You are on page 1of 27

A LEVEL SET BASED

FINITE ELEMENT ALGORITHM


FOR THE SIMULATION
OF DENDRITIC GROWTH

MICHAEL FRIED

Abstract. Dendritic growth is a nonlinear process, which falls into the cat-
egory of self-organizing pattern formation phenomena. It is of great practical
importance, since it appears frequently and, in the case of alloys, affects the
engineering properties of the resulting solid. We describe a new finite element
algorithm for the two–dimensional Stefan problem, where the free boundary
is represented as a level set. This allows to handle topological changes of the
free boundary. The accuracy of the method is verified and several numerical
simulations, including topological changes of the free boundary, are presented.

1. Introduction

Dendritic growth is an important specimen of self–organizing pattern formation


phenomena. Dendrites which arise during the solidification of a metallic melt affect
engineering properties such as strength and durability of the later solid. Of less
practical interest, but of fascinating beauty are the hexagonal structured and often
intricated patterns of snowflakes, the dendrites of water.
Here we consider the solidification of a pure, undercooled liquid. We base on
a well–known generalization of the classical Stefan problem, cf. [2, 22, 24, 26, 39].
In order to fix notations, we recall the two–dimensional case of this model. Let
Ω ⊂ R2 be a region filled with a pure substance, which may be in solid or liquid
phase. The solid and liquid regions are then denoted by ω s and ω l , respectively.
Together with the one–dimensional phase boundary γ they give a time-dependent
partition p = (ω s , γ, ω l ) of the region Ω. This phase distribution p = p(t) and the
dimensionless temperature θ = θ(x, t) satisfy the following evolution laws:
θt − ∆θ = 0 in ω s (t) ∪ ω l (t),
 l
∂θ
v=− on γ(t), (1.1)
∂n s
β(n) v + a(n) k + θ = 0 on γ(t),

where n is the unit normal to γ into the liquid region ω l , v is the normal velocity
l
of γ and k is the curvature of γ with a positive sign when ω s is convex. By [[ · ]]s we

Research supported by the DFG Schwerpunkt “Dynamik: Analysis, effiziente Simulation und
Ergodentheorie”, Project: Effiziente Simulation und Numerische Analysis der Dynamik von Den-
driten (G. Dziuk and A. Schmidt).
1
2 MICHAEL FRIED

denote the jump across the phase boundary γ with sign convention given by “liquid”
− “solid”. The functions β and a denote material quantities: the kinetic coefficient
β is defined on the unit sphere S, the interface potential a can be calculated from
a smooth surface tension φ = φ(n) by the formula a(n) = D2 φ(n) n⊥ · n⊥ , where
D2 φ(n) is the Hessian of the positively homogeneous extension of φ on R2 and ⊥
denotes the counterclockwise rotation by 90◦ . We assume that the temperature θ is
scaled such that θ = 0 is the melting temperature of the regarded substance, that
is the temperature of a planar phase boundary in equilibrium.
In contrary to the classical Stefan problem, the above model includes a gen-
eralized Gibbs–Thomson law with undercooling, which allows for a shift in the
temperature θ at the free boundary due to kinetic and curvature effects. Since
the coefficients β and a depend on the normal n, the generalized Gibbs–Thomson
equation also includes the modeling of directional anisotropies. In the sequel we
suppose that the material quantities fulfill

β∗ := inf β(n) > 0 and a∗ := inf a(n) > 0. (1.2)


n∈S n∈S

In the present article we consider the numerical solution of the two–dimensional


modified Stefan problem (1.1) using a finite element method. The most difficulties
for the numerical treatment of Problem (1.1) arise from the generalized Gibbs–
Thomson law. In view of this equation, the evolution of the free boundary γ(t)
may be interpreted as an anisotropic mean curvature flow with the temperature θ
as a “driving force”. Therefore a common starting point to develop a numerical
scheme for the dendritic growth consists of an algorithm for the mean curvature
flow problem. Once such an algorithm is given, and if, on the other hand, we
have a method to calculate the temperature θ provided the phase distribution p(t)
is known, we may combine those methods to achieve a numerical scheme for the
whole Stefan problem.
We chose a numerical method for the mean curvature flow part of Problem
(1.1) which uses a level set representation of the phase boundary γ(t) as it was
first proposed as a computational procedure by Osher and Sethian for problems
involving the curvature dependent evolution of an interface, cf. [28]. The basic idea
behind such a level set method is to interpret the interface γ as the zero level of some
continuous real valued function u. This enables us to handle topological changes of
the phase boundary as they will occur for example if different parts of a dendrite are
growing towards each other finally touching one another. In case of mean curvature
flow like problems, the level set approach leads to a degenerate nonlinear partial
differential equation for the function u, which has to be understood in a viscosity
sense, cf. [3], [15], [16], [17], [18], [21].
A broad variety of algorithms has been developed to numerically approximate
the curvature dependent evolution of a hypersurface. A description of such methods
could be found in [13] and the references therein.
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 3

