the marker points are easily underlying fluid velocity are not a problem, but
reordered as points are added or deleted. The data for more complicated interactions, such as an ob-
structure is simpler in the second approach and it lique shock reflection, then a local grid refinement
is more commonly used in 3-D calculations. may be necessary to reduce the splitting errors.
These hybrid methods, combining and interface
tracking algorithm with local grid refinement, are
also excellent when there is a boundary layer in
2.4. Interactions
the solution adjacent to the interface. This might
In many calculations the topology of the inter- be caused by the energy generated in a flame front
faces is constantly changing. The interfaces can or governed by an internal mixing length. If each
interact with each other, spontaneously disappear, region is being separately solved on an interface-
or new ones can be created when an existing fitted curvilinear coordinate system, then standard
interface bifurcates or is formed by a compression adaptive methods can be used to resolve the
wave, a flame ignition, or a phase change. Multiple boundary layer [2, 241.
J. M. Hyman / Numerical methods for tracking interfaces 401
Because of the complexity of handling interac- be poorly defined and sensitive to small errors.
tions and adaptively refining the interface in 3-D, This often happens in expansion regions unless
to my knowledge, the only surface tracking codes new particles are continually added and deleted
that resolve multiple interactions are in l- and during the calculation.
2-D. In most 2-D codes with interface interactions, To improve the efficiency of the MAC method,
and in all 3-D codes I know of, the interfaces are we can scatter initial marker particles more densely
tracked by a volume tracking method. (or only) near the interfaces or, if there are only
two materials, only use marker particles to identify
one of the materials.
3. Volume tracking methods The MAC methods do not require special logic
for colliding surfaces but need many marker par-
The volume tracking methods are less capable ticles per computational cell to get a well-defined
than the surface tracking methods in providing interface. Also, numerical errors in transporting
subgrid-scale resolution, but they can simply and the marker particles can cause an artificial numeri-
accurately account for the interactions of many cal diffusive mixing near the interface resulting in
different smoothly varying interfaces. Here, the a fuzzy interface. The fuzzy interface makes it
interface tracking equations have the same dimen- harder to generalize the MAC methods for com-
sion as the underlying PDEs and, therefore, are plicated interfaces such as detonations, flame fronts
potentially more expensive than the surface track- or multiphase flow problems. To reduce the fuzzi-
ing methods. In practice, however, the interface ness far more marker particles than computational
data need only be stored and the equations solved cells are needed, increasing the computational cost.
in the cells along or near the interface. This tactic The fractional marker volume methods were de-
increases the computational complexity but im- veloped to overcome these drawbacks.
proves the efficiency when the interfaces are well
separated.
3.2. Fractional marker volumes
a b C
Fig. 2. The original interface separating two regions and the associated volume fractions in each computational cell are shown (Za). In
figs. 2b and 2c are two possible reconstructed interfaces using the rectangular and piecewise-linear fractional volume methods,
respectively.
I I I
a b C
Fig. 3. The reconstructed interfaces for an unsplit algorithm (a), in the x-sweep (b), and in the y-sweep (c).
tional step algorithms are more attractive than the The PDEs are then on the new grid
surface tracking methods for almost all interacting treating points on the interface
surfaces in 2-D and, more so, in 3-D calculations. moving boundary;
The unsplit fractional volume method is more dif-
ficult to implement because of the odd shaped values for the possibly multi-
regions that must be accounted for, such as the valued solution easily
one shown in fig. 3a. selected.
In a fractional step method, the interface recon- points near the
structed in the x-sweep may be different from interface easily derived on each time step [19].
what it is in the y-sweep, (figs. 3b,c). By alter- If the fmite is being used, the
nating the sweep direction, averaging the results, solution across the interface
or using conservative limiters that preserve the
symmetry, much of the sweep dependence can be global solution algorithms
reduced. are available
Fig. 4. A simple local adjustment of the mesh aligns it with the front, so the interface is located only along cell edges or diagonals. a)
The original underlying mesh and interface; b) the locally adjusted mesh is aligned with the interface.
Depending upon the goal of the calculation, any are required to move with the interface. The grid
one of several moving-mesh methods can be used points away from the interface are distributed to
away from the interface [6-10, 13, 171. Unfor- prevent mesh tangling or to better resolve the
tunately, unless the moving cells are allowed to structure of the smooth solution. In addition, the
change their nearest neighbors in a calculation, mesh points along the interface can be redistrib-
then continual rezoning is necessary to prevent the uted along the interface using the surface tracking
mesh distortion, in even the simplest flows, from rezone methods described earlier.
destroying the accuracy of the calculation. This is The mesh tangling is not caused by the
shown for the very early stages of a Lagrangian Lagrangian equations but is an artifact of the
Rayleigh-Taylor calculation on a logically rectan- mesh data structure. A more flexible data structure
gular grid in fig. 5. Soon after the time of this plot, that prevents mesh tangling is the neighborhood
numerical errors caused the grid lines to cross and grid, where pointers are kept to all the nearest
the calculation to fail, neighbors of each mesh point. These are used to
The grid distortion can be somewhat alleviated approximate the equations locally using a finite
if only the mesh points along specified interfaces element or finite volume discrete approximation.
