JOURNAL OF COMPUTATIONAL PHYSICS 98,231-253 (1992)

Crystal Growth and Dendlritic Solidification

JAMES A. S~THIAN*

Department of Mathematics, University of California, Berkeley, California 94720

AND

JOHN STRAIN?

Courant Institute of Mathematical Sciences, New York University,

251 Mercer Street, New York, New York IO012

Received September 19,199O; revised May 29, 1991

review of the physics may be found in Langer [19] and

We present a numerical method which computes the motion of com- Chalmers [6], and detailed analysis may be found in many
plex solid/liquid boundaries in crystal growth. The model we solve papers, including Caginalp and Fife [3], Gurtin [13],
includes physical effects such as crystalline anisotropy, surface tension, Mullins and Serkerka [25,26], Nash and Glicksman
molecular kinetics, and undercooling. The method is based on two [27, 281, Laxmanan [21], Verdi and Visintin [44],
ideas. First, the equations of motion are recast as a single history-
dependent boundary integral equation on the solid/liquid boundary.
Wollkind and Notestine [45], Cahn and Hilliard [4],
A fast algorithm is used to solve the integral equation efficiently. Duchon and Robert [S].
Second, the boundary is moved by solving a “Hamilton-Jacobi”-type The physical motivation for this problem is as follows.
equation (on a fixed domain) formulated by Osher and Sethian for a Begin with a container of the liquid phase of the material
function in which the boundary is a particular level set. This equation under study, water for example. Suppose the box is
is solved by finite difference schemes borrowed from the technology of
hyperbolic conservation laws. The two ideas are combined by con-
smoothly and uniformly cooled to a temperature below its
structing a smooth extension of the normal velocity off the moving freezing point, so carefully that the liquid does not freeze.
boundary, in a way suggested by the physics of the problem. Our The systemis now in a “metastable” state, where a small dis-
numerical experiments show the evolution of complex crystalline turbance-such as dropping a tiny seed of the solid phase
shapes, development of large spikes and corners, dendrite formation into the liquid-will initiate a rapid and unstable process
and side-branching, and pieces of solid merging and breaking off
freely. Q 1992 Academic Press, Inc
known as dendritic solidification. The solid phase will grow
from the seed by sending out branching lingers into the
distant cooler liquid nearer the undercooled wall. This
growth processis unstable in the sensethat small perturba-
The moving boundary problems of crystal growth and
tions of the initial data can produce large changes in the
unstable solidification have attracted considerable interest
time-dependent solid/liquid boundary.
over the last few years. An outstanding problem of current
Mathematically, this phenomenon can be modeled by a
research is to find the time-dependent shape of a crystal of
moving boundary problem. The temperature field satisfiesa
a pure substance, growing from an undercooled bath of its
heat equation in each phase, coupled through two bound-
liquid phase. The goal is to understand the role played by
ary conditons on the unknown moving solid/liquid bound-
such physical parameters as the crystalline anisotropy,
ary, as well as initial and boundary conditions. Derivations
surface tension and molecular kinetics of the material, the
are given by Langer [19], Gurtin [13], Caginalp and Fife
undercooling imposed at the container walls, and the initial
[3], and much asymptotic analysis has been done (see
temperature distribution, and the solid shape. A good
Langer [19], Chadam and Ortoleva [S], Strain [40],
* Supported in part by the Applied Mathematics Subprogram of the Kessler and Levine [17], Ben Amar and Pomeau [l],
Oftice of Energy Research under Contract DE-AC03-76SFOOO98, and the Caginalp and Fife [3], Fix [ll], Langer [18], Pinus and
support of the National Science Foundation and DARPA under grant Taylor [31]), but no rigorous mathematical existence
DMS-8919074.
+ Current Address: Dept. of Mathematics, Princeton University,
theory is availabe at present. This is because the moving
Princeton, New Jersey 08544. Supported by DARPA/AFOSR Contract boundary conditions explicitly involve geometric properties
E-49620-87-C-0065 and an National Science Foundation Mathematical of the boundary itself, such as the local curvature and the
Sciences Postdoctoral Research Fellowship. normal direction, as well as the temperature field. Hence the
231 0021-9991/92 $3.00
Copyright 0 1992 by Academic Press, Inc.
All rights of reproduction in any form reserved.
232 SETHIAN AND STRAIN

transformation into a fixed domain, which has been used to Hence, the moving boundary can develop corners and cusps
analyze the classical Stefan problem for melting, breaks and undergo topological changes quite naturally. Second,
down. Even asymptotic calculations are difficult, because we reformulate the equations of motion as a boundary
surface tension constitutes a singular perturbation of the integral equation for the normal velocity as is also done in
Stefan problem. Langer [19], Meiron [22], Strain [38], Kessler and Levine
Models for crystal growth have also been studied through [17], Brush and Sekerka [2]. We then extend the normal
numerical calculations. One approach is to solve the heat velocity smoothly to the whole container, as required by the
equation numerically in each phase and to try to move the level set Osher-Sethian algorithm. These two ideas combine
boundary so that the two boundary conditions are satisfied to yield a fixed-domain formulation of the equations of
(seeChorin [7], Smith [37], Kelly and Ungar [16], Meyer motion, which may be of analytical interest itself. Our
[23], Sullivan et al. [43]). However, it is difficult to impose numerical method is based on solving a regularization of
the boundary conditions accurately on a time-dependent this equation, with entropy-satisfying upwind differencing
and complicated boundary. Thus these calculations have for the level set equation of motion and a fast algorithm for
been used mainly to study small perturbations of smooth evaluating the normal velocity. The method generalizes
crystal shapes.Another approach is to recast the equations immediately to higher order accurate schemes and, more
of motion as a single integral equation on the moving importantly, to three-dimensional problems. This general-
boundary and solve the integral equation numerically, as is ization will be presented elsewhere (Sethian and Strain
done in Meiron [22], Strain [38], Karma [lS], Kessler C361).
and Levine [17], Langer [19]. This approach can yield The paper is organized as follows. Section 1 reviews the
more accurate results for smooth boundaries, as well as equations of motion and some of the physical background.
agreement with linear stability theory. On the other hand, a Section 2 gives an overview of the numerical method, and
parametrization of the boundary must be computed at each Section 3 gives a flow chart. Section 4 describes the level set
time step. The curvature and normal vector are then equation for moving boundaries. Section 5 discussesa fast
derivatives of the parametrization, and these methods algorithm for applying the integral operator which occurs in
usually break down in the interesting cases where the the boundary integral equation. Section 6 contains some of
boundary becomes complicated and loses smoothness, as the many details of the numerical method, and Section 7
discussedin Sethian [32, 333. In particular, the calculations presents the numerical examples. Our conclusions are
presented in Strain [38] indicate that corners and cusps discussedin Section 8.
may form and pieces of the boundary may intersect; in In the numerical examples, we compute the motion of a
either case, the boundary cannot be parametrized by a growing crystal under a variety of conditions. We begin with
single smooth function. Finally, we note that most known calculations of crystal motion under purely geometric effects
numerical methods for crystal growth are difficult to extend and demonstrate the effectsof crystalline anisotropy on the
to problems in three space dimensions. evolving crystal. We then study a crystal growing from a
In this paper, we present a numerical method for crystal perturbed circular seed.For one set of values for the kinetic
growth problems which avoids these difficulties and com- coefficient, surface tension, and latent heat of solidification,
putes complex crystal shapes. Our method follows growth the resulting motion generates large smooth limbs. With
from an initial seed crystal or crystals (of arbitrary shape, another set of physical parameters, our model produces
size, and location) in a standard model which includes the intricate shapes which exhibit lingering, tip splitting, and
crystalline anisotropy, surface tension and molecular side branching. In all cases, these crystalline boundaries,
kinetics of the material, the undercooling imposed on the even when highly complex, remain virtually unchanged
container walls, and the initial state of the crystal/tem- under refinement of the numerical parameters, demon-
perature system. Our numerical calculations exhibit com- strating the robustness of our numerical algorithm. In
plicated shapes with spikes and corners, topological addition, we change physical parameters smoothly to show
changes in the solid-liquid boundary, dendrite formation, the evolution of tip-splitting and lingering. Finally, we study
and sidebranching. the interaction of an initial array of crystals, showing the
The method relies on two main ideas. First, we represent complex topological changes that result from various
the solid/liquid boundary as the zero set of .a function d choices of the physical parameters.
defined on the whole container. The boundary is then
moved by solving a nonlinear pseudo-differential equation
of suggestive of a Hamilton-Jacobi on the whole container. 1. THE MODEL EQUATIONS OF MOTION
This level set “Hamilton-Jacobi” formulation of moving
interfaces was introduced in Osher and Sethian [30] and We wish to model the growth of a solid crystal from a
allows us to compute geometric properties of highly com- seed in a undercooled liquid bath. We consider a model
plicated boundaries without relying on a parametrization. which includes the following physical effects:
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 233

(1) Undercooling. The sub-freezing temperature of

the walls and the initial liquid state. This is the driving force
for the growth of the solid.
(2) Crystalline anisotropy. The tendency of the crystal
to grow along certain axes of symmetry. The selection of a
particular symmetry is due to the microscopic structure of
the material; we treat it in a standard phenomenological
way in this macroscopic model.
(3) Surface tension. A curvature-dependence of the
boundary temperature which regularizes the problem by
smoothing corners and preventing excessive stretching of
the solid/liquid boundary.
(4) Molecular kinetics. A velocity-dependence of the
boundary temperature, which imposesan energy penalty for as
excessivelyfast motion of the boundary.
(5) Initial conditions. The shape, size, location, and
initial growth velocity of the initial seedcrystal or crystals,
and the initial temperature distribution in the liquid and
solid.
[I
xau =-Hv on r(r)

u(T,r) = - c&)C - rv(n)V for P on I-@)