In Section 2 we outline a level set based finite element algorithm for the general-
ized mean curvature flow problem as it is given by the generalized Gibbs–Thomson
law. Here we suppose for the moment the temperature θ to be known. Using a
regularized formulation of the anisotropic and inhomogeneous level set problem,
we combine time discretization by a semi implicit Euler method with the standard
procedure to get a finite element method for the function u, and such for the gen-
eralized evolution of the free interface γ, which is defined by the zero level of u.
Some numerical results are presented to illustrate the applicability of the algorithm
for the anisotropic mean curvature flow of level sets. While an explicit calculation
of the curvature of the level sets is not needed, in case of several applications it
is desirable to know the curvature of the moving interface γ. Particularly with
regard to our algorithm for the Stefan Problem (1.1), where we have to know the
curvature explicitely, we end Section 2 by short describing a numerical method to
approximate the mean curvature of a given finite element function’s level sets.
Vice versa we assume in Section 3 the phase boundary γ(t), its normal velocity v
and its curvature k to be known. Then a numerical method for approximating the
second unknown, the temperature, is obtained by a similar finite element method
as it was introduced by Schmidt in [31] and [32]. The difference lies in the fact that
Schmidt used a parametric ansatz for the representation of the phase boundary,
while we represent γ using a level set method.
To obtain a numerical method for the whole modified Stefan problem we combine
in Section 4 the finite element methods for the free boundary and for the temper-
ature in the following manner: for each timestep tm , we first calculate the new
finite element approximation U m of the (regularized) level set function u, using the
discrete temperature Θm−1 from the previous timestep. The zero level of U m gives
us the new discrete phase boundary γ m . With this, we apply the above mentioned
finite element algorithm for the temperature, to compute the update of the discrete
temperature Θm . Since this algorithm makes use of the curvature k of γ m , we have
to calculate k. This is done by employing the finite element approximation K of
the curvature of level sets of U m , which is defined in Section 2.3. is depicted in
Section 4.1. Finally, the computation of the temperature Θm finishes timestep tm .
We end by presenting various numerical results obtained by our method. They
include a verification of its accuracy with the help of an example with explicitly
known exact solution and an example with topological changes of the phase bound-
ary.

2. A Level Set Formulation for the Mean Curvature Flow

In this section we suppose the temperature θ to be given. Then the last equa-
tion in Problem (1.1), the generalized Gibbs–Thomson law, leads to the following
generalized mean curvature flow for the evolution of the phase boundary γ(t):
4 MICHAEL FRIED

Problem 2.1. Find a family of interfaces {γ(t)}t∈[0,T ) such that

β(n)v + a(n)k = −θ on γ(t), t ∈ (0, ∞), (2.1)

where γ(0) = γ0 .

The purpose of this section is to develop a finite element algorithm for the
anisotropic and inhomogeneous mean curvature flow given by the above problem,
which can be extended to an algorithm for the Stefan Problem (1.1). The latter
will be done in Section 4.
As mentioned in the introduction, we consider a level set approach to the mean
curvature flow problem 2.1, assuming the free boundary γ to be represented by the
zero level of some continuous real valued function u. In addition we assume that all
level sets of u evolve according to equation (2.1). A formal calculation then leads
to the following non–linear degenerate and singular parabolic partial differential
equation for the function u, compare [15]:
   
∇u ut ∇u ∇u
β −a ∇· = θ. (2.2)
|∇u| |∇u| |∇u| |∇u|
Conversely, if u is a viscosity solution of (2.2) in R2 , then each level set of u evolves
in a weak sense as it is required by (2.1). In a series of publications ([15], [16], [17],
[18]), Evans and Spruck investigated problem (2.2) for β = a ≡ 0 and θ ≡ 0 together
with an initial condition u(·, t) = u0 in arbitrary spatial dimensions ≥ 2, and proved
existence and uniqueness of a viscosity solution u of this problem. Here, the initial
function u0 is an arbitrary continuous function whose zero level represents the
initially given hypersurface γ0 . By Chen/Giga/Goto and Giga/Goto/Ishii similar
results were obtained, even for anisotropic and inhomogeneous situations, cf. [3],
[21]. In view of Problem 2.1, the above authors showed that not only the viscosity
solution u is uniquely determined by (2.2) with the above initial condition, but also
the generalized evolution of γ, defined by {x ∈ R2 |u(x, t) = 0}, does only depend
on the initial hypersurface γ0 , and therefore is independent of the special choice of
the initial value u0 . These results allow to reformulate Problem 2.1 in a level set
context:

Problem 2.2. Given some function u0 with

γ(0) = {x ∈ R2 | u0 (x) = 0},

find the continuous function u : R2 × [0, T ) −→ R such that


   
∇u ut ∇u ∇u
β −a ∇· = θ,
|∇u| |∇u| |∇u| |∇u|
with initial condition

u(·, 0) = u0 ∈ R2 .

There are various methods to numerically solve Problem 2.2, see for example [5],
[28], [34], [40].
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 5

One of the main difficulties in discretizing this problem arises if the gradient
∇u vanishes. To avoid such difficulties, we turn to the following regularization of
Problem 2.2:

Problem 2.3. For fixed ε > 0, find uε ∈ C 0 (R2 × [0, ∞)) such that
uεt ∇uε
βε (∇uε ) − aε (∇uε ) ∇ · = θ in R2 × (0, ∞), (2.3)
Q(uε ) Q(uε )
with initial condition

uε (·, 0) = u0 ∈ R2 ,

where Q(uε ) is given by


p
Q(uε ) = ε2 + |∇uε |2 (2.4)

The functions βε and aε denote smooth extensions βε , aε ∈ C ∞ (R2 ) of the given


anisotropies β and a, respectively:

p

 β( |p| ),
 |p| ≥ ε
p ε2 p
βε (p) := e [β0 − β( |p| )] exp(− ε2 −|p| 2 ) + β( |p| ), 0 < |p| < ε (2.5)

|p| = 0,

β0 ,
with β0 = 12 (minS β + maxS β), and the analogous definition for aε . In the
homogeneous situation, i.e. θ = 0, we have

Theorem 2.1. Suppose β, a ∈ C ∞ (S 1 ) fulfilling (1.2) and u0 ∈ Cη (R2 ) ∩ C ∞ (R2 )


for some η ∈ R are given. Suppose θ ≡ 0.
Then for every 0 < ε < 1 there exists a unique bounded solution uε of Problem
2.3 with

sup kuε , ∇uε , uε,t kL∞ (R2 ×[0,∞)) ≤ C, (2.6)


0<ε<1

with a positive constant C which only depends on u0 .