As the mesh moves, the nearest neighbor pointers
are continually updated. Usually these reconnec-
tions alone will not completely solve the problem
of grid resolution and regularity, and some form
of rezoning is till necessary. But here, adding and
deleting points is much easier than on a logically
rectangular or cuboid mesh.
There has been considerable success in using the
reconnecting moving neighborhood mesh in fluid
flows with material interfaces [7-9, 271. These free
Lagrangian codes, as they are called, could easily
be adapted for more complicated flows where.the
Fig. 5. Distorted Lagrangian mesh in the early stages of a interface moves with the shock or flame velocity
Rayleigh-Taylor instability calculation. rather than with just the underlying fluid velocity.
J. M. Hyman / Numerical methodr for tracking interfaces 405
II x ! x x X
x x X X
*
x I x X X
a
GL x x
iif
X
b
X
X
C
Fig. 6. The dual computing mesh is advanced according to the appropriate moving mesh equations. At the end of each time step it is
regularized as in 6c. a) The reference mesh and the dual computing mesh; b) the new dual mesh after one time step; c) the regularized
dual mesh at the new time.
The major disadvantage of neighborhood meshes appropriate moving-mesh equations for the inter-
is the data structure. Most of the fast and accurate face depicted by a solid line in fig. 6. If desired, an
numerical methods (and multidimensional plotting underlying reference mesh can now be chosen to
packages) are for logically rectangular and cuboid resolve the solution. The mesh is then regularized
grids. The accuracy of local discrete approxima- (fig. 6c) by adding new dual mesh points to empty
tions of second derivative operators is particularly cells or combining them in reference cells with
poor on neighborhood grids. more than one dual mesh point. In this way, the
One of the more promising reconnecting mesh mesh points where the solution is computed can
algorithms retains a logically rectangular or cuboid continue changing their nearest neighbors while
data structure by approximating the solution on maintaining a logically rectangular data structure.
the dual mesh of a logically rectangular or cuboid In addition, the dual mesh computing points
reference mesh [13, 171. The dual mesh consists of can be tagged as special interface points (fig. 7a).
one mesh point within each cell of the reference In fig. 7b the dual mesh points, separating materi-
mesh (fig. 6a). als q, x and 0, have been advanced forming
At the beginning of every time step, the mesh is some new mixed cells in the center. In the regulari-
regular (one grid point per reference cell). The zation stage (fig. 7c) these are combined conserva-
dual mesh is advanced (fig. 6b) according to the tively to form two new triple points.
0 i Y 0 X
d x I
42E$il#
ID
.Elia
0 ,D-- -- 0
/ I/ x /
-_ _g--_ I
---D
._d x x x
@-a. x -p;- \\ x
x x
x -@- + *- \ 0 -@--
1 --@-
x
I
0 .O
x 4 O x i O
a b C
Fig. 7. Interface tracking on the dual mesh. a) Original dual mesh marker points and interface; b) predicted new dual mesh points; c)
regularized dual mesh and interface.
406 J.M. Hyman/ Numerical method for tracking interfaces
5. Summary and conclusions Problems with Interfaces, Elliptic Problem Solvers, Martin
Schultz, ed. (Academic Press, New York, 1981), pp.
247-254.
Most 161J.K. Dukowicz, A Simplified Adaptive Mesh Technique
Derived from the Moving Finite Element Method, Los
Alamos National Laboratory report LA-UR-82-3664.
[71 J.K. Dukowicz, Lagrangian Fluid Dynamics using the
Voronoi-Delatmay Mesh, Los Alamos Scientific Labora-
tory report LA-UR-81-1421.
181 M.J. Fritts, Lagrangian Simulations of the Kelvin-Helm-
holtz Instability, Science Applications Inc. report SAI-76-
632-WA, September 1976.
[91 M.J. Fritts and J.P. Boris, The Lagrangian Solution of
Transient Problems in Hydrodynamics using a Triangular
Mesh, J. Comp. Phys. 31 (1979) 173-215.
WI R.J. Gelinas and S.K. Doss, The Moving Finite Element
Method: Applications to General Partial Differential
Equations with Multiple Large Gradients, J. Comp. Phys.
40 (1981).
WI J. Glimm, E. Isaacson, D. Marchesin and 0. McBryan,
Front Tracking for Hyperbolic Systems, Adv. Appl. Math.
2 (1981) 91-119.
