JOURNAL OF COMPUTATIONAL PHYSICS 11, 348-359 (1973)
A Simple Scheme for Generating General Curvilinear Grids
A. A. AMSDEN AND C. W. HIRT
University of California, Los Alamos Scientific Laboratory,
Los Alamos, Nen?Mexico 87544
Received March 29, 1972
An intuitively simple approach is presented for the computer generation of two-
dimensional curvilinear grids suitable for finite difference solutions of problems in the
field of continuum dynamics. An iterative process is employed to transform uniform
networks of rectangular zones into more complex configurations. Ease of use and
optimal adjustment are stressed, and numerous examples are given.
1. INTRODUCTION
Various techniques have been developed over the years for generating two-
dimensional grids that are orthogonal in certain regions or in their entirety. Many
of these methods are specifically oriented to the production of grids for the finite
difference solutions of particular problems in continuum dynamics, and a number
of investigators, such as Chu [l] have turned their attention to creating finite
difference approximations for general grids.
Indeed, with the exception of very complicated boundary geometries, meshes in
arbitrarily shaped regions are easily and rapidly constructed by the computer.
Winslow has described a technique [2, 31 for creating a grid of triangular zones
by direct numerical solution of Laplace equations. Other numerical methods
have been described by Barfield [4, 51 for generating orthogonal or nearly ortho-
gonal grids. The principal limitation with all of these, however, is the relatively
rigid constrictions each places on the final results. For example, in some applica-
tions it is often desirable to enhance the grid resolution in selected regions. The
technique described here offers a conceptually simple and logical approach to
accomplish this, allowing the creation of a wide variety of computing grids or
grid sections, through the use of elementary mathematics, letting the high-speed
computer calculate the desired coordinates of the mesh vertices through an
iterative process. Ease of use and optimal adjustment are emphasized, at the
expense of orthogonality if necessary.
348
Copyright 0 1973 by Academic Press, Inc.
All rights of reproduction in any form reserved.
CURVILINEAR GRIDS 349
2. THE METHOD
Let us begin by considering a network of rectangular zones. To simplify this
discussion, we shall define the zones to be of uniform dimensions 6x and 8~
throughout, although the concept is not limited by this restriction. The calculation
of the desired grid, conforming to a specified boundary shape, proceeds by moving
the vertices of the original grid in a sequence of small steps or iterations toward
their final positions. The entire field of vertices is repeatedly swept until all vertices
are being moved less than some small predetermined distance E, at which time
the field is said to have relaxed and the generation is complete.
In the logic scheme, each vertex is distinguished as being one of two possible
types-boundary or interior. The boundary vertices are the crucial ones to move
correctly, since the interior vertices are moved to adjust to the boundaries in a
smooth fashion. Various prescriptions may be used to move the vertices, and it
is often necessary to experiment before an optimum choice is obtained. It is, in
fact, this flexibility of choice that we wish to emphasize.
We will study a number of specific examples, of increasing complexity, to
demonstrate possible treatments of the boundary and interior vertices.
In the first example, the initial grid consists of square computing zones 20 wide
by 10 high (i.e., 21 vertices by 11 vertices), the index i counting vertices horizontally
and the index j counting vertices vertically. The zones have dimensions 6.x = 8~ =
0.1, and are shown in Fig. la. The problem is to deform this rectangular grid into
the semicircle of Fig. 1b. Such a grid might be useful for calculating the slosh of a
liquid in a round-bottomed tank, using, for example, the Arbitrary Lagrangian-
Eulerian (ALE) computing method [6], which can handle curved boundary regions.
The transition of the geometry from Fig. 1a to that of Fig. 1b can be accomplished
if we move the boundary vertices on the left, bottom, and right radially inward,
toward the center of the semicircle; that is, each boundary point at (x, JJ) is shifted
during each iteration pass to
where
p = /3&x -- x,)2 + (y ~ y,)2 - t-21.
Here x0 and y, are the coordinates of the circle center (vertex 11, 11), and r is
the desired circle radius. In this example, x0 = y0 = r = 1.0. The coefficient ,6$
is chosen to restrict the magnitude of x and 4’ changes to a reasonable value during
a given iteration to reduce the possibility of vertices crossing one another. The
vertices should move gradually and persistently to their final positions through
the course of the iteration. The value of /3,, = 0.01 was chosen to limit x and ,J’
350 AMSDEN AND HIRT
FIG. 1. (a) Initial 20 x 10 computing grid; (b) semicircle formed by iterative Process.
changes to no more than approximately 6x/10. All boundary vertices not on the
top edge (j = 11) are moved by the above prescription, whereas the vertices at
j = 11 are never moved. This leaves only the interior vertices to be considered,
and in this example they are moved by the simple prescription of placing each at
the average position of its eight closest neighbors,
where, for simplicity, the most updated values of x and y available are usedon the
right side of these equations. Thus, we now have a prescription defined for every
vertex, and it is an easy matter to construct a logic path for the computer to follow.