In case of the isotropic mean curvature flow of level sets, this was proved by
Evans and Spruck, cf. [15]. Their proof directly carries over to the above anisotropic
situation. Furthermore, Evans and Spruck showed that for ε −→ 0, the solution
uε converges locally uniformly to the viscosity solution of Problem 2.2. Of special
interest is a result of Deckelnick, who proved the estimate

sup |(u − uε )(x, t)| ≤ Cεα ,


x∈R2 ,0≤t≤T

for all ε > 0, all α ∈ (0, 12 ) and a constant C which only depends on u0 , T and α.
(cf. [6]). Hereafter, we intend to approximate the viscosity solution u by numerically
solving the regularized equation (2.3) for uε . Traditionally, the mean curvature flow
of level sets is solved in the whole space and for all times. However, for practical
reasons it is necessary to restrict ourself to a bounded convex domain Ω and a finite
time interval [0, T ). Hence boundary conditions at the boundary ∂Ω are required.
In the homogeneous situation with compact initial surface, the choice of boundary
6 MICHAEL FRIED

conditions on ∂Ω won’t affect the evolution of the interface as long as it remains


strictly inside of the domain Ω: If u0 ∈ C 0 (R2 ) is chosen such that u0 is constant
outside of the domain Ω, then also the viscosity solution u remains constant outside
of Ω, cf. [3], [15].
In the inhomogeneous case, the free interface γ may grow and touch the boundary
of the domain even if we start with a convex initial surface γ0 . Specifying Dirichlet
data then corresponds to prescribing the location at which the interface touches the
boundary ∂Ω, while specifying Neumann data means that we fix the angle between
γ and the boundary at the common points.
We consider Dirichlet as well as Neumann boundary conditions, our regularized
mean curvature flow problem on a bounded domain Ω then is given by:

Problem 2.4. For fixed ε > 0, find uε ∈ C 0 (Ω × [0, T )) such that


uεt ∇uε
βε (∇uε ) − aε (∇uε ) ∇ · = θ in Ω × (0, T ), (2.7)
Q(uε ) Q(uε )
with initial condition

uε = uε,0 on [Ω × {0}], (2.8)

where uε,0 is some continuous function, and


• Dirichlet boundary condition

uε = uε,0 on ∂Ω × (0, T ), (2.9)

or
• Neumann boundary condition
∂uε
=0 on ∂Ω × (0, T ), (2.10)
∂ν
where ν denotes the normal on ∂Ω.

As the generalized evolution of γ is independent of the choice of the initial


function u0 , we are free to choose any function whose zero level represents the
interface γ0 . However it is convenient to choose u0 such, that

u0 > 0 in ω l (0), u0 < 0 in ω s (0),

where ω s (0) denotes the part of the domain Ω which lies inside of the initial interface
γ0 , and ω l (0) := Ω \ (ω s (0) ∪ γ0 ).

Remark 2.1. For ε = 1 Problem 2.4 describes the situation of the anisotropic mean
curvature flow of a graph. By a scaling argument, we see that a solution uε of
Problem 2.4 leads to a solution of the corresponding problem of a moving graph,
and vice versa. Hence, results on moving graphs may be transfered to the solution
of the regularized level set problem. Various results on existence and uniqueness
of moving graphs on a bounded domain were obtained for instance by Huisken
[25], Lieberman [27] and Veeser [38]. With regard to our algorithm the results of
Deckelnick and Dziuk on convergence for a semi–discrete finite element method
[7], [8] and for a fully discrete numerical scheme [10], and an error estimate and
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 7

stability results for several time discretizations obtained by Dziuk [12] and for are
of particular interest.

2.1. A Finite Element Scheme for the Mean Curvature Flow. Our finite
element method is based on this weak formulation of the regularized problem: find
uε (·, t) ∈ H 1 (Ω) such that
∇uε ∇φ
Z Z Z
βε (∇uε ) u φ θφ
p ε,t + p = ∀φ ∈ H01 (Ω),
aε (∇uε ) ε2 + |∇uε |2 ε2 + |∇uε |2 aε (∇uε )
Ω Ω Ω
(2.11)
with initial conditions (2.8) and Dirichlet boundary data.
In order to state the discrete problem, we define the finite element spaces V by

V := {Φ ∈ C 0 (Ω) | Φ is a polynomial of degree ≤ 2 on each S ∈ T },

and V̊ := V ∩ H01 (Ω), where T is a conforming non-degenerate simplicial triangu-


lation of the polygonal domain Ω. Denoting by τ > 0 the timestep size, and using
the notations
 
T
Φm (x) = Φ(x, tm ), tm = mτ (m = 0, 1, . . . , M = )
τ
for a given function Φ : Ω × [0, T ] −→ R, a discretization of Problem 2.4 is given
by:

Problem 2.5. For given U0 (·, t) ∈ V, t ∈ [0, T ), find U m ∈ V (m = 1, . . . , M )


such that
βε (∇U m−1 ) U m − U m−1 ∇U m ∇Φ θm
Z Z Z
1
m−1 m−1
Φ + m−1
= Φ (2.12)
τ aε (∇U ) Q(U ) Q(U ) aε (∇U m−1 )
Ω Ω Ω

∀Φ ∈ V̊ ,

with initial and boundary conditions

U 0 = U0 , U m (·) − U0 (·, tm ) ∈ V̊ ,

where U0 (·, t) ∈ V denotes a finite element representation of u0 (·, t).

Remark 2.2. In case of the isotropic and homogeneous situation, the Time Dis-
cretization Scheme 2.5 is equivalent to the semi–implicit scheme for the mean cur-
vature flow of graphs proposed by Dziuk in [12], compare Remark 2.1. Using linear
finite elements Deckelnick and Dziuk proved in [9] a convergence result for such a
finite element scheme.