WI J. Glimm, D. Marchesin and 0. McBryan, A Numerical
Method for Two Phase Flow with an Unstable Interface,
J. Comp. Phys. 39 (1981) 179-200.
1131 A. Harten and J.M. Hyman, A Self-Adjusting Grid for the
Computation of Weak Solutions of Hyperbolic Consetva-
tion Laws, J. Comp. Phys. 50 (1983) 235-269.
[I41 C.W. Hirt and B.D. Nichols, A Computational Method
for Free Surface Hydrodynamics, American Society of
Mechanical Engineers 80-C2/PVP-144, April 1980.
[I51 C.W. Hirt and B.D. Nichols, Volume of Fluid (VOF)
Method for the Dynamics of Free Boundaries, J. Comp.
Phys. 39 (1981) 201-225.
Ml J.M. Hyman, Accurate Monotonicity Preserving Cubic
Acknowledgements Interpolation, Los Alamos Report No. LA-8796-MS, to be
published in SIAM J. Sci. and Stat. Comp.
I thank Andy White for pointing out the distinc- P71 J.M. Hyman, Adaptive Moving Mesh Methods for Partial
Differential Equations, Advances in Reactor Computa-
tion between a numerical method and a numerical
tions (American Nucl. Sot., La Grange Park, IL, 1983) pp.
scheme. 24-43.
W31 J.M. Hyman, Adaptive Static Rezoning Methods, to be
published in the proceedings AMS-SIAM Conference on
References Large Scale computations in Fluid Mechanics (1983).
[I91 J.M. Hyman and B. Larrouturou, The Numerical Differ-
111P.K. Barr and Wm.T. Ashurst, An Interface Scheme for entiation of Discrete Functions Using Polynomial Interpo-
Turbulent Flame Propagation, Sandia National Laborato- lation Methods, Appl. Math. and Comp., ~01s. lo-11
ries report SAND828773 (1982). (1982) pp. 487-506.
121J.U. Brackbill and J.S. Saltzman, Adaptive Zoning for PO1 P. Lotstedt, A Front Tracking Method Applied to Burgers
Singular Problems in Two Dimensions, J. Comp. Phys. 46 Equation and Two-Phase Porous Flow, J. Comp. Phys. 47
(1982) 342-368. (1982) 211-228.
[31 A.J. Chorin, Flame Advection and Propagation Algo- WI B.D. Nichols, C.W. Hirt and R.S. Hot&kiss, SOLA-VOF:
rithms, J. Comp. Phys. 35 (1980) l-11. A Solution Algorithm for Transient Fluid Flow with Mul-
[41 P. Colella, P. Concus and J.A. Sethian, Some Numerical tiple Free Boundaries, Los Alamos Scientific Laboratory
Methods for Discontinuous Flows in Porous Media, Report LA-8355, August 1980.
Lawrence Berkeley Laboratory report LBL-15932, April WI B.D. Nichols and C.W. Hirt, Methods for Calculating
1983. Multi-Dimensional, Transient Free Surface Flows Past
[51 J.E. Dendy and J.M. Hyman, Multigrid and ICCG for Bodies, Los Alamos Scientific Laboratory report LA-UR-
J.M. Hyman/ Numerical methods for tracking interfaces 407
75-1932; Proc. First Int. Conf. on Numerical Ship Hydro- Vessels, Lawrence Berkeley National Laboratory report
dynamics, Gaithersburg, MD, October 1975. LBL-15744, Feb. 1983.
[23] W.F. Noh and P. Woodward, SLIC (Simple Line Interface (271 H.E. Trease, A Two-Dimensional Free Lagrangian Hydro-
Calculation), in Lecture Notes in Physics, vol. 59: Proc. dynamics Model, Ph. D. Thesis, Sept. 1981.
Fifth Int. Conf. on Numerical Methods in Fluid Dy- [28] Bram van Leer, Towards the Ultimate Conservative Dif-
namics, A.I. van de Vooren and P.J. Zandbergen, eds. ference Scheme. V. A Second-Order Sequel to Godunovs
(Springer-Berlin, 1976). Method, J. Comp. Phys. 32 (1979) 101-136.
[24] J.F. Thompson, ed., Numerical Grid Generation (North- [29] J.E. Welch, F.W. Harlow, J.P. Shannon and B.J. Daly, The
Holland, Amsterdam, 1982). MAC Method: A Computing Technique for Solving
[25] E.A. Overman II and N.J. Zabusky, Evolution and Merger Viscous, Incompressible, Transient Fluid Flow Problems
of Isolated Vortex Structures, Inst. for Computational Involving Free Surfaces, Los Alamos Scientific Laboratory
Mathematics and Applications Report ICMA-81-31(1981). report LA-3425, 1966.
[26] J. Sethian, Turbulent Combustion in Open and Closed