We now state the equations of motion, Consider a square f&n) = C&l -Acos(k~e +&)I
container, B= [0, l] x [0, 11, filled with the liquid and
solid phases of some pure substance. The unknowns are c”(n) = cv(1 - Acos(k~e +00))
the temperature u(x, t) for x E B, and the solid-liquid
FIG. 1. The equations of motion and physical domain.
boundary ZJt).
The temperature field u is taken to satisfy the heat
temperature effects;setting H = 0 reducesthe crystal motion
equation in each phase (see Fig. 1), together with an initial
to pure geometry.
condition in B and boundary conditions on the container
The second boundary condition on f(t) is the classical
walls. Thus
Gibbs-Thomson relation, modified to include crystalline
anisotropy and molecular kinetics as well as the surface
u,=Au in Boffr(t)
tension:
4% t) = %3(x) in Bat t =0 (1.1)
u(x, t)= -Ec(n)C-E,(n) V
44 t) = UB(X) for XE~B.
for x on f(t). (1.3)
Since the position of the moving boundary r(t) is unknown,
Here C is the curvature at x on f(t), taken positive if
two boundary conditions on f(t) are required to determine
the center of the osculating circle lies in the solid. The
u and f(t). Let n be the outward normal to the boundary,
anisotropy functions are modeled by
pointing from solid to liquid. The first boundary condition
is the classical Stefan condition:
q-(n)=E=(l -A cos(k,O+&,)) (1.4)
[&4/&z] = -HV on f(t). (1.2) .s,(n)=s,(l -A cos(k,8+8,)), (1.5)

Here [au/an] is the jump in the normal component of heat where 8 is the angle between n and the x-axis, and sC, E “, A,
flux au/& from solid to liquid across I(t), V is the normal k,, and B0are constants depending on the material and the
velocity of f(t), taken positive if liquid is freezing, and H is experimental arrangement. For example, if E,= = 0 (E y = 0),
the dimensionless latent heat of solidification, which is a there are no surface tension (molecular kinetic) effects.For
constant. The signs of geometric quantities are chosen so A = 0, the system is isotropic, while if A >O, the solid is
that if au/an < 0 in the liquid phase and au/&r = 0 in the k,-fold symmetric with a symmetry axis at angle f& to the
solid phase, then [au/an] is negative and V > 0, indicating x-axis. Typically A < 1. The equations of motion under
that the solid phase is growing. Physically, this means that study are (l.l), (1.2), and (1.3), with anisotropy functions
undercooling drives solid growth. The latent heat of given by (1.4) and (1.5).
solidification controls the balance between geometry and For appropriate initial data u0 and boundary data us, an
234 SETHIAN AND STRAIN

exact solution exists in which Z(t) is a line and u is an I

exponentially decaying traveling wave

“(x,r)=~~(....)drw
Ol-

r(t)={y= V?,XE(O, l)} (1.7)

44 Y>f) = e- Q- .,p3 (1.8)

-l-&,V,
l-0)
+
u, =Au
where V is a positive constant and E,,= E,,(O). Linear
stability of this solution has been examined in Langer [ 191,
Chadam and Ortoleva [S], Pinus and Taylor [33], Mullins
and Sekerka [26], Strain [41]. If .sC= .sy= 0, then this
solution is linearly unstable with a disturbance cos(kx)
growing like e Vlklt. If E=> 0, only a finite band of low
wavenumbers are unstable; the high-k instability is damped
by the introduction of surface tension. Making Ey > 0 does
Level set f$=O
,, ..-. . ..\

! I
_, ___...-’ :
!
!. L(
$=-I
....’
,. -. _,’
,: :>-.
: :’ ‘i .,
V

!‘.. . -..-.
/. ,.__..... .’ 3’ -..... : ; ;....--~ . .... .. ,,

6
not change the picture qualitatively. _,_ ,,$=I ,.._ / J j j, ‘.. :
,,. _.... .‘,,,’
‘... _,, ‘.,
More general models have been widely discussed in the ‘:: : : . . . . . ....
., : : ..’
i 0, ! j
literature. The assumption of equal heat diffusion coef- ‘:, ‘...,
/!
‘..: ,.: ;
+ .. . . ..-.
:
ficients in solid and liquid, for example, is roughly valid only
for certain plastic crystals like succinonitrile, and certainly
highly invalid for the ice/water system. It is also interesting
to treat problems in which the domain is infinite; this FIG. 2. The flow of the algorithm: (a) position of solid liquid bound-
ary in B; (b) boundary integral formulation; (c) level set formulation;
removes the effects of the finite container size and makes it (d) extension of velocity field to level set formulation.
possible to compare numerical results with some classical
similarity solutions of the Stefan problem, for example. The
results presented in this paper are not directly comparable domain into an equation which involves only the moving
with known exact solutions like expanding circles, so our boundary and its previous history.
only check on the method will be convergence under mesh In Step 3, we represent the boundary r(t) as the level set
refinement. However, our method can be extended to treat { ~5= 0) of a function 4 defined on all of B (see Fig. 2~). We
these more general problems; we explain it in this case construct a nonlinear pseudodifferential equation which
mainly to simplify the explanation of what is already a evolves 4 in such a way that the zero set { 4 = 0}, at each
rather complicated algorithm. time t, is the moving boundary r(t) (seeFig. 2d). To do this,
in Step 4 we construct a “speed function” F, which is equal
to the normal velocity V on r(t) and smoothly extends V to
2. OUTLINE OF THE NUMERICAL ALGORITHM all to B. We then move all the level sets of 4 with normal
velocity F. The resulting level set equation for 4, which is
In this section, we sketch the numerical method we use to reminiscent of a Hamilton-Jacobi equation, may be solved
solve the equations of motion. We present the algorithm in numerically by finite difference schemes borrowed from
four steps.In the first two steps,we transform the equations hyperbolic conservation laws. A major advantage of this
of motion into a boundary integral formulation. In the last formulation is that the equation for 4 can be solved on a
two, we describe the level set formulation for moving the uniform mesh on the box B; the level setsare moved without
boundary and the necessary extension of V. The details of constructing them explicitly.
each step may be found in the sections that follow. However, We now describe the algorithm in more detail.
the fundamental philosophy behind the algorithm is most
easily conveyed through a series of figures.
In Fig. 2a, we show a typical solid/liquid boundary. Step 1. Subtraction of Initial and Boundary Conditions
According to Eqs. (1.1)-( 1.6), the temperature field u must
Recall the equations of motion (l.l)-( 1.6). The tem-
satisfy the heat equation in each phase, as well as two
perature field u and moving solid/liquid boundary r(t)
boundary conditions on the phase boundary. In Steps 1 and
satisfy
2, we transform the equations of motion into a single
boundary integral equation on the moving boundary (see U,=dU in B--T(t)
Fig. 2b). Thus, we transform equations of motion which
require computation of the temperature field u on the whole 4-T t) = uo(x) for t=O, (2.1)
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 235

4x, t) = UB(X) for xCC?JB B = [0, l] x [0, l] with Dirichlet boundary conditions on
the box walls is given by
[au/an] = -HV on r(t) (2.2)
u(x,t)= -EC(n)C--EY(n)v K(x, x’, t’) = f f eC(kf+ki)n2rsin(k,rcxi)
for x on r(t). (2.3) k,=l kz=l

x sin(k,7rx2) sin(k,nx;) sin(k,nx;) (2.9)

First, we subtract the temperature field due to the initial
conditions u0 and boundary conditions us. Let U(x, t) be =& 2 f c c 0102
the solution to the heat equation k,=-m kz=-cc a,=+1 cq=*l

x e- [(xl --blxi -=I)~+ (-a2+2k2)21/4f, (2 10)

U,=AU in B
U(x, 0) = uo(x) at t=O (2.4) where x = (x1, x2) and x’ = (xi, xi). The first expression for
K can be calculated by Fourier series, the second by the
U(x, f) = u,(x, f) for xEaBand t>O. method of images. The equivalence of the two is a special
caseof the Poisson summation formula (Dym and McKean
Let W= u - U. Then W satisfies
C91).
The function SV(x, t) defined by Eq. (2.8) is a continuous
W,=AW in B-T(t) function on B x [0, T], vanishing for t = 0 or x on aB, and
W(x, t)=O at t =O, (2.5) satisfying the heat equation everywhere off rT. Across r(t),
however, SV(x, t) has a jump in its normal derivative
W(x,t) = 0 for XEI~B equal to V. Thus, W(x, t) = H. SI’(x, t) is the solution to
[awlan] = -HV on T(r) (2.6) Eqs. (2.5)(2.6). All that remains is to satisfy the second
boundary condition Eq. (2.7). This is equivalent to the
W(x,t) = -&c(n)C- &v(n)v boundary integral equation
- w, t) for x on r(t). (2.7)
c,(n)C+ cV(n) V+ U
Equations (2.5)-(2.7) are equivalent to the original equa-
tions of motion. Given the solution U, at any time we can K(x, x’, t - t’)
+Hs’.i0 l-(f)
add W to U to produce the solution u to the original
problem. The “free” temperature field U(x, t), which is x T/(x’, t’) dx’ dt’ = 0 (2.11)
defined on a fixed domain, may be found analytically for
simple initial and boundary conditions, or by numerical for x E r(t). Equation (2.11) is an integral equation for the
calculation in practical situations. normal velocity of the moving boundary. We note that the
velocity V of a point x on r(t) depends not only on the
Step 2. Transformation to a Boundary Integral Equation position of r(t) but also on its previous history. Thus, we
The next step is to transform Eq. (2.5)-(2.7) into a have stored information about the temperature off the mov-
integral equation on the boundary r(t). We use the kernel ing boundary in the previous history of the boundary.
K of the heat equation to express the solution W to
Eqs. (2.5)-(2.6) as a single layer heat potential. Given a Step 3. Level Set Formulation of Moving Boundaries
function V on Numerical approximations to the boundary integral
equation based on marker particle techniques have been
used to solve solidification problems similar to the one
under investigation (see Meiron [22], Strain [38], Kessler
and Levine [ 171, Karma [IS], Brush and Sekerka [2] ). In
the single layer heat potential SV is defined for (x, t) in this approach, the boundary r(t) is represented by a set of
BxCO,Tlby marker points. A discretized version of the integral equation
is then solved for the velocity of each point, and this velocity
is then used to update the positions of the markers. There
W-T t) = j; s,,,) K(x, x’, t - t’) V-(x’, t’) dx’ dt’. (2.8) are several difficulties with this approach. First, at each time
step one must evaluate an integral over all previous history.
Here the x’ integration is over the curves comprising r(t’), Thus, given 1 < i< M points on the curve at each of
and the Green function K of the heat equation in the box 1 d n < N time levels n At, it costs O(M2N2) operations just
236 SETHIAN AND STRAIN