2.2. Numerical Experiments. We implemented the above algorithm with the


help of ALBERT, an adaptive multi–level finite element toolbox. ALBERT uses
bisectioning refinement and error control based on residual techniques, and pro-
vides also a wide range of linear system solvers, cf. [33]. The following numerical
experiments may illustrate the capabilities of our finite element algorithm for the
anisotropic mean curvature flow of level sets. The first example deals with an
8 MICHAEL FRIED

isotropic and homogeneous situation, where the exact solution is known. This
enables us to verify the algorithm by calculating the experimental order of conver-
gence. By the second experiment, we again investigate an isotropic and homoge-
neous case, to show the applicability of our scheme in situations where γ changes its
topological type. The level set method not only covers topological changes by break-
ing or merging of the moving interface but also situations where the free boundary
develops a nonempty interior, we show this by computing an inhomogeneous ex-
ample introduced by Paolini and Verdi, cf. [30]. Using a threefold anisotropy, we
approximate in the forth experiment the anisotropic mean curvature flow of a circle
toward the Wulff shape of the given anisotropy. As an outlook we sketch in the last
example the applicability of our algorithm in case of three dimensional situations
like the shrinking of a torus.
The Shrinking Circle. A known solution of the homogeneous and isotropic
p
mean curvature flow is the shrinking sphere γr0 (t) := {x ∈ R2 ||x| = r02 − 2 ∗ t}
with initial radius r0 > 0. For r0 = 1, the domain Ω = [−3.5, 3.5]2 and the initial
function
(
− cos( π9 r2 ) + cos( π9 ) for r ≤ 3,
u0 (r) = π
1 + cos( 9 ) for r > 3,
we have the radial symmetric solution
(
− cos( π9 (r2 + 2(n − 1)t)) + cos( π9 ) for (r2 + 2(n − 1)t) ≤ 9,
u(r, t) =
1 + cos( π9 ) else,
where r = |x|. Here we choose Neumann boundary conditions and calculate the
finite element approximations U using several uniform triangulations Tj . Starting
from a given triangulation T0 , the triangulation Tj is obtained from Tj−1 by twice
bisecting each triangle in Tj−1 , hence the gridsize hj is given by hj = 21 hj−1 . For
this numerical tests of convergence we choose the time step size τ to be τ = h2 .

h Err EOC Err EOC Err EOC


ε=h ε=h ε = h2 ε = h2 ε = h3 ε = h3
2.475 4.005 – 4.005 – 4.005 –
1.237 2.811 0.511 2.821 0.506 2.828 0.502
0.619 1.844 0.608 1.518 0.894 1.150 1.298
0.309 0.991 0.897 0.402 1.915 0.183 2.653
0.155 0.491 1.013 0.101 1.998 0.037 2.313
1) 1)
0.077 0.261 0.912 0.028 1.869
1) 1)
0.039 0.136 0.940 0.008 1.868
Table 1. EOC for ε = h, ε = h2 and ε = h3 . Time step size
τ = h2 . 1) Calculation stopped.

Table 1 shows the rates of convergence obtained using different relations between
the mesh size hj and the regularization parameter ε. The experimental order of
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 9

convergence is defined by
Errj
ln( Errj−1 )
EOCj := , (2.13)
ln2
with error Errj := max ku(·, tm ) − U m kL2 (Ω) .
1≤m≤M
For ε = h, we observe first order rates of convergence with respect to h while for
ε = h2 and ε = h3 we found second order rates of convergence. Therefore, we used
in the following experiments always the relations ε ≤ h2 and ε ≤ h2min for uniform
and not uniform meshes, respectively.
Topological Changes. One advantage of the level set approach is the ability to
handle topological changes during the evolution of the free boundary. To illustrate
this feature of our algorithm, we approximate in the following experiment the mean
curvature flow of a lemniscate. We consider γ0 to be given by the zero level of the
function
1
u0 (x) = ((x2 + (ax2 )2 )2 − 2(x21 − (ax2 )2 )), x = (x1 , x2 ) ∈ R2 ,
30000 1
with a scaling parameter a ∈ R+ . The zero level of this function looks like an “∞”,
compare Figure 1.

Figure 1. Lemniscate

Depending on the initial choice of the parameter a ∈ R+ , we found two qualita-


tively different types of topological changes: If a > 1, the initial lemniscate γ0 is
thin and long, thus the outer angles α0 are bigger than the inner ones, cf. Figure
1. Given such initial data, the evolving curve breaks into two tear shaped parts,
which independently of each other shrink to two points before vanishing. Figure 2
shows the evolution in the case a = 1.5.

If, on the other hand, we choose the parameter a to be a < 1, then the angles
α and α0 of the initial lemniscate are α0 < α, and the interface breaks at the self
intersection such, that it stays one connected curve which shrinks down to one point
10 MICHAEL FRIED

Figure 2. Shrinking lemniscate with parameter a = 1.5

at the origin, cf. Figure 3.

Figure 3. Shrinking lemniscate with parameter a = 0.5

Fattening. Evans and Spruck conjectured that instead of breaking, an evolving


interface may develop a nonempty interior, if at some time the interface is not
the boundary of a smooth open set, cf. [15], [17]. This phenomenon is called
fattening. Simply speaking, fattening is the “getting flat” of the function u around
its zero level. In our numerical experiments concerning the evolution of a given
interface, we always checked for the appearance of such an effect in a rather rough
way by testing if the finite element function U vanishes on any simplex of the
triangulation. Since this did not promise very strict results, in the next example we
used a more accurate method to numerically check for fattening. It was invented
by Paolini and Verdi, cf. [30], and is based on the following idea: If the zero level
of the function U develops a nonempty interior, one will find a jump in the volume
At (λ) := vol({x ∈ Ω|U (x, t) ≤ λ}) at λ = 0. Paolini and Verdi calculated this
quantity to show the numerical evidence of fattening using a phase field approach
to approximate the inhomogeneous mean curvature flow of two circles. At certain
critical time during the evolution, the two circles are touching each other. We
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 11

investigated this example by the help of our level set based algorithm. In contrast
to the previous examples we are now in a situation with non vanishing right hand
side θ.

20

19

18

17

16

15

14

13

12
-0.01 -0.008 -0.006 -0.004 -0.002 0 0.002 0.004 0.006 0.008 0.01

Figure 4. Fattening. Volumina A0.45 (λ) for −0.01 ≤ λ ≤ 0.01.