The iterative sweeping of all vertices is performed until no boundary or interior
vertex has an x or y change greater than E (we have chosen E = lo-“). The transi-
tion of the grid of Fig. la to that of Fig. lb required 291 iterations, a task involving
only a few seconds of computer time on the CDC 6600, and very little program-
ming.
It was found on this example that the number of iterations required for relaxa-
tion was at a minimum for /?,,values in the range 0.10-0.92. The number dropped
rapidly as ,B,,was increased from 0.01 to 0.08, where only 75 iterations were
required. For /$, values up to 0.94, the number of iterations required was almost
CURVILINEAR GRIDS 351
constant, with a minimum of 67 for /3, values of 0.90-0.92. Above ,& = 0.92,
however, the number of iterations again rose quite rapidly, with 227 required for
p,, = 0.98, and with /$, = 1.0, the solution never relaxed. This experimentation
with various values of /I, was tried only with this first example, but it is evident
that it is worthwhile to find an optimum p,, when similar grids are to be generated
repeatedly.
It should be noted that it is generally not necessary to relax the boundary vertices
to their final positions insofar as grid generation itself is concerned (see, for
example, Ref. [2]). A similar calculation was made in which the boundary vertices
were simply placed in their final semicircular configuration, and the interior points
were forced to relax to this boundary using Eqs. (3). In the early stages the grid
points were strongly crossed, but they eventually uncrossed and the convergence
was as fast as the optimum-/& run above, requiring only 68 iterations using the
same E = IO-” convergence criterion.
If desired, the possibility of crossover can be eliminated by always expanding
the grid outward to form the desired curvilinear grid, then resealing the final set
of coordinates to the actual zone size desired.
The reason we wish to emphasize the technique of relaxation of the boundary
points is that in many cases it is preferable to generate the curvilinear grid within
and during a continuum dynamics calculation. Thus, one iteration pass of the
grid relaxation is performed during each time step of the calculation, and effectively
provides a rezoning technique. (In a grid qualitatively like that of Fig. 1, this may
be useful for calculating spherically diverging shocks.) To allow even slight crossing
of mesh lines would be detrimental or catastrophic to the results if it occurs during
a calculation. Moreover, the expansion-and-resealing technique to eliminate
crossover is not appropriate if a calculation is being performed simultaneously.
In combining the grid generation as a rezoning technique within the continuum
dynamics calculation, it was seen that some overrelaxation of the interior vertices
was required to keep them moving inward ahead of the exceptionally fast motion
of the edge vertices. Even with a p,, of only 0.01, the grid will virtually reach its
final configuration in only 25 iterations, and optimum speed of grid generation
may cease to be a desired goal.
If desired, the area of the final semicircle can be made equal to the original grid
area by an appropriate choice of 6x and 8~ in the original rectangle. For example,
to obtain a final semicircle radius of 1, with the circle center again at vertex
(1 I, 1I), the initial 6x and 8~ of the rectangle are seen to be 0.0886, and the initial
coordinates of vertex (I ,I), thus, lie at (0.1138, 0.1138) rather than at (0.0, 0.0).
In the iteration the grid vertices will automatically move inward or outward as
required. The vertices along the top boundary, other than the corners, should be
evenly spaced in x between neighbors on the left and right. The resulting specified-
area grid has the same qualitative appearance as that of Fig. 1b, and required 286
352 AMSDEN AND HIRT
iterations to generate with ,$, = 0.01. The reduced number of iterations from the
earlier example, using the same /3,,, may be attributed to the fact that vertices do
not have to be moved as far.
The choice of interior vertex positioning as governed by Eqs. (3) is roughly
equivalent to making the grid locally orthogonal. More accurate formulas are
described by Winslow [2, 31. However, it is the purpose of this paper to indicate
how other choices for grid generation can be realized as well.
In the examples discussed thus far, the approximate orthogonal spacing of
interior vertices has been sufficient, but the approach will handle any arbitrary
bias or weighting toward a particular direction that one desires, giving more
flexibility in this regard than the usual orthogonal approach. For example, unequal
weighting of the four principal neighbors is often useful, as illustrated in the
examples of Fig. 2. Figure 2a shows the initial notched grid of equal rectangular
cells. The problem is to move the boundary vertices lying in the notch into a
semicircular configuration. Remaining boundary vertices not in the notch are
moved to the average position of their two boundary neighbors on either side,
while the four outside corners are held forever fixed. Figure 2b shows the resulting
grid when all eight neighbors are weighted equally for each interior vertex, accord-
ing to Eqs. (3). The interior vertices can be drawn closer to the semicircle with a
weighting (W) for each neighbor, proportional to its distance from the semicircle.