to evaluate SV at the given points. This becomes very given in [24]. Let x(t) be the trajectory of a particle located
expensive for large M and N, as documented in Strain [ 381 on this level set, so
and Greengard and Strain [12]. Independent of the
expense, fundamental problems with marker particle algo- W(t), t) = c. (2.15)
rithms arise as the moving boundary r(t) becomes more
complex (see Sethian [32, 331). Sharp corners and cusps The particle speed ax/& in the direction n normal to r(t) is
can develop which are difficult to resolve with marker par- given by the speedfunction F. Thus,
ticles, and instabilities often result when the boundary
becomes interesting. Furthermore, it is difficult to use ax
marker particles when the boundary changes topology, for --n=F, (2.16)
example, when two boundaries collide or a piece of solid
at
breaks off. Finally, the extension to three dimensional
problems is challenging. where the normal vector n is given by n = V&lV#l. By the
An alternative to marker particles for moving boundaries chain rule,
is provided by the level set technology introduced in Osher
and Sethian [30]. With this approach, a complex boundary
can be advanced. Sharp corners and cusps are handled
(,+$hJkO
naturally, and changes of topology in the moving boundary
require no additional effort. Furthermore, these methods and substitution yields
work in any number of space dimensions. Recently, this
technique has been applied to interface problems in mean d,+FIVbI =O
curvature flow [33], singularity formation and minimal
4(x, t = 0) = given.
surface construction [34], and compressible gas dynamics
[24]. In addition, theoretical analysis of mean curvature
flow based on the level set model presented in Osher and Equation (2.18) yields the motion of Z(t) with normal
Sethian [30] has recently been developed in [lo]. velocity V on the level set 4 = 0. We refer to Eq. (2.18) as the
We now describe the level set algorithm in the general level set “Hamilton-Jacobi” formulation. It is not strictly
casewhen a curve or union of curves f(t) moves with speed a Hamilton-Jacobi equation except for certain speed
V normal to itself. The essential idea is to construct a functions F, but the flavor of Hamilton-Jacobi equations is
function 4(x, t) defined everywhere on B, such that the level present.
set { 4 = 0} is the set r(t); that is, The initial value problem (2.18) can be solved numeri-
cally by finite difference schemes. Because 4 can develop
corners and sharp gradients, numerical techniques
r(t) = {x$(x, t) =O}. (2.12) borrowed from hyperbolic conservation laws are used to
produce upwind schemesfor 4 which track sharp corners
We now derive a partial differential equation for 4, which accurately and employ the correct boundary conditions at
holds on Bx [0, T]. Suppose we can construct a smooth the edge of the computational box. These schemes are
function F(x, t) defined on all of B such that presented in Section 4.

V(x, t) = F(x, t) for x E r( t ). (2.13) Step 4. Construction of the SpeedFunction F

We now describe how to extend the velocity V to a
Then V = F on Tr, and we call F a smooth extension of V off globally defined speed function F. Such an extension is
r(t). We shall postpone until the next section the extension necessaryin order to use the level set formulation.
for the case of crystal growth. The most natural extension makes direct use of the
What is the equation of motion for d? Obviously, this integral equation (2.11) namely,
equation must reduce to normal propagation by speed V on
the level set 4 = 0. Suppose we initialize 4(x, 0) such that E,(n)C+ Ey(n) V+ U

4(x, 0) = + distance from x to r(t), (2.14) K(x, x’, t - t’)

+“idLW
where the minus (plus) sign is chosen if x is inside (outside) x V(x’, t’) dx’ dt’ = 0 (2.19)
the initial boundary r(t = 0). Now consider the motion of
some level set {4(x, t) = C}. Here, we follow the derivation for x E r(t). Each term in Eq. (2.19) can be evaluated
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 231

anywhere in B, once V is known on r(t’) for 0 d t’ 6 t and following regularized extension of V. Let F be defined
4 is known on B. Thus, given the set f(t), plus all its previous everywhere in B by
positions and velocities for 0 < t’ 6 t, one could first solve an
integral equation to find the velocity Yfor all points on f(t) ~c(n)C+~,(n) F(x, t) + W, t)
and then find F(x, t) by solving the equation + H $j7t F(x, t) + HSBY= 0. (2.24)

EC(n)c+Ev(n)mt)+w, t) We can then solve this equation for F to produce

-1
+H1;s,(,,)
ax, x’,t- t’) F=
dn)+ffJG
[Ec(n)C+ U+HS5V]. (2.25)

x V(x), t’) dx’ dt’ = 0 (2.20)

Thus, F is an explicit function of 4, the velocity at previous
times t’ < t - 6, and the free temperature field U.
for 1; throughout B. Here, x is a point in B, while C = C(x, t) Note that F is defined even if EV= 0. Furthermore, F is
is the curvature, and n is the outward normal vector to the equal to V+ O(S3j2)on f(t), and does not have a boundary
level set passing through x; layer as we move off the curve, as long as 6 is not too small.
Finally, in the limit 6 -+ 0, Eq. (2.25) reducesto Eq. (2.20) as
it should.
c=v. g =V.n
c > Remark. We have reduced the equations of motion,
(2.21) with an O(a3j2) error, to a pair of equations on a fixed
+v(s
domain B:
n=$q
&,(n)C+.s,(n)F+ U(x, t)

and these expressions make sense everywhere in B. The +HJ$iF+HS,,V=O (2.26)

speedfuncion F given by Eq. (2.20) is defined throughout B 4, + F IV4( = 0. (2.27)
and equal to V on the solid/liquid boundary f(t).
For our purposes, it is easier to work with a regulariza- These equations constitute a fixed-domain approximation
tion of the speed function. We split the single layer heat of the crystal growth equations of motion. No analysis of
potential into a history part Sd V and a local part S, V as the original equations of motion has been carried out,
follows: mostly because there is no fixed-domain formulation. An
approximate fixed-domain formulation may be a viable
wx>t)=y lrct,,
K(x, x’, t - t’) V(x’, t’) dx’ dt’ substitute.

K(x, x’, t - t’) V(x’, t’) dx’ dt’ 3. OUTLINE OF THE ALGORITHM

In this section, we describe the general flow of the algo-

~S,V/+S,V. (2.22)
rithm from one time step to the next. We have a pair of
equations for the speedfunction F and the level set function
Here 6 is a small regularization parameter. Heuristically, we 4 defined throughout B, namely
try to separate the local part, which is causing the jump in
the normal derivative of the potential, from the history part .+(n)C+&,(n)F+ U(x, t)
which is smooth and independent of the current velocity.
In Section 5, we show that the local part S, V can be +H&F+H&V=O (3.1)
approximated by q5,+FIV&=O. (3.2)

s, V(x, ?)= Js/71V(x, t) + o(63’2) (2.23) Here, the curvature C and the normal vector n are func-
tionals of 4. U is computed by solving the heat equation on
B. Sd V is computed from the previous history of r(t).
at points x on r(t). The history part SdV depends only on To describe the algorithm, we imagine that at time step
values of V at times t’ bounded away from the current time, n At we have the following information:
since t’ d t - 6 in S6 V. Since the leading term fi V in the
expression for the local part S, V varies smoothly in the (1) the level set function #z defined at discrete grid
direction normal to the curve if V does, this suggests the points xii
238 SETHIAN AND STRAIN

(2) the free temperature field 17; defined on the same methods based on parametrizing r(t). The first is that
discrete grid 4(x, t) always remains a function, even if the level surface
(3) the previous positions of the boundary Z(m dt), 4 = 0 corresponding to the boundary T(t) changes topol-
m = n - 1, n - 2, .... n - d (where d= h/At), stored as points ogy, breaks, merges, or forms sharp corners. Parametriza-
on each curve. tions of the boundary becomemultiple-valued or singular in
these cases. As an example, consider two circles in R2
Given this information, we proceed from one time step to expanding outward with normal velocity V= 1 (Fig. 3a).
the next as follows: The initial function 4 is a double-humped function which is
ALGORITHM. continuous but not everywhere differentiable (Fig. 3b). As 4
evolves under Eq. (4.2), the topology of the level set 4 = 0
Step 1. At each grid point xii, compute the extended changes.When the two circles expand, they meet and merge
speed function F; from (3.1). This is done as follows: into a single closed curve with two corners (Fig. 3~). This is
(a) Expressions for the discrete curvature C, and reflected in the change of topology of the level set 4 =0
normal vector nij may be computed from the (Fig. 3d).
discrete level set function 4; without explicit Another advantage of this formulation concerns numeri-
construction of the particular level set passing cal approximation. Because4(x, t) remains a function as it
through the grid point xii (see Section 4). evolves, finite difference approximations may be used to
(b) The history part S6 V is updated by using the approximate the spatial and temporal derivatives. For
stored boundary at previous time levels, This part example, given a uniform mesh with spacing h, with grid
of the calculation is described in Section 5. nodes xii and time step At, let 4; approximate &x,, n At).
An approximation to Eq. (4.2) is given by
Step 2. Calculate 4;’ ’ from 4; and F; using the
upwind, finite difference schemedescribed in Section 4. qy.’ l - 4;.
Step 3. Calculate U;+ ’ from Us by solving the finite L +F”, IV/&q =o.
At
difference approximation to the heat equation discussed in
Section 6.A. Here, we have used forward differences in time, and some
Step 4. Find point on the new boundary r((n + 1) At) appropriate finite difference operator V,dz. to approximate
by constructing the level set 4 = 0 from di”, (see the spatial gradient. In this section, we shall discuss how
Section 6.C). Store the position x and velocity V of each accurate and efficient approximations V,d; are obtained by
point, found by interpolating F from the values on the grid. exploiting the close link between Hamilton-Jacobi equa-
Note that these points on the boundary do not move. They tions and hyperbolic conservation laws. Details may be
serve only as quadrature points for updating the history found in [30].
integral S6 V.
Step 5. Replace n by n + 1 and return to Step 1. a b I
In Section 4, we present the details of the level set algo-
rithm and in Section 5 we give the details of the evaluation
of the history part of the single layer potential. Section 6
presents other numerical details.
Level Set 4 = O\ x
4. SOLVING THE LEVEL SET EQUATION

We have reformulated the problem of moving r(t) with

normal velocity V by representing r(t) as the zero set
z-(t) = {x 1fj(x, t) = O}. (4.1)