Paolini and Verdi considered the mean curvature flow of two circles with centers
at (xm , 0) and (−xm , 0), and initial radius r0 . Choosing the function θ = θ(t) :=
2 − t, the evolution of each of those circles isolated from the other one is given by
the ordinary differential equation
1
r0 (t) + = θ(t), r(0) = r0
r(t)

for the radius r(t). For r0 smaller than a critical radius rcrit , also the evolution of
both circles is ruled by the above ordinary differential equation. Until the time t = 2
the radius r(t) is growing, while for later times it shrinks until the circles vanish.
If the initial radius r0 > rcrit , then there is a time t∗ at which r(t∗ ) = xm holds.
In this case, the evolution of one isolated circle differs from that of two circles,
which will touch each other and merge to one single connected curve. At the time
t∗ , we are in a similar situation as we were at the initial situation of the previous
example. Due to the results of Paolini and Verdi, we expect that the interface will
fatten up, developing a nonempty interior. Calculating the volumina At∗ (λ) for
λ ∈ [−0.01, 0.01], we indeed found the numerical evidence for fattening of a level
set which evolves by its mean curvature, cf. Figure 4. This result corresponds to
the results obtained by Paolini and Verdi, cf. [30].
12 MICHAEL FRIED

Anisotropic Mean Curvature Flow. The following example illustrates the


application of our algorithm in case of the anisotropic mean curvature flow. To
keep it simple, we suppose β to be constant β = 1 and a to be given as

a(n(α)) = f (α) + f 00 (α) = 1 + c · cos(3α),

where α denotes the angle between the normal n and the x1 –axis, c is a given
parameter c > 0 and f (α) = 1 − 8c cos(3α). A method to visualize the amount of
anisotropy imposed by the function a is given by depicting the the Wulff shape

Wa := ∂{p ∈ R2 |hp, n(α)i ≤ f (α) ∀α ∈ R}.

In some sense, the Wulff shape Wa can be viewed as the unit sphere of the
anisotropic metric given by the anisotropy a. In the isotropic situation c = 0,
the Wulff shape Wa fulfills

Wa = ∂B1 (0).

Due to the assumption (1.2) the anisotropy a is convex in the sense of Gurtin,
cf. [23], and in case of a convex anisotropy it is known that a convex initial curve
γ0 during the evolution approximates the Wulff shape Wa , cf. [20], [29]. For the
following anisotropic experiment, we choose c = 0.8. Figure 5 shows the Wulff
shape with threefold symmetry belonging to that choice of the parameter c.

Figure 5. c = 0.8: Wulff shape of the anisotropy a

We start with the unit circle and calculate the anisotropic evolution on the
domain Ω = [−4, 4]2 choosing Neumann boundary conditions at ∂Ω. The shrinking
circle converges to the Wulff shape rather quickly, compare Figure 6.
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 13

Figure 6. c = 0.8: Anisotropic shrinking circle

2.3. Computing the Curvature. Using the level set method to approximate the
mean curvature flow of a given interface, an explicit calculation of the interface’s
mean curvature is not needed. However, for certain applications – as for instance
for the modified Stefan problem – one would like to know the curvature of γ(tm )
explicitly, as it appears in the computation of some other quantities, compare Sec-
tion 3.2. Here, γ(tm ) is a level set of a piecewise polynomial function U m ∈ V , and
unfortunately there is no straightforward definition of the curvature of such a level
set. We find a remedy by defining a discrete curvature K ∈ V of the level sets of
the finite element function U : In case of a smooth function u with non vanishing
gradient ∇u, the curvature of the level sets of u is given by
∇u
k(x, t) = ∇ · .
|∇u|

Recalling the way of regularizing the mean curvature flow equation (2.2), we again
p
replace the term |∇u| by the square root term Q(u) = ε2 + |∇u|2 . Multiplication
with a test function φ ∈ H01 (Ω), integration over the domain Ω and integration by
parts lead to
∇u ∇φ
Z Z
kε φ = − , (2.14)
Q(u)
Ω Ω

with suitable boundary values kε = k0 on ∂Ω, see below. From this, we define a
discrete curvature K ∈ V of the level sets of the function U m ∈ V by
∇U m ∇Φ
Z Z
KΦ = − ∀ Φ ∈ V̊ , (2.15)
Q(U m )
Ω Ω
14 MICHAEL FRIED

with boundary values K = K0 on ∂Ω, where K0 ∈ V is an finite element approx-


imation of the boundary data k0 . A typical choice of k0 would be k0 (x) = |x|−1 ,
which represents the curvature of the level sets of a rotationally symmetric function.
In principle, U m is not a rotationally symmetric function, but a discrete distance
function from γ(tm ) (compare Section 4.2 below), and so the level sets of U m near
the boundary ∂Ω are approximately circular, at least as long as the interface γ(tm )
is far away from ∂Ω.
Numerical Examples. The following numerical experiments illustrate numer-
ical convergence as well as possible problems of the above finite element approxi-
mation of the curvature. Firstly we consider the function u(x) = − cos( π9 |x|2 ) + π9 .
As the function u1 is rotational symmetric, the curvature of its level sets is given
1
by k1 (x) = |x| . We calculate the finite element approximation K1 of the interpo-
lation U ∈ V of u on the domain Ω = B2 (0) and compute the experimental order
of convergence EOC using linear, quadratic, and cubic finite elements. Again, we
define the EOC by (2.13) using L2 –errors. Since the curvature k1 has a singular-
ity at the origin, we compute the L2 –error on the domain D = Ω \ B0.5 (0). For
all computations we fixed the parameter ε to be ε = 0.0001. Table 2 shows the
obtained results.