A possible form is
w,v = I + [a/(1 + 20d2)1,
where
d = [(x,v - xJ2 i- (y, - yo)2 - r2]/r2,
01is a constant, and the subscript N denotes the vertex label of a particular neighbor.
The set of resulting weights are then used to define new interior vertex positions to
replace Eqs. (3),
Yij= i1 wNYN/i
1
WN .
Figure 2c shows the grid when 01= I, while the extreme deformation of Fig. 2d
is obtained with 01= 3. The solutions required 144, 168, and 225 iterations for
Figs. 2b, 2c, and 2d, respectively, with the convergence criterion /$ = 0.01.
A grid similar to that in Fig. 2c was successfully used in several ALE-method
calculations of flows about plane cylinders and spheres. The weighting treatment
was necessary to increase grid resolution near the curved boundary for the accurate
calculation of boundary layer formation in this region.
CURVILINEAR GRIDS 353
a)
bl
d)
FIG. 2. (a) Initial grid; (b), (c), (d) are resultant semicircles with various amount of weighting
of interior vertices.
Another possible mesh for such calculations has the horizontal grid lines run
up over the cylinder, rather than intersecting it, thus more closely following stream-
lines of the flow. A grid of this type can be formed by starting with the full
rectangular grid as in Fig. 3a and “pushing” a portion of the boundary inward to
form the semicircle, where some number (K) of vertices along the bottom are
581/r I/3-4
354 AMSDEN AND HIRT
chosen to define the semicircle. In this example, the angle 0 at the circle center
(x0 , y,,) subtended by neighboring pairs of vertices along the semicircle is constant
and is given by 0 = z-/(K - 1). Both 13and the coordinates of the boundary vertices
to be pushed in are easily specified; the k-th point around the semicircle will have
final coordinates
XJc= x, - r cos[(k - 1) 01,
(4)
y, = y. + r sin[(k - 1) 01,
where r is again the semicircle radius. A vertex point (i,,j) to be pushed in to the
k-th circle point can be moved by the simple relaxation equations
b)
FIG. 3. (a) Initial grid (b) grid; pushed in to form semicircle; (c) crossove:r eliminated by
unequal weighting.
CURVILINEAR GRIDS 35.5
When interior vertices and other bounda~ vertices are moved to average positions
among adjacent eight or two neighbors, respectively, the resulting grid is shown
in Fig. 3b. A total of 342 iterations were necessary to obtain this configuration,
with &, = 0.02. However, moving interior vertices to average positions of four
neighbors instead of eight gives virtu~ly identical results, and is faster. One of the
problems of the pushing-in technique is evident in Fig. 3b-the semicircle is
properly formed, but the j = 2 line has crossed over and lies below the semicircle
formed from the j = 1 line. This results from the logic chosen for moving the
interior vertices and not from the fineness of the iteration. (A run with /I0 reduced
to 0.004 gave identical results after 1322 iterations.) To prevent crossover the
treatment of the interior vertices must be modified. The more accurate equipoten-
tial method of Winslow [2, 31 is one method that would prevent this crossover.
Alternatively, it may be noted that crossover really is manifested in the y direction,
and evidently an equal weighting of neighbors for a new JJcoordinate does not
Initial grid; (b) corner pushed in, exhibiting crossover of interior vertices.
356 AMSDEN AND HIRT
give a proper balance. The validity of this argument is shown in Fig. 3c where
crossover has been eliminated by making 80 ‘A of y-coordinate weighting dependent
on the two principal vertical neighbors, and only 20 “/I on the two lateral neigh-
bors,
Jg = ()4(y;+l $ y;-‘> + O.l(y%,.1 + y:-1).
The interior x weighting remained equal for the four neighbors. The grid of
Fig. 3c also required 342 iterations to create.
Thus, it is seen that a judicious choice of interior vertex weights is crucial in
strongly concave regions if crossover is to be eliminated in the final configuration.
Usually, unequal weighting is helpful, and the strongest “springs” should pull
normal to the concave boundary, as approximated in the above 80% and 20%
equation.
Despite its usefulness, the pushing-in technique cannot be used in all instances.