Here 4 solves the initial value problem

d,+f’IV4 =O
(4.2) Level Set 4 = 0
4(x, t = 0) = f distance from x to r( t = 0)
FIG. 3. Formulation of level set approach: (a) propagation of two
and F is a smooth extension of V off r(t). There are at least circles; (b) initial value of level set function; (c)merger of propagating
two advantages of this level set formulation compared to circles; (d) change of topology of level set 4 = 0.
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 239

A. Corner Formation and the Entropy Condition a

To illustrate, suppose we wish to follow an initial cosine
curve r(t = 0) propagating with normal velocity
V = 1 - EC, where C is the local curvature of the boundary.
The curvature of the level curve passing through a point
(x, y, t) is given by

Once again, the minus sign occurs because we have

initialized the surface so that Vd points inwards, and we
want C to be positive for a circle. The smooth extension of
V to F is straightforward, and Eq. (4.2) becomes

(4.5)
4(x, y, t = 0) = f distance from (x, y) to r( t = 0). FIG. 4. Comer formation and the entropy condition: (a) propagating
curve until singularity forms; (b) entropy-violating swallowtail solution;
As shown in [32], for E> 0, the parabolic right-hand side (c) entropy-satisfying solution from Huyghen’s construction.
diffuses sharp gradients and forces 4 to stay smooth for
all time. This is not true for E= 0 and F= 1. In Fig. 4, we the fact [32] that our entropy condition for propagating
show the level set 4 = 0 corresponding to the propagating boundaries is identical to the one for hyperbolic con-
boundary r(t). For E = 0, the boundary moves with unit servation laws, where stable, consistent, entropy-satisfying
speed,and a corner must develop, since the two sides come algorithms have a rich history. Our next step is to exploit
together at a constant rate (see Fig. 4a). Thus, singularities this link.
in the curvature C develop from smooth initial data.
What happens once a corner develops? It is not clear how B. Hamilton-Jacobi Equations and Hyperbolic Conservation
to propagate in the normal direction, since the derivative is Laws in One Dimension
not defined at the corner. One possibility is the “swallow-
tail” solution found by letting the boundary pass through As motivation, consider the one-dimensional version of
itself, see Fig. 4b. However, the boundary consists only of the level set equation, with constant normal velocity V= 1.
those points located a distance t from the initial curve, Then Eq. (4.2) becomes a standard Hamilton-Jacobi
which is the Huygens’ principle construction. Roughly equation
speaking, we may obtain this solution by removing the
“tail” from the “swallowtail.” In Fig. 4c, we show this alter- 9, + H(4,) = 0, where H(#,) = - (d,T)1/2
native “weak” solution. Another way to characterize this (4.6)
6(x, 0) = given.
weak solution is through the following “entropy condition,”
see [32]: If the boundary is viewed as a burning flame, then Let u = 4,. Taking the derivative of Eq. (4.6) with respectto
once a particle is burnt it stays burnt. Careful adherence to x, we have
this stipulation produces the same weak solution as the
Huygens’ principle construction. Furthermore, this physi-
cally reasonable weak solution has an equally appealing
u,+ CH(u)l,=O, where H(u) = - (u’)I/~
(4.7)
mathematical quality. It is the formal limit of the smooth u(x, 0) = given.
solutions for E> 0 as E-+ 0 and the curvature term
vanishes [ 321. Equation (4.7) is a single scalar hyperbolic conservation law
Thus, our goal is to produce approximations to the in one space variable. Solutions can develop discontinuous
spatial derivative that (1) do not artificially smooth sharp jumps, known as shocks, even with smooth initial data
corners and (2) pick out the correct entropy solution [20]. Jumps in the slope u = 4, are equivalent to corners in
when singularities develop. Our schemesare motivated by I$ [32]. In order to make senseof the solution after shocks
240 SETHIAN AND STRAIN

form, we study an integral version of the conservation law have used the standard definitions of the forward and back-
(4.7) which admits discontinuous solutions. Thus, consider ward difference operators, namely,
an arbitrary interval [a, 61. We integrate both sides of
Eq. (4.7) to produce

(4.14)
f 1” 24(x,t) dx = H[u(a, t)] - H[u(b, t)]. (4.8)
a

We say that u is a weak solution of the conservation law if Finally, we need an appropriate numerical flux function
it satisfies the above integral equation. Note that u need not g. In the special case where H(u) may be written as a
be differentiable to satisfy the integral form of the conserva- function of u’, i.e., H(u) = f(u’) for some functionf, we may
tion law. use the Hamilton-Jacobi flux function given in [30]:
How can we be sure that a numerical algorithm will
approximate the correct, entropy-satisfying solution to d”i- l/23 Ui+ l/2 ) = g(D; di, 0,’ 4i)

Eq. (4.8)? We begin with a definition, which is simply a

numerical version of Eq. (4.8): =f((max(D; 4, 0))2 + (min(Dz 4, 0))‘).
(4.15)
DEFINITION. Let u’ be the value of u at mesh point i Ax
at time n dt. A three-point difference schemeis said to be in This conservative monotone schemeis an upwind method in
conservation form if there exists a function g(u,, u2) such that it differences in the direction of propagating charac-
that the schemecan be written in the form teristics. This is important, since we must impose boundary
conditions on the walls of a finite-sized computational box.
An upwind scheme automatically differences in the out-
ward-flowing direction at the box walls if the boundary is
expanding, thus information flows out. In the special case
where g(u, U) = H(u). (4.9) when V= 1, so thatf(u2) = -(u’)“~, we have

This definition is natural; the scheme must approximate &“I = 47 -At ((max(D; 4, 0))2 + (min(D,+ 4, O))2)1’2.
the hyperbolic conservation law, subject to the consistency (4.16)
requirement g(u, U) = H(u). In order to guarantee that the
scheme picks out the correct entropy-satisfying weak solu- This algorithm produces the correct entropy-satisfying
tion, we require monotonicity, that is, that u;” be an weak solution to the moving boundary problem. More
increasing function of the arguments ~1~i, ~1, u;+, . We can details about shock schemesand conservation laws may be
now state the main fact: A conservative, monotone scheme found in [29].
produces a solution that satisfies the entropy condition.
How do we convert Eq. (4.9), which is a schemefor the C. Multimensional Problems and More Complicated Speed
slope U, into a scheme for 4 itself7 We begin by writing Functions
Eq. (4.6) with a forward difference in time, namely,
We now extend the previous discussion to problems in
&+‘=#;-AtH(u). (4.10) more than one spacedimension. Recall that we are solving
the “Hamilton-Jacobi”-type equation
Since the numerical flux function g approximates H, the
solution to Eq. (4.10) may be approximated by &+FIVqhI =O. (4.17)

We begin by decomposing F into two components:

on-17-l &,I-& F=F,+Fo. (4.18)

x&y-Atg Ax , Ax 7 (4.12)
>
Here, FA is an advection term containing that part of the
= 47 - At g(D; di, 0,’ $i), (4.13) velocity which is independent of the moving boundary, and
FG contains those terms which depend on geometric proper-
where g is an appropriate numerical flux function and we ties of the boundary, such as the curvature and normal. We
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION

begin by splitting the influence of F, and rewrite Eq. (4.17) a

as

4t= - (FAIV41+ Fo IV41)- (4.19)

In two space dimensions, one can easily extend the scheme

given in Eq. (4.12) by differencing in each direction to
produce

4;’ ’ = q5;- F, At ((max(D; dii, O))*

+ (min(DZ ii,, O)Y+ (max(DJ: 4v, 0))’
+ (min(Dt d,,, O))*)‘j2 - At F, IV&. (4.20)

Here, we have not approximated the final term F, IVq5l;one

may use a straightforward central difference approximation
to this term. This is the first-order multidimensional
algorithm described in [30].
In Fig. 5, we show this technique applied to the caseof a
star propagating outwards with speedF = 1 - EC, where C
is the curvature as in Eq. (4.5). Here, we take At = 0.00025 FIG. 6. Grid effects in anisotropic motion: (a) grid effects: sixfold
and a mesh size of 160 points in each direction on a unit symmetry, phase angle = 0.0; (b) grid effects: sixfold symmetry, phase
box (Ax = dy = l/159). The curve evolves smoothly and angle = 22.5”; (c) grid effects:sixfold symmetry, phase angle = 45.0”.
becomescircular as it evolves. The orientation of the under-
lying grid used to advect the circle is difficult to detect.
suffers from grid effects. In Fig. 6, we show the effects of
F. “Grid-Free” Extension to Multi-Dimensions the algorithm applied to a circle moving under a sixfold
anisotropy (k, = 6) speed function. Here, we have taken
Unfortunately, when the speed function F is more
EC=O.Oi, &y = 0.00. In Fig. 6a, the phase angle 8, = 0, thus
complicated, the above extension to multi-dimensions
the preferred modes of growth are along the axes at O”, 60”,
is inadequate. In particular, for the anisotropic speed
and 120”. In Fig. 6b, the phase angle is shifted to 0,, = 7c/16;
function
thus the preferred growth direction axes are shifted accor-
dingly. In Fig. 6c, the phase angle is shifted to Q0= n/8. It is
F= 1 -s&l -A cos(k,8+8,))C, (4.21)
clear that grid effectsare pronounced, as shown in the asym-
metry of the evolving shapes.
which is the crystal growth speed function given in
This problem arises because of our construction of the
Eq. (2.25) in the limit H+ 0, 6 + 0, the above method
length of the gradient. We have used the approximation

IV41= (4: + 4y2

z ((max(D; 4, O))‘+ (min(D,f 4, 0))’
+ (max(D, 4, 0))2 + (min(D: 4, O))2)“2. (4.22)