hj Errj EOCj Errj EOCj Errj EOCj


linear linear quadratic. quadratic. cubic cubic
1 7.6639 – 2.3349 – 1.0229 –
0.5 9.6335 -0.33 0.9301 1.33 0.2022 2.34
0.25 12.0309 -0.32 0.4652 0.99 0.0465 2.12
0.125 12.4556 -0.05 0.2428 0.94 0.0138 1.75
0.0625 12.6097 -0.02 0.1228 0.98 0.0038 1.86
0.03125 12.6950 -0.01 0.0615 0.99 0.00098 1.96

Table 2. L2 –errors and EOCs for the approximation of k1

Obviously using linear finite elements the method did not converge, we need to
choose at least piecewise quadratic elements.
Further problems may arise if the gradient of the function vanishes on a whole
level line γc ∈ Ω. We choose the function
(
−dist(x, ∂B1 (0) ∪ ∂B2 (0)) for 1 ≤ |x| ≤ 2
d(x) :=
dist(x, ∂B1 (0) ∪ ∂B2 (0)) else,
on the domain Ω = B2 (0). Again the function is a rotationally symmetric, but
this time the gradient ∇d vanishes on the level line d = 12 in between the both parts
of the zero level set. Evaluating the L2 –errors on the domain D as before, even for
cubic elements we obtain no convergence of the finite element approximation K2
to the curvature k2 of the level lines of d. But the error could be localized around
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 15

the level line with vanishing gradient, by omitting the ring R := B1.75(0) \ B1.25 (0)
and calculating the error on the domain D0 := D \ R we found the results shown
in Table 3.

hj Errj EOCj Errj EOCj


in D in D in D0 in D0
1.0 23.0293 – 2.4 –
0.5 28.4363 -0.30 2.5759 -0.10
0.25 45.4699 -0.68 0.5952 2.11
0.125 54.8327 -0.27 0.2482 1.26
0.0625 78.9481 -0.53 0.1736 0.52

Table 3. L2 –errors and EOCs for the approximation of k2

3. Finite Element Method for the Heat Equation

The previous section dealt with the anisotropic mean curvature flow of level sets,
assuming the right hand side θ to be known. Vice versa, in this section we suppose
that the free boundary γ(t), its normal velocity v and curvature k are given. Then,
together with boundary and initial conditions, the equations (1.1) determine the
temperature θ.

Problem 3.1. Find the temperature θ such that

θt − ∆θ = 0 in [Ω \ γ(t)] × (0, T ), (3.1)

 l
∂θ
v=− on γ(t), (3.2)
∂n s

β(n) v + a(n) k + θ = 0 on γ(t), (3.3)

θ = θ0 on [Ω × {0}] ∪ [∂Ω × (0, T )] (3.4)

In order to derive a weak formulation of Problem 3.1, we pursue the method


proposed by Schmidt, cf. [32], multiplicate equation (3.1) by a test function φ ∈
H01 (Ω) and integrate over the domain [Ω \ γ(t)]. After integration by parts of the
principal term, we utilize equations (3.2), (3.3) and replace [Ω \ γ(t)] by Ω to get
for all φ ∈ H01 (Ω) and all t ∈ (0, T )
Z Z Z Z
1 a
θt φ + ∇θ ∇φ + θφ = − kφ (3.5)
β β
Ω Ω γ(t) γ(t)

We discretize equation (3.5) with the help of an implicit Euler method in time and
the finite element spaces V and V̊ as defined in Section 2.1. This leads to
16 MICHAEL FRIED

Problem 3.2. Given smooth interfaces γ(tm ) with curvature k and boundary val-
ues Θ0 (·, t) ∈ V, t ∈ [0, T ), find Θm ∈ V (m = 1, . . . , M ) such that
Z Z Z
1 m m−1 m 1 m
(Θ − Θ ) Φ + ∇Θ ∇Φ + Θ Φ (3.6)
τ β
Ω Ω γ(tm )
Z
a
=− kΦ ∀Φ ∈ V̊ ,
β
γ(tm )

with initial and boundary conditions

Θ0 = Θ0 , Θm (·) − Θ0 (·, tm ) ∈ V̊ ,

where Θ0 (·, t) ∈ V denotes a finite element interpolation of θ0 (·, t).

In the following section, Discretizations 2.5 and 3.2 for the mean curvature prob-
lem and the heat equation, respectively, are combined to get a method to solve the
Stefan problem (1.1).

4. Finite Element Method for the Stefan Problem

We now turn back again to the full Stefan problem (1.1). In contrast to the above
assumptions, neither the temperature nor the phase boundary and its curvature are
known.

4.1. Combining the Algorithms. As mentioned in Section 1, we combine the


both algorithms in the following way:
In each time step tm , given the old values Θm−1 and U m−1 , we first compute
the new free boundary by solving equation (2.12) for U m , where we replaced the
temperature θ on the right–hand side by the old discrete temperature Θm−1 . The
zero level of U m then gives us the new discrete interface γ m .
To compute the new temperature Θm we choose a modified version of equation
(3.6), replacing the interface γ(tm ) by γ m and the curvature k by the finite element
approximation K ∈ V of the curvature of γ m as it was defined in Section 2.3. Now
the temperature Θm is computed according to
Z Z Z
1 m m−1 m 1 m
(Θ − Θ ) Φ + ∇Θ ∇Φ + Θ Φ (4.1)
τ β
Ω Ω γm
Z
a
=− KΦ ∀Φ ∈ V̊ ,
β
γm

which ends the time step tm .

Remark 4.1. Replacing the curvature k by the approximation K ∈ V forces us to


choose at least quadratic finite elements, compare Section 2.3.