An extreme example is shown in Fig. 4. The upper right corner of the grid in
Fig. 4a was pushed in to form a notch (Fig. 4b). As in the previous examples,
final boundary point coordinates were defined a priori, with the vertical boundary
points on the right side laid over flat, and the points along the top turned down
vertically. The crossover is not incurable, but the zoning around the notch would
be highly irregular regardless of the treatment of the interior points in that region,
and consequently of little value as a finite difference grid. Generally, such a notched
shape would be better formed from an initially notched rectangular region.
The general iterative scheme is ideally suited to the formation of a grid for
computing bifurcated flows. One possible technique makes use of a grid with a
row of zones of zero height along the line of bifurcation. The initial grid shown
in Fig. 5a is actually 11 zones high rather than the 10 zones it appears to be. Here,
the y coordinates of the j = 7 grid line are initially identical to the y coordinates
of thej = 6 grid line, and the mesh is split along this central line into the configura-
tion shown in Fig. 5b.
The desired final coordinates of only the four corner vertices on the right side,
(19, 1), (19, 6), (19, 7), and (19, 12), need to be known and specified beforehand.
These four vertices are then iterated to their final positions, the two lower ones
moving downward, and the two upper ones moving upward. Other boundary
vertices that will lie at corners of the final grid (1, I), (7, l), (9, 6), (9, 7), (I, 12),
and (7, 12), and also vertices on the j = 6 and j = 7 grid lines out to i = 9, are
never moved. Remaining boundary vertices are then moved to the average posi-
tions of their appropriate two neighbors, and interior vertices to the average of
their eight neighbors. Vertical grid lines in the central zones (j = 6) stretched in
the bifurcation have been deleted from the plot. In a finite difference calculation
using this grid, the j = 6 vertices on the left eight zones must be identified with
CURVILINEAR GRIDS 357
12
I/
IO
9
a
a) 6,7
5
4
i-
b)
FIG. 5. (a) Initial 18 x 11 grid, where j = 6 row has zero height; (b) bifurcated flow grid
formed by iterative process.
the j = 7 vertices above. The grid in Fig. 5b required 360 iterations to create,
with /3” = 0.02, E = 10-4, and 6x = Sy = 0.2.
The generation of the grid in Fig. 6b from the grid of Fig. 6a followed the same
general concepts as the previous example. Here, however, we specified the final
coordinates of all 16 boundary vertices forming each of the four curvilinear
boundaries. These were calculated by the computer from specified radii (1.4 and
2.4) and circle centers (1.O, 0.0) and (1.O, 4.8). Remaining vertices were averaged
as in the previous example. This 90” circular-turnout grid required 386 iterations
in its generation.
These are but a few examples of rectangular grids that have been distorted into
a variety of different shapes, requiring only elementary algebra and trigonometry
on the part of the user. At times some experimentation must be used to arrive at
optimum grids if orthogonality is not a suitable criterion. Rather than being a
disadvantage, it is this feature that makes the present technique particularly useful,
as for example, in the caseof Fig. 2 where finer zoning was desired near the curved
358 AMSDEN AND HIRT
a)
b)
FIG. 6. (a) Initial 20 x 11 grid, where j = 6 row has zero height; (b) bifurcated flow grid
formed by iterative process.
boundary. The possibilities of these simple, intuitive, iterative techniques for the
generation of complex grids are unlimited, since the boundary and interior vertices
may be moved by any reasonable prescriptions that produce the desired grid
structure.
ACKNOWLEDGMENT
The authors acknowledge the assistance of F. H. Harlow and H. M. Ruppel for their helpful
discussions and advice on the manuscript. This work was performed under the auspices of the
United States Atomic Energy Commission.
REFERENCES
1. W. H. CHU, Development of a general finite difference approximation for a general domain.
Part 1: Machine transformation, J. Camp. Phys. 8 (1971), 392.
2. A. M. WINSLOW, “ ‘Equipotential’ Zoning of Two-Dimensional Meshes,” Lawrence Radia-
tion Laboratory, Report UCRL-7312, 1963.
CURVILINEAR GRIDS 359
3. A. M. WINSLOW, Numerical solution of the quasilinear Poisson equation in a nonuniform
triangle mesh, J. Comp. Phys. 2 (1967), 149.
4. W. D. BARFIELD,Numerical method for generating orthogonal curvilinear meshes, J. Comp.
Phys. 5 (1970), 23.
5. W. D. BARFIELD,An optimal mesh generator for Lagrangian hydrodynamic calculations in
two space dimensions, J. Comp. Phys. 6 (1970), 417.
6. C. W. HIRT AND A. A. AMSDEN,An arbitrary Lagrangian-Eulerian computing method for ail
flow speeds, J. Camp. Phys., to be published.