As the family of level sets sharpen along the lines of

anisotropy, they develop sharp corners. At those corners,
the approximation of jV& is biased with respect to the grid
lines. As can be seenfrom Fig. 6, when the preferred lines of
growth do not lie along the x and y axis, the problem is
substantial.
Grid effects may be significantly diminished by the
following alternative multi-dimensional scheme. At each
grid point, we find the normal vector to the level curve
(4 = Const) passing through that grid point by averaging
FIG. 5. Expanding star, F(K) = 1 -0.01 K. the one-sided limits of the normal vectors. We then perform
242 SETHIAN AND STRAIN

one-dimensional upwind differencing in the calculated The discrete unit normal nq calculated above can then be
normal direction. used to construct a relatively “grid-free” multi-dimensional
We now explain this strategy in detail. If 4 is smooth, then scheme for moving boundaries. We begin by constructing
the unit normal at point (x,, y,,) to the level curve passing the forward and backward differences, approximating the
through (x,,y,) is Vd(x,)/lV+(x,,)l. However, when 4 directional derivative in the direction no at the grid point
develops corners, care must be taken in the construction of (i Ax, j Ay ), namely,
this normal, since the speedfunction F is highly sensitive to
the exact direction of the normal in the anisotropic case D+i, =cwx+A s iv, j Ay + As pii) - dii
given by Eq. (4.21). At a corner, the direction of the normal I/ (4.24)
AS
undergoes a jump. This suggests the following technique.
We form the one-sided difference approximation to the unit D-~,=~“.-~(iAx-AsBi,,jAy-AsPd)
(4.25)
normal in each possible direction. We then average all four As
limiting normals to produce the approximate normal at the
corner, seeFig. 7: with step size As = min(Ax, Ay). The values of 4 in between
grid points are found by interpolation from neighboring
grid values. The above forward and backward difference
operators are used to construct an upwind approximation
to the gradient based on the one-dimensional schemegiven
in Eq. (4.12), namely,

qS;+‘=q$-F,,,At
CD,+dii, D; d&l
+ (CD.: hjJ2 + (0; &)2)1’2) x (max(D-4+ 0)2 + min(D+@ii, O)*))“’

(D; bij, D; 4J -At CFci,IV41J. (4.26)

+ ((D, &I2 + (0; c&,-)*)“~)
To summarize, we first construct the normal vector and
CD; dp D, dij)
+ ((D; c&J2 + CD.; hj)2)“2) 1.
(4.23)
then use one-dimensional upwind differencing in the
calculated normal direction. Note that this methodology
may be extended to any number of spacedimensions. Equa-
If any of the one-sided approximations to IV41 in the. tion (4.26) is the schemeused in this paper.
denominator vanish, we ignore that component and
re-weight accordingly. This discrete unit normal vector a
ng = (a,, pii) is then used to compute 8, in the anisotropy-
based speedfunction Fv given in Eq. (4.21).

(i-lj) ii (i+lJ)
\
\
/ \\

FIG. 8. Modified approach: (a) Modified method: sixfold symmetry,

phase angle = 0.0”; (b) Modified method: sixfold symmetry, phase
FIG. 7. Construction of modified method. angle = 22.5”; (c) Modified method: sixfold symmetry, phase angle = 45.0”.
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATlON 243

on a M x M grid in B. The heat kernel K is given (seeDym

and McKean [9] and Strain [39]), by

K(x, x’, t)= 2 5 ~-‘k:+k:fn2zsin(k,nx,)

k,=l kz=l

x sin(k,nx*) sin(k, XX’,)sin(k,zx;) (5.2)

x e- [(XI -o(.T~ - 2k1)*+ (~2-02.~-22k2)~1/4f

(5.3)
A straightforward numerical approach to calculating
ST/(x, t) is to use a quadrature formula to approximate the
integral. To compute the values of SV at M2 points at N
times n dt would then require O(M3N2) work if there are
O(M) points on f(t) at each time t. Since complicated
moving boundaries can develop in solidification problems,
accurate approximations require large values for M
and N and straightforward evaluation schemes become
prohibitively expensive, as shown in [ 38, 121for evidence of
this.
As an alternative, we can use a fast algorithm which
evaluates SV(x, t) on a M x M grid in O(M ‘) work per time
step, to any specified precision. The constant in O(M2)
depends on the precision and the smoothness of the data
f(t) and V. The basic idea behind this algorithm is as
follows. Suppose we split the time integral in SV(x, t) at a
L
cutoff time 6:

FIG. 9. Calculation of motion under various crystalline anistropies:

(a) Threefold symmetry; (b) Fourfold Symmetry; (c) Fivefold symmetry;
wx, t)= 1;-”s,,,,)K(x, x’, t - t’) V(x’, t’) dx’ dt’
(d) Sixfold symmetry; (e) eightfold symmetry; (f) tenfold symmetry
“, ”

K(x, x’, t - t’) V(x’, t’) dx’ dt’

In Fig. 8, we show the results of this new algorithm + Jr.-6?,,,,
applied to the previous example of a circle using the fourfold =SgV+SLV. (5.4)
anisotropy speedfunction. In Fig. 8a, the phase angle 0 = 0;
thus the preferred modes of growth are along the x and y The key to evaluating Eq. (5.4) is to use the two different
axes. In Fig. 8b, the phase angle 8 = 7c/8;thus the preferred expressions, (5.2) and (5.3), for the kernel K. Substitution of
modes of growth are along the diagonals. In both cases,the Eq. (5.2) into the history part S6V yields a Fourier series
circle is transformed into a square, and the grid effectsare which can be approximated by a small number of terms. In
considerably diminished. In Fig. 9, we show this modified addition, the Fourier coefficients of S, V can be updated in
algorithm applied to problems with threefold, fourfold, 0( M 2, work per time step, using Rokhlin’s non-equidistant
fivefold, sixfold, eightfold, and tenfold symmetries. In each fast Fourier transform (Rokhlin, unpublished manuscript;
case, grid effects have been essentially eliminated by using Strain [39]). On the other hand, substitution of the method
the modified approach. of images sum (5.3) into the local part S, V produces a
rapidly convergent sum, which we can approximate by
5. RAPID EVALUATION OF THE Taylor seriesexpansion of the first term.
SINGLE LAYER HEAT POTENTIAL A less robust fast algorithm which evaluates SV(x, t) for
x in r(t) using M quadrature points on r(t) and requiring
The boundary integral formulation of the problem only O(MN) work for N time steps has been constructed in
required the evaluation of the single layer heat potential [ 121. However, we do not need this algorithm here, because

SV(x,
t)=
’
ss 0 l-(C)
K(x, x’, t - t’) V(x’, t’) dx’ dt’, (5.1)
we are evaluating only S6V on a M x M grid and we are
evaluating it with O(M) quadrature points given in f(t’).
This simplifies the problem considerably.

5x1 ‘9R.2.5
244 SETHIAN AND STRAIN

A. The History Part S, V We can now use the integral test to bound the error E,
We begin with the history part S6V. We have incurred by truncating the series(5.9) after the first p* terms:

K(x, x’, t - t’) V(x’, t’) dx’ dt’. f f + f

k,=l k2=p+l kL=p+l kz=l

(5.5) x Ick(t,
6) sin(rck,x,) sin(rtk,x,)l

IV(,jOUe~n2k2ddkjme~~2k2”dk
The Fourier series for K is given by <8
1 0 1
K(x, x’, t - t’) <81VI, &;*;j
-- = o(.q (5.13)
[

if n2p26> 1. The final bound is derived by observing that

x sin(k,rrx*) sin(k,rrx,) sin(k,rc~~) sin(k,zx;), (5.6)
e (5.14)
where lkl* = k: + kz. Now 6 < t - t’ by construction, and
therefore we may bound the summands in Eq. (5.6) by if z2p26 >, 1. This error bound corrects an erroneous state-
e-X2’ki2g.Thus, the Fourier series converges quickly for ment in [ 121.
6 > 0 and can be used to provide a good approximation to The bound given in Eq. (5.14) shows that the series con-
the history part. Inserting Eq. (5.6) into the history part verges exponentially fast as (p + l)* 6 -+ co, even if 6 + 0.
Eq. (5.5) gives Thus the Fourier series truncation error requires that
dp*+mas6+Oandp+co.
At this point, it is not obvious that evaluating the series
given in Eq. (5.9) is any less expensive than the original
x sin(k,rcx,) sin(k2rcx2) sin(k,nx;) O(N*) cost of history-dependence. This is because at each
x sin(k,rcx;) V(x, t’) dx’ dt’. (5.7) time t every coefficient ck(t, 6) must be computed by
integrating from 0 to t - 6. Fortunately, the coefficients can
Interchanging summation and integration gives a Fourier be updated one step at a time by recursion, rather than
series representation recomputed all the way from t = 0. To simplify notation,
define Mk(t’) by
s, V(x,t) (5.8)
Mk(f’)= sin(k, nx;)(k2nx;) V(x’, t’) dx’.
j (5.15)
= f f ck(t,6) sin(rck,x,) sin(zk,x,) (5.9) r(f)
k,=l kz=l Then we have the recursion relation
with COefiiCientS ck(t, 6) given by
“-“) Mk(t’) dt’
ck(t, S) =4 Jb’-” lrI,,) e-n21k12(f-r’)sin(rck,x;)
+4 j’-” ,-n21k12(t-r’) Mu dt’
x sin(nk,x;) V(x’, t’) dx’ dt’. (5.10) I--6-Al
t-At-6
n2/kf2&
Here, x = (x,, x2), and x’ = (xi, x;). For 6 > 0, this series =4e- e-~21W2(~-~‘) M,(f) dt’
s0
converges rapidly. We first note that since 6 < t - t’, the
coefficients C,(t, 6) decay like Gaussian functions of IkJ, +4 j’-” e-~21k12(~-~‘)Mk(f) dt’
r-S-Al
ck(t, 6)<4e-“2’k’2S [VII, (5.11)
=e -rr2’k’2ArC,(t - At, 6)
where ,-.*tk12(t-f) Mk(f) dt’.
+4 j’-” (5.16)
f--b-Al
) V/)1= j;-’ jrct,) 1I/(x’, t’)l dx’ dt’. (5.12)
Thus ck( t, 6) can be computed from ck( t - At, 6), given the
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 245