Summarizing, this gives the following level set based finite element discretization
of the Stefan problem (1.1):
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 17

Problem 4.1. For given U0 (·, t), Θ0 (·, t), K0 ∈ V, t ∈ [0, T ), find U m and Θm ∈
V (m = 1, . . . , M ) such that
βε (∇U m−1 ) U m − U m−1 ∇U m ∇Φ Θm−1
Z Z Z
1
m−1 m−1
Φ + m−1
= Φ (4.2)
τ aε (∇U ) Q(U ) Q(U ) aε (∇U m−1 )
Ω Ω Ω

∀Φ ∈ V̊ ,

and
Z Z Z
1 1
(Θm − Θm−1 ) Φ + ∇Θm ∇Φ + Θm Φ (4.3)
τ βε (∇U m )
Ω Ω γm

aε (∇U m )
Z
=− KΦ ∀Φ ∈ V̊ ,
βε (∇U m )
γm

with initial and boundary conditions

U 0 = U0 , U m (·) − U0 (·, tm ) ∈ V̊ ,
Θ0 = Θ0 , Θm (·) − Θ0 (·, tm ) ∈ V̊ ,

where

γ m := {x ∈ Ω|U m (x) = 0},

and K ∈ V is given by
∇U m ∇Φ
Z Z
KΦ = ∀ Φ ∈ V̊ ,
Q(U m )
Ω Ω

with boundary condition

K − K0 ∈ V̊ .

Remark 4.2. Though the analytical determination of the level set γ m of U m is


straightforward, the numerical determination is costly. In practice, we therefore
use a polygonal approximation Γm of γ m . The basic idea for an algorithm to
approximate γ m is to find the intersections of γ m with the edges of the triangulation
T , and then to connect them conveniently while preserving the topological type of
γ m . A detailed description of this algorithm goes beyond the scope of this article,
it could be found in [19].

4.2. Numerical Algorithms. In this section we describe some of the numerical


methods which were used to implement the Finite Element Method 4.1.

4.2.1. An Adaptive Algorithm. From numerical and physical experiments, it is


known that the temperature θ varies rapidly in the liquid phase near the inter-
face γ, but is flat far away from γ and inside the solid phase. In order to obtain
a good finite element approximation of the temperature, a fine grid is required, at
least in the region where θ varies heavily. A globally fine mesh, whose grid size is
small enough to ensure an approximation as accurate as needed for the numerical
simulation of dendritic growth, would lead to high computational efforts. Due to
18 MICHAEL FRIED

the observed behavior of θ, an adapted mesh, very fine around the phase boundary
and coarser far away from it, can be a let-out. Of course, since the phase boundary
moves in time, the spatially adapted mesh has to change in time, too. This require-
ments can be met by a time–dependent adaptive strategy, which at each time step
tm automatically adjusts the triangulation such, that an estimation of the error
kΘm − θ(·, tm )kL2 (Ω) is below a given tolerance tol but not much smaller than a
certain fraction of tol. Our method consists of three parts: First, some a–posteriori
error indicators, which provide information on the local (i.e. element wise) and the
global error, and a prescribed tolerance for the (indicated) global error. The second
part is a marking strategy. Instead of building up a completely new triangulation
each time the global error indicator says that the mesh has to change, this algo-
rithm selects elements of the actual mesh for refinement or coarsening. The last
part of our method consists of refinement and coarsening algorithms for the modi-
fication of the grid, which have to make sure that the modified triangulation is also
a conforming one.
As error indicators, we apply a simplified version of an a–posteriori error esti-
mator for the Poisson equation, which was introduced by Eriksson and Johnson in
[14], namely the local indicators
 1
X Z  m  2 2
1 ∂Θ
ηT (Θm ) :=  h3e

 ,
2 ∂ne
e∈∂T ∩Ω̊ e

where the sum is over all edges e of the triangle T , which belong to the interior of
Ω, and the global error indicator
! 12
X
m 2 m
η(Θ ) := ηT (Θ ) .
T ∈T

For marking, we use the guaranteed error reduction strategy proposed by Dörfler in
[11]. The idea is to refine a subset A of the triangulation that produces a consid-
erable amount of the indicated global error, and to coarsen another subset B such
that the indicated additional error after coarsening is not larger than some fixed
amount of the prescribed tolerance. A subset A could be found by the algorithm:

Algorithm 4.1 (Guaranteed error reduction strategy[11]).


Start with given parameters p1 , p2 ∈ (0, 1)

ηmax := max(ηT , T ∈ T )
sum := 0
κ := 1
while sum < (1 − p1 )2 η 2 do
κ := κ − p2
for all T in T do
if T is not marked
if ηT > κ ηmax
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 19

mark T for refinement


sum := sum + ηT2
end if
end if
end for
end while

A similar method leads to the subset B of triangles, which could be coarsened,


compare [11]. Finally, the modification of the mesh is done by a bisectioning algo-
rithm, which is taken from [1].
The initial triangulation is generated by an application of this adaptive procedure
to the stationary problem
Z Z Z Z
0 1 0 0
∇Θ ∇Φ + Θ Φ = − Θt Φ − γ K Φ, ∀Φ ∈ V̊ , (4.4)
β
Ω Γ(0) Ω Γ(0)

with given initial data Γ(0) and Θ0t and boundary value Θ0 . This method was pro-
posed by Schmidt, [31]. Beside the generation of an initial adapted grid, defining Θ0
as the solution of Equation (4.4) ensures that the initial temperature is compatible
with the discrete interface Γ(0) and its curvature K, compare [31].
The finite element algorithms for the approximation of the regularized level set
function uε are analogous to the methods described for the heat equation. Namely,
in both problems we apply the same finite element spaces, and consequently the
same triangulation. Hence the adaptive process is completely done using the error
indicator for the heat equation presented above.