values of the moment Mk(l’) over the time step. Since this Since t- t’ < 6, the sum in Eq. (5.19) converges very
requires constant work per time step, we can evaluate rapidly. Suppose we truncate Eq. (5.19) after one term. The
Ck(n At, 6) for n = 1, 2, ... in 0( NW,) work, where W, is the error in this approximation depends on the distance d of the
work involved in computing the moment integrals Mk(f’). boundary r( t’) from the box walls, since this measureshow
This eliminates the 0( N *) cost of the history-dependence. close the nearest contributing images can lie. Note that this
To achieve our goal of O(M’) work per time step, we distance d is independent of the numerical parameters. All
choose p = O(M) so we have O(M *) coefficients C, to but the first term (in which k, = k, =O, 6, = o2 = 1) are
evaluate. Then we can evaluate each one by quadrature then bounded by 0((1/6) ewd2@).Then as the numerical
with O(M) quadrature points, doing O(M’ log M) work parameters are refined and 6 --) 0, we can approximate S, I/
altogether, with Rokhhn’s non-equidistant fast Fourier by the corresponding free-spaceheat potential
transform. Since the Fourier seriestruncation error requires
bp*+co as S-PO and p-+co, this means GM’+co so
6 > 0 (dt) where dt is the time step. This means that 6 will
be, in practice, a fixed small number of time steps, and the
storage requirements of our method are therefore small.
This algorithm uses all the points given on r(t) and can x V/(x’, t’) dx’ dt’. (5.20)
therefore be expected to be more robust than the algorithm
given in [ 121. (This is just Kac’s “principle of not feeling the boundary”
After evaluating the coefficients Ck, we must evaluate a [14-j.) Since the Gaussian decays so fast for t’ small, we can
p*-term Fourier series for S, V at the M 2 grid points in E. assumethat S, V(x, t) is mostly determined by (x’, t’) near
The coefficients are zero-padded and an M x M FFT (x, r). Thus, we can perform a Taylor expansion, and, to
applied, for a total of O(M2 log M) work. lowest order, replace V(x’, t’) by V(x, t) and I-(?‘) by its
Finally, we need an accurate quadrature rule to compute tangent plane at (x, f). A brief calculation, carried out in
these integrals over space and time. Recall that we wish to [ 121, shows then that
compute the time integral (Eq. (5.16)) given by
l-6 s, V(x, t) = J5ji V(x, t) + O(cv2) (5.21)
e-dk12(f--r’) M,(f) dt’. (5.17)
4s f-d-Al
for x E T(r). A higher order Taylor expansion can be carried
The exponential can vary with moderate rapidly over the out, if desired, as in [ 121, but (5.21) will be accurate enough
range of integration, when Ikl* is large. On the other hand, for our purposes.
M, is relatively smooth. Thus we interpolate M,Jt’) by a Thus, we may summarize the fast algorithm described in
linear function oft’ passing through the endpoints t - 6 - At this section as
and t - 6. We can evaluate the resulting integral exactly.
Gauss-Legendre quadrature over each line segment is (5.22)
wx, t) = S,(x, t) + S,(x, tX
used for the spatial integration over r(t) required to
evaluate the moments Mkr with typically six points being where
ample. This reducesthe integral over r( t ) to a weighted sum
of sines, to which the non-equidistant fast Fourier transform s, V(x, = &ii V(x, r) + O(S3’2)
t) (5.23)
applies. A more detailed description of this calculation can
be found in Strain [397.
s, V(x, t) = i i: c,(t, S) sin nk,x,
k,=l k,=l
B. The Local Part SL V
x sin zk,x, + O(e-n2p2”/6) (5.24)
Now we derive the approximation to the local part S, V.
Ck(t, 6) = e-rr21k’2ArC,(t - Ar, 6)
From the method of images expression (5.3) for the heat
kernel K in 3, we have I-6
+4 ,-n21k12(I-r’fMk(lt)dt’ (5.25)
qx, x’, t - t’) I f-d-At

= ,f i: c 1 dlQ2
Mk’J sin(k,nx;)(k,rrx;) V(x’, t’) dx’. (5.26)
k,=-x k>=-m u,=&l oz=~l
rw
I
-[(XI -CL Xi -Z&I)‘+ (x2- CT~X~ - 2kt)2J/(4(t - I’)) This decomposition allows us to evaluate S, V in O(M’)
X4A(M)e work per time step at M* points with O(M) quadrature
(5.19) points on T(t) at each time t = n At.
246 SETHIAN AND STRAIN

6. NUMERICAL DETAILS B. Splitting the Velocity Field and Computing the Curvature

In this section, we discuss some of the numerical details As discussedin Section 4.C., we split the speedfunction F
that arise in the course of the algorithm. into two components, namely,

F=F,+FG, (6.2)
A. The “Free” Temperature Field U;
Computing the speed function FG. using Eq. (2.20) where FA is advection and FG depends on the geometry of
requires evaluating the free temperature field UG which is r(t). For crystal growth, the speed function is given in
the solution to Eq. (2.4). To update U;, we use an finite Eq. (2.20), and we let
difference approximation, for example, +l
FA = cu+s,v1 (6.3)
Edn)+Hfi
-1
FG= dn)C. (6.4)
X
qj+,+ u;jpl -I- u;,,,i+
h2
u:pl,,-4u;j
19
(6.1)
E&)+Hfi
We then use Eq. (4.26) to update 4. The curvature term in
Eq. (6.4) is approximated using central differences applied
to the expression for the curvature of the level curve of 4
where h = Ax = Ay. (For easeof notation, we shall write all passing through x, namely,
formulae assuming a uniform grid). The initial conditions
Uz.=” are found by solving the consistency conditions
implied by the original moving boundary problem, which
requires that, at t = 0, the two boundary conditions
(Eqs. (1.2k( 1.3)) are satisfied on r, as well as the boundary Since C is approximated by central differences,we maintain
conditions on the box wall for the initial temperature field consistency by choosing a central difference approximation
(Eq. (1.1)). Thus we must have to IV41 in the last term of Eq. (4.26). Following the
philosophy of Section 4, we could construct a grid-free
uo= -Ec(n)C--Ev(n)V, approximation to the curvature similar to our approxima-
tion to the normal, but have not done so here.
[au/an] = - HV on r,
C. Finding the Front r(t)
and As outlined in Step 4 of Section 3, at each time step, we
must find the level set 4 = 0. We construct a piecewiselinear
ug = ug. approximation to r(t) as follows. Given a cell ij, if max(4i, jr
di+l,j? di,j+l, #i+l,j+l)<O or min(4i,j, di+i,.j~ tii,j+ly

dj+ ,,j+l) >O, then the cell cannot contain r(t), so we

on the box walls. Note that the normal velocity V is also ignore that cell. Otherwise, find the entrance and exit points
unknown at time t = 0. This problem is underdeterminated, where 4 = 0 by linear interpolation; this provides two nodes
since u. is arbitrary away from ZYIt is physically and mathe- or “marker points” on r(t) and thus one of the line
matically reasonable to complete the determination of u.
segmentswhich form our approximation to r(t). The collec-
by imposing the Laplace equation Au, = 0, subject to the tion of all such line segmentsconstitutes our approximation
given boundary conditions on r and 8B. Physically, it is to r(t), which is then stored for future evaluation of the
reasonable to require u. to solve the steady-state heat equa- history integrals M,. The line segments need not be
tion, while mathematically this imposes smoothness and ordered, nor is there need for any further information about
requires u0 to satisfy the maximum principle. This equation the curve. Note that this construction finds O(M) points on
may be solved efficiently by a potential-theoretic method. r(t) on a A4 x M grid, as assumed in the fast evaluation
First, u0 is represented as the sum of a harmonic function in scheme. Note also that it generalizes, like the rest of our
B with the right boundary values on aB and a classical method, to the three-dimensional case.
single layer potential with density Vspread over r, resulting
in an integral equation for V on IY Once this integral equa- D. Differing Time Steps
tion is solved for V, u. may be evaluated everywhere by an
appropriate fast algorithm. We refer the reader to [39] for The stability requirement for the explicit method for
the details and efficiencies involved. solving the our level set equation (4.26) is At = U(Ax2), due
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 241

to the parabolic-like curvature dependence. This forces perform a study of the relative influence of the physical
a small time step for line grids. However, the boundary parameters, analyzing the effect of the surface tension coef-
propagation algorithm is relatively inexpensive compared ficient sC, the kinetic coefficient E“, coefficient of anisotropy
to the evaluation of the heat potential required in A, crystalline symmetry kA, latent heat of solidification H,
computing F;. Becauser(t) moves little during each time and initial seedon the evolving crystal boundary.
step d t, the change in F is small. More precisely, there is no 1. Smooth growth. We begin with a calculation per-
stability requirement for the velocity evaluation. Thus, we formed with surface tension coefficient sC= 0.001, kinetic
provide the option of employing two time steps,d tFront and
coefficient &y = 0.01, no crystalline anisotropy (coefficient of
At Vel, with k A tFronf = A beI. That is, we take k steps of size
anisotropy A = 0), and latent heat of solidification H = 1.
At FK7nt in the level set algorithm before we recompute the
The calculation were performed in a unit box, with a
velocity field FZ and update the stored boundary by time
constant undercooling on the side walls of us = - 1.
step Atve,. This option both speeds up the computations,
The initial shape was a perturbed circle with average
since it allows us to increase 6, keep fewer Fourier series
radius R = 0.15, perturbation size P = 0.08, and L = 4 limbs.
terms, and requires less storage, since fewer past time steps
That is, the parametrized curve (x(s), y(s)), 0 bs d 1,
need be stored. describing the initial position of the crystal is given by

(x(s), y(s)) = [(R + P cos 27cLs)](cos 27q sin 27~s) (7.1)

7. NUMERICAL RESULTS

In this section, we show a series of calculations with the with R=0.15, P=O.O8, and L=4.
numerical method presented in this paper. In Fig. 10, we show a series of calculations performed to
study the effect of relining the grid size and time step on the
A. Input Parameters computed solution. We begin in Fig. 10a with mesh size
48 x48, and time step At =0.005. The boundary grows
The numerical input parameters are smoothly outward from the initial limbs. Each limb is
h = grid size in the unit box drawn towards its wall by the effectof the undercooling. The
value of the kinetic coefficient Ey is large enough to keep the
At = time step
evolving boundary smooth (in contrast to those calcula-
6 = size of history cutoff ( 30 (At) for consistency)
p = number of Fourier modes (p >/ 0 (At -I’*) for con-
sistency) 3- t
k = number of times boundary moves per velocity evalua-
tion.
The physical input parameters are
EC,eV= anisotropy constants
k, = crystalline symmetry
8, = angle of crystalline symmetry
u0 = initial values of temperature field
us = boundary values of temperature field on box walls
Size, shape, and position of initial seed.

B. Results
In this section, we present the results of our numerical
calculations using our solidification algorithm. We begin
with a seriesof calculations to check numerical consistency;
that is, to verify that the computed solution convergesas the
numerical parameters are relined. (A further verification
comparing the computed solution using the level set FIG. 10. Smooth crystal: Effect of relining both grid size and time step:
approach with an accurate boundary integral calculation in (a)48x48mesh,d?=0.005,H=1.0,A=0.0,s,=0.01,~,=0.01,k,=0;
(b)64x64 mesh, dt=0.0025, H=l.O, A=O.O, ~,=0.001, ~,,=0.01,
their common range of validity is in progress, see [42].) k,=O; (c)96x96 mesh, dt=O.OOl, H=l.O, A=O.O, ~~=0.001,
After demonstrating the robustness of the algorithm, we &y=O.OO1,k~=O.
248 SETHIAN AND STRAIN

tions discussed below). In Fig. lob we repeat the above First, each limb flattens out. Then, tip splitting occurs as
calculation with mesh size 64 x 64 and At =0.0025. In spikes develop from each limb. Finally, side-branching
Fig. lOc, we take mesh size 96 x 96 and At =O.OOl. The begins as each multi-tipped arm is both pulled toward the
figures are unchanged, indicating the robustness of the closest wall and also drawn by the walls parallel to the limb.
algorithm. In Fig. 1lb, we perform the same calculation, but decrease
Although the shape remains smooth as it evolves, several the time step to At = 0.0025, maintaining the 48 x 48 mesh.
effects can be seen. First, while the undercooling pulls the In Fig. 1Ic, we perform the same calculation, but decrease
limbs towards the walls, the other walls act to thicken the the time step to At = 0.001, maintaining the 48 x 48 mesh.
limb, creating a highly smoothed version of side-branching. The position of the evolving boundary is unchanged under
Second, the tips of the limbs, where the curvature is positive, these time refinements, indicating the robustness of the
move very fast compared to the indented pockets between numerical algorithm.
limbs, where the curvature is negative. While the evolving boundary is complex, it should remain
unchanged as both the space and time step are refined.
2. Fingered growth. Next, we perform a similar calcula- In Fig. 12, we study the effects of such refinement, using
tion, but change the kinetic coefficient to Ey = 0.001. Once the same physical parameters, that is, kinetic coefficient
again, we take surface tension coefficient ~~=0.001, no sy= 0.001, surface tension coefficient .sC= 0.001, no crys-
crystalline anisotropy (coefficient of anisotropy A = 0), and talline anisotropy (k, = 0, A = 0), and latent heat of
latent heat of solidification H= 1. The constant under- solidification H= 1, with constant undercooling - 1. In
cooling on the side walls is - 1, and the initial shapeis again Fig. 12a, we take a 32 x 32 mesh with At =0.005. In
a perturbed circle with average radius 0.15, perturbation Fig. 12b, we take a 48 x 48 mesh with At =0.005.
size 0.08, and four limbs, as in Eq. (7.1). In Fig. 12c, we take a 96 x 96 mesh with At = 0.00125. In
In Fig. 11, we show the results using these values for the Fig. 12d, we take a 128 x 128 mesh with At = 0.00125. On
physical parameters. We begin with a set of calculations per- the coarsest mesh (32 x 32), only the gross features of the
formed with decreasing time step and fixed mesh size. In fingering and tip splitting processare seen.As the numerical
Fig. 1la, we show the evolution of r(t) on a 48 x 48 mesh parameters are refined, the basic pattern emerges.It is clear
with time step At=0.005. Starting from the smooth per- that the resulting shapes are qualitatively the same, and
turbed circle, the evolving boundary changes dramatically. there is little quantitative difference between Fig. 12~ and

3- t, a
br

FIG. 12. Fingered crystal: Effect of relining both grid size and time
FIG. 11. Fingered crystal: Effect of refining time step: (a) 48 x48 step: (a) 32 x 32 mesh, LIZ= 0.005, H = 1.O, A = 0.0, cc = 0.001, zy = 0.001,
mesh, dt=O.OOS, H=l.O, A=O.O, ~~=0.001, E~=O.OOI, k,=O; k,=O; (b)48x48 mesh, dt=O.OOS, H=l.O, A=O.O, ~~=0.001,
(b)48x48 mesh, ~lr=0.0025, H=l.O, A=O.O, ~,=0.001, ~,=0.001, ~,=0.001, k,=O; (c)96x96 mesh, dl=O.O0125, H=l.O, A=O.O,
k, =O; (c)48 x48 mesh, dr =O.OOl, H= 1.0, A=O.O, ~c=0.001, ~~=0.001, ~~=0.001, k,=O; (d) 128 x 128 mesh, dt=0.00125, H= 1.0,
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 249

12d. Even when computing the highly complex boundaries a ,r

seenin these figures, the algorithm remains robust.
3. Latent heat of solidification. Next, we study the effect
of changing the latent heat of solidification H. Recall that H
controls the balance between the pure geometric effectsand
the solution of the history-dependent heat integral. We con-
sider the same initial shape and physical parameters as in
Fig. 12 (again, sv=O.OO1, sC=O.OO1,kA=O, A=O, with
constant undercooling - 1). In all calculations, we use a
96 x 96 mesh with time step At = 0.00125.However, we vary
H smoothly: in Fig. 13a, H = 0.25, in Fig. 13b, H = 0.5, in
Fig. 13c, H = 0.75, and in Fig. 13d, H = 1.0. (
The evolving boundaries are all plotted at the sametime,
and the differences are substantial. With H =0.25
(Fig. 13a), the dominance of geometric motion serves to
create a rapidly evolving boundary that is mostly smooth.
As the latent heat of solidification is increased, the growing
limbs expand outwards less smoothly, and instead develop
flat ends. As seenin Fig. 12, these flat ends are unstable and
serve as precursors to tip splitting. We also note that the
influence of the heat integral slows down the evolving
boundary, as witnessed by the fact that all the plots are
FIG. 14. Small-scale refinement of latent heat of solidification H:
given at the same time. The mechanism operating here is (a)H=O.75,A=0.0,~~=0.001,~,=0.00l,k,=0,96~96,At=0.00125;
presumably that increasing latent heat decreasesthe most (b)H=0.833,A=0.0,~~=0.001,~~=0.~1,k,=0,96x96,Ar=0.~125;
unstable wavelength, as described by linear stability (c) H= 0.916, A = 0.0, tC= 0.001, E,, =o.OOl, k,, = 0,96 x 96, At = 0.00125;
theory [ 191. (d) H = 1.0, A = 0.0, eC = 0.001, cy = 0.00.1, k, = 0, 96 x 96, At = 0.00125.

To further understand the transition from smooth

b
crystals to complex ones, in Fig. 14 we perform a small-
scale refinement of the latent heat of solidification for
values H=0.75 (Fig. 14a), H=0.833 (Fig. 14b), H=0.916
(Fig. 14c), and H= 1.0 (Fig. 14d). In this set of figures, the
evolving boundaries are not given at the sametime. Instead,
the final shape is plotted when r(t) has reached to within
0.02 of the box walls (recall that the box has width 1.0). In
these figures, we can see how lingering and tip splitting
emergefrom the essentially smooth shape given in Fig. 14a
to the complex shape of Fig. 14d. - -

cr C
4. Coefficient of anisotropy. Next, we study the effect of
changing the coefficient of anisotropy. We again start with
the basic crystal shape given in (Eq. (7.1)), with the same
choices for radius and perturbation. We take Ey = 0.001,
sC= 0.001, H = 1.0, a 96 x 96 mesh with time step
At =0.00125 and constant undercooling - 1. In these
calculations, we assume a fourfold anisotropy (k, = 4). In
Fig. 15a, we take the coefficient of anisotropy A = 0.2, in
Fig. 15b, A =0.4, in Fig. 15c, A =0.6, and in Fig. 15b,
A = 0.08. Our results show that increasing the coefficient of
FIG. 13. Effect of smoothly varying latent heat of solidification H:
(a)H=0.25, A=0.0,~~=0.001,e~=0.001, k,=0,96x96, At=0.00125;
anisotropy has several effects.First, it increasesthe speedof
(b)H=0.50, A=0.0,~~=0.001,~~=0.001, k,=0,96x96, At=0.00125; propagating interface. Second, it controls the shape of the
(~)H=0.75,A=0.0,~~=0.001,~~=0.001,k~=0,96x96,Af=0.00125; evolving linger from each limb. Third, it influences the
(d)H=1.0,A=0.0,~~=0.001,~~=0.001,k,=0,96x96,At=0.00125. amount of side branching that occurs.
 SETHIAN AND STRAlN

FIG. 15. Effect of changing coefficient of anisotropy A: (a) A = 0.2, FIG. 17. Changing initial perturbation, anisotropy coefficient A, and
~~=0.001, ~~=OXlOl,k~=4, H=l.O, 96x96, At=0.00125; (b)A=0.4, crystalline symmetry k, : (a) L = 6, A = 0.0, k, = 6, cc = 0.001,Ey = 0.001,
so=O.OOl, ~~=0.001, k,,=4, H=l.O, 96x96, At=0.00125; (c)A=0.6, H=l.O, 96x96, At=0.00125; (b)L=S, A=O.O, k,=8, ~~=0.001,
&o=O.OOl, ~~=0.001, k,=4, H=l.O, 96x96, Ar=O.O0125; (d)A=O.S, e,,=O.OOl, H=l.O, 96x96, At=0.00125; (c)L=6, A=0.4, k,=6,
eC=0.001,cv=0.001,k,,=4,H=1.0,96x96.At=0.00125. ~~=0.001, s,,=O.OOl, H=l.O, 96x96, At=0.00125; (d)L=S, A=0.4,
k,=8,eC=0.001reY=0.001,H=1.0,96x96,Ar=0.00125.

b r-
5. Changing initial conditions, anisotropy coefficient A,
and crystalline symmetry k, . In Fig. 16 and 17, we com-
pute a collection of shapes by altering the initial seed,
anisotropy coefficient, and degree of crystalline anisotropy.
All calculations are performed using a 96 x 96 grid, with
time step At = 0.00125, and E,= 0.001, sC= 0.001, H= 1.0,
and constant undercooling - 1.
In Fig. 16, we concentrate on changing the anisotropy
i coefficient and the crystalline anisotropy. In Fig. 16a, we
take an initial shape with four limbs, fourfold crystalline
anisotropy, and A = 0.4. In Fig. 16b, we take an initial shape
with four limbs, sixfold crystalline anisotropy, and A = 0.4.
In Fig. 16c,we take an initial shapewith four limbs, fourfold
crystalline anisotropy, and A = 0.8. In Fig. 16b, we take an
initial shape with four limbs, sixfold crystalline anisotropy,
and A = 0.8. The results depend dramatically on the choice
of these values.
In Fig. 17, we concentrate on changing the anisotropy
I J
FIG. 16. Changing initial perturbation, anisotropy coefficient A, and
coefficient and the initial shape of the crystal. In Fig. 17a, we
take an initial shape with six limbs, sixfold crystalline
anisotropy, and A = 0. In Fig. 17b, we take an initial shape
crystalline symmetry k,: (a) L=4, A=0.4, k, =4, ~~=0.001, .sy=O.OO1, with eight limbs, eightfold crystalline anisotropy, and A = 0.
H=l.O, 96x96, At=0.00125; (b)L=4, A=0.4, k,=6, e,=O.OOl,
ey=O.OO1, H=l.O, 96x96, Ar=O.O0125; (c)L=4, A=O.S, k,=4,
In Fig. 17c, we take an initial shape with six limbs, sixfold
~~=0.001, ey=O.OO1, H=l.O, 96x96, Ar=0.00125; (d)L=6, A=O.S, crystalline anisotropy, and A = 0.4. In Fig. 17d, we take
k,=6,sC=0.001,ev=0.001,H=1.0,96x96,At=0.00125. an initial shape with eight limbs, eightfold crystalline
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 251

a b

FIG. 18. Array of crystals: H = 0, purely geometric motion: (a) Time = 0.&0.025,96 x 96 mesh,At = 0.00125,H= 0.0, A = 0.4, cc = 0.001,ey = 0.001,
k, =4; (b) Time=0.025-0.05, 96 x96 mesh, At =0.00125, H=O.O, A=0.4, ~~=0.001, e,,=O.OOl, k, =4; (c) Time=0.05-1.0, 96 x96 mesh,
At=0.00125, H=0.0,A=0.4,&,=0.001,~,=0.001,k,=4

anisotropy, and A = 0.4. The resulting shapes display a determined by purely geometric effects, the presence of
variety of intricate behavior. neighboring crystals is not known until collision occurs.
Once neighboring boundaries meet, they all join together,
6. Arrays of Crystals. To demonstrate the versatility of leaving only a collection of pockets which soon collapse.
our method, in Fig. 18 and 19 we consider an array of The complexity of this problem underscores the ability of
crystals. In both cases, we consider a 96 x 96 mesh, with the level set approach to track involved topological changes.
time step At =0.00125. In Fig. 18, we show the case of In Fig. 19, we perform a similar calculation, but this time
purely geometric motion (H = 0, and motion according to we turn on the full effects of the equations of motion.
Eq. (4.1)) with sC= 0.01, E~=O.OO.We consider a 4 x 4 We take H= 0.75, sC= 0.001, .sy= 0.001, with fourfold
array of crystals, and take fourfold anisotropy, with anisotropy and coefficient of anisotropy A = 0.4. Again, we
anisotropy coefficient A = 0.4. We plot the figure at various start with a 4 x 4 array of initially circular crystals. In this
times, as shown in Figs. 18a, 18b, and 18~.In each of these case, the evolving crystals are quite different. Neighboring
figures, we show several time levels of the moving boundary, crystals “hear” each other due to the temperature boundary
as well as the solid crystal at a early time, depicted by conditions on each interface. Thus, we seethat the exterior
shading in those points for which 4 > 0. edge of crystals grow towards the walls, under the influence
To begin with, the crystals evolve under the influence of of the undercooling and the anisotropic effects. However,
the fourfold anisotropy. Since the motion of each crystal is growth towards the center, and in particular for the interior
 SETHIAN AND STRAIN

the numerical parameters produces a converged solution,

even for the most complex interfaces. In [35], we use our
algorithm to generate crystal boundaries under a wide
range of conditions. Finally, the method presented here
generalizesin an obvious manner to three spacedimensions.
We shall report on that work elsewhere [36].

ACKNOWLEDGMENTS

We thank OIe Hald and Alexandre Chorin. All calculations were

performed at the University of California, Berkeley, the Lawrence Berkeley
Laboratory, and the Courant Institute of Mathematical Sciences.

REFERENCES

1. M. Ben Amar and Y. Pomeau, Europhys. Left. 2, 3307 (1986).

2. L. N. Brush and R. F. Sekerka. J. Cry.rt. Growrh, to appear.
3. G. Caginalp and P. C. Fife, SIAM J. Appl. Marh. 48, 506 (1988).
4. J. W. Cahn and J. E. Hilliard, .I Chem. Phys. 28,258 (1958).
5. J. Chadam and P. Ortoleva, IMA J. Appl. Math. 30, 57 (1983).
FIG. 19. Array of crystals: H=0.75, full effects: (a) Time =O.OOZS,
6. B. Chalmers, Principles of Solidijication (Wiley, New York, 1964).
48x48 mesh, At=0.005, H=0.75,,4=0.4,~~=0.001,~~=0.001,k~=4;
7. A. J. Chorin, J. Comput. Phys. 57,472 (1985).
(b) Time = 0.005, 48 x 48 mesh, At = 0.005, H = 0.75, A = 0.4, ec = 0.001,
ey=O.OOl, k,=0.4; (c)Time=O.Ol, 48x48 mesh, At =0.005, H=0.75, 8. J. Duchon and R. Robert, Evofution d’une Interface par Capillarite et,
A =0.4, eo=O.OOl, e,,=O.OOl, k, =4; (d) Time =0.015, 48 x48 mesh, Diffusion de Volume I. Existence Locale en Temps, Ann. Inst. Henri
Af=O.O05, H=0.75, A=0.4,ec=0.001,e.=0.001,k,=4. Poincure 1,361 (1984).
9. H. Dym and H. P. McKean, Fourier Series and Integrals (Academic
Press, New York, 1972).
10. L. C. Evans and J. Spruck, J. DQ” Geom. 33,635 (1991).
crystals, is severely limited, as the release of latent heat by
11. G. J. Fix, “Numerical simulation of free boundary problems using,
the neighbors warms the liquid. phase field methods,” in The Mathematics of Finite Elements
The calculations presented above are all performed in a and Appkafions IV (Oxbridge, 19&I), p. 265 (Academic Press,
square container. The influence of the box walls is signili- London/New York, 1982).
cant on the growing front, and it would be interesting to 12. L. Greengard and J. Strain, Commun. Pure Appl. Math 43, 949 (1990).
extend our work to other bounding containers, as well as 13. M. E. Gurtin, Arch. Ration Mech. Anal. %, 199 (1986).
infinite domains. The extension to other bounding regions is 14. M. Kac, Am. Math. Mon. 73, No. 4, 1 (1966).
not overly difficult; our focus in this paper has been to 15. A. Karma, Phys. Rev. Left. 57, 858 (1986).
explain and test the fundamental algorithm. The extension 16. F. X. Kelly and L. H. Ungar, Phys. Rev. B34, 1746 (1986).
of our approach to problems in infinite domains, while not 17. D. A. Kessler and H. Levine, Phys. Rev. Leif. 57, 3069 (1986).
as straightforward, would be valuable. 18. J. S. Langer, Phys. Rev. A, 33, p. 435441, 1986.
19. J. S. Langer, Rev. Mod. Phys. 52, (1980).
8. CONCLUSION 20. P. D. Lax, Commun. Pure Appl. Math. IO, 537 (1957).
21. V. Laxmanan, Acta Metall. 33, 1023 (1985).
In this paper, we have presented a numerical algorithm 22. D. I. Meiron, Physica D 23, 329 (1986).
for computing the motion of complex solid/liquid interfaces 23. G. H. Meyer, SIAM J. Numer. Anal. 10, 552 (1973).
in crystal growth. Our model includes such physical effects 24. W. Mulder, S. J. Osher, and J. A. Sethian, J. Compuf. Phys., to appear.
as crystalline anisotropy, surface tension, molecular 25. W. W. Mullins and R. F. Sekerka, Appl. Ph.w. 34, 323 (1963).
kinetics, and undercooling. Our algorithm relies on a 26. W. W. Mullins and R. F. Sekerka, J. Appl. Phys. 35,444 (1964).
boundary integral formulation of the equations of motion 27. G. E. Nash and M. E. Glicksman, Acta Metall. 22, (1974).
coupled to a level set method for advancing the propagating 28. G. E. Nash and M. E. Glicksman, Acto Metall. 22, 1291 (1974).
interface. The method is able to follow the evolution of 29. S. J. Osher, SIAM J. Numer. Anal. 21, 27 (1984).
extremely intricate shapes, exhibiting complex behavior 30. S. J. Osher and J. A. Sethian, J. Cornput. Phys. 79, 12 (1988).
such as lingering, tip splitting, and side branching, as well as 31. V. K. Pinus and P. L. Taylor, Phys. Rev. B 32, 5362 (1985).
profound topological changes. In addition, our numerical 32. J. A. Sethian, Commun. Math. Phys. 101, 4 (1985).
techniaue is shown to be highlv robust. in that refinement of 33. J. A. Sethian. J. Diff Geom. 31. 131 (1989).
 CRYSTAL GROWTH AND DENDRITIC SOLIDIFICATION 253

34. J. A. Sethian and D. L. Chopp, “Singularity Formation and Minimal 39. J. Strain, Physica D 30, 297 (1988).
Surfaces,” J. Dzj$ Atom., to appear. 40. J. Strain, SIAM J. Appl. Math. 50, 1 (1990).
34. J. A. Sethian and J. Strain, “Numerical Simulation of Complex Crystal 41. J. Strain, “A Fast Boundary Integral Method for Crystal Growth”
Growth,” J. Off: Geom., to appear. (unpublished).
35. J. A. Sethian and J. Strain, “Three-Dimensional Crystal Growth” 42. J. M. Sullivan, D. R. Lynch, and K. 0. O’Neill, J. Compur. Phys. 69,81
(unpublished). (1987).
36. J. B. Smith, J. Comput. Phys. 39, 112 (1981). 43. C. Verdi and A. Visintin, Boll. U.M.I. (7) I-B, 795 (1987).
37. J. Strain, J. Comput. Phys. 85, 342 (1989). 44. D. J. Wollkind and R. D. Notestine, IMA J. Appl. Math. 27, 85 ( 1981).
38. J. Strain, Berkeley Center for Pure and Applied Mathematics Report,
PAM-494, April 1990; J. Cornput. Phys., submitted.