4.2.2. Redistancing. Remembering the equation (2.14) the weak curvature K, we


see that this quantity depends on the gradient ∇U m of the level set function U m . In
practice, very steep or flat gradients ∇U m would lead to inaccurate approximations
∇U m
of the curvature K and the normal n = k∇U m k , as well as this could cause stability

problems like tip splitting. Therefore it is desirable to prevent having steep or flat
gradients develop in U m , whenever this would be possible. For instance, if U m is a
distance function to γ(tm ), we will avoid such ugly gradients. However even if we
choose u0 to be the signed distance function


 dist(x, γ0 ) if x ∈ ω s (0)

dγ0 (x) :=


−dist(x, γ0 ) if x ∈ ω l (0)

from γ0 , the level set function will cease to be a distance function after one timestep.
This may be fixed by reinitializing the function U m to be a distance function from
the discrete phase boundary γ(tm ) at each timestep. There are several possibilities
of redistancing the level set function, see for example Chopp [4], Strain [36] or Suss-
man, Sekerka and Osher [37]. In view of the fact that we already know a polygonal
approximation Γm of the interface γ(tm ), we use a direct method, calculating a
Lagrange interpolant of the distance function from the polygonal curve Γm .
20 MICHAEL FRIED

4.3. Numerical Results. We implemented a version of the above described algo-


rithm using the finite element tool-box ALBERT, which was developed by Schmidt
and Siebert, [33]. Here, we present some numerical results obtained by this imple-
mentation.

4.3.1. Convergence of the Numerical Method. As a first test we compare the solu-
tion of our finite element method with the known solution in case of an isotropic
test problem. This enables us to check the accuracy of our numerical method.
Using the notations

p
R(t) := R02 + 2t, R0 ∈ R+ , f (t) := Ẇ (t) = R(t) 3

2β √ Rs z2
e− 2
W (t) := − R(t) , β ∈ R+ , T (s) := − e z dz.
1

the problem reads as follows (c. f. Schmidt in [31]):




 θt − ∆θ = f for t ≥ 0, x ∈ R2 \ Γ(t),





 h
 i
∂θ
∂νΓ + VΓ = 0 on Γ(t),

 Γ(t)






θ + βCΓ + βVΓ = 0 on Γ(t).

The difference between the modified Stefan problem (1.1) and the above problem
is the right hand side f in the heat equation, which is chosen in such a way, that
an explicit solution is known. This solution is given by
(
θ(x, t) :=
W (t)   for |x| ≤ R(t)
|x|
W (t) + T R(t) else
1
and the free boundary Γ(t) = {x ∈ R2 | kxk = R(t) }, where v(t) = Rt (t) = R(t)
1
and k(t) = R(t) are the normal velocity and curvature of the free boundary.
Using the parameter β = 0.1 and the regularization parameter ε = 10−6 , we
investigate the behavior of error between the computed temperature Θ and the
exact temperature θ, solving the same problem with different globally refined spatial
meshes and time step sizes. The error was measured by

E∞,2 (Θh ) := max k Θ(·, ti ) − Θih kL2 (Ω) .


i=0,...,n

Table 4 shows the results of these calculations.

h 1.0 0.5 0.25 0.125 0.0625


∆t 0.25 0.0625 0.0156 0.0039 0.00098
E∞,2 (Θh ) 0.8319 0.3546 0.0657 0.0256 0.0133

Table 4. Exact Solution: Growing Circle. Evolution of the Error


while Uniformly Refining the Grid
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 21

4.3.2. Numerical Experiments. In this section, we show the results of different nu-
merical experiments in the case of the Stefan problem (1.1), where an explicit
solution is not known. For all of the following experiments, we choose the bound-
ary values Θ0 (·, t) = T0 to be constant on ∂Ω for all t ∈ [0, T ], and the functions a
and β of the form

β(n) = a(n) = δ(1 + A cos(kα + α0 )),

where α = αn denotes the angle between n and e1 – axis. The initial temperature
Θ0 is computed from Equation (4.4) with Θ0t = 0.
The first numerical experiment is a further isotropic calculation choosing the pa-
rameter A = 0. As opposed to the above test problem, we investigate the evolution
of an initially non–convex phase boundary, namely
( ! )
π cos(α)
Γ(0) = [0.15 + 0.4 cos(4α + )] | α ∈ [0, 2π] . (4.5)
4 sin(α)
The investigation of such a curve was proposed by Sethian and Strain in [35]. Here,
we are interested in the influence of the parameter δ. In some sense this parameter
plays the role of the anisotropic functions β(n) and a(n), and in a realistic setting,
it could be a very small parameter. We compare the numerical results for different
values of the parameter δ. Figure 7 shows the evolution for the parameters δ = 0.006
and δ = 0.0009 respectively.

Figure 7. Isotropic Growth for δ = 0.006 (left) and δ = 0.0009

For larger values of δ as in the first case (figure 7 left), we found a smooth
evolution of the phase boundary Γ, while the evolution for smaller δ ≤ 0.0009 tends
to be unstable. Figure 7 (right) shows a typical pattern with so called tip splitting:
the evolution becomes instable and develops small fingers and side branches. For
both situations we took the spatial domain to be Ω = B1 (0) the unit circle with
boundary value T0 = −1, and calculated the numerical solutions using the above
adaptive methods. If we choose a smaller tolerance for the adaptive method as well
22 MICHAEL FRIED

as finer time step sizes, the evolution is again stable, and shows qualitatively the
same pattern as in the case δ = 0.006. Therefore, we believe that the observed tip
splitting is due to grid effects.
We now come to results obtained by simulating the anisotropic growth of a seed
crystal into an undercooled melt. For the initial interface Γ(0) given by equation
(4.5), we calculated the evolution using the parameters δ = 0.006, A = 0.4 and
k = 4. Choosing the domain Ω = [−4, 4]2 and T0 = −1 we found results as
depicted in figures 8 and 9.

Figure 8. Fourfold Anisotropy: Evolution of the Phase Bound-


ary Γ and typical Adaptive Grid

Despite the anisotropic situation there are only small secondary dendrites. Fig-
ure 9 shows the graph of the temperature Θ. Here we observe the expected behavior,
nearly constant temperature inside the solid, but a fast decay in the liquid.

Figure 9. Fourfold Anisotropy: Graph of the Temperature Θ


with Grid

Starting with a rotational symmetric seed crystal will give us a better feeling of
the anisotropy’s influence on the growth. Therefore we took in a further calculation
the initial curve to be circular Γ(0) := ∂B0.1 (0). Figure 10 shows the result obtained
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 23

with the parameters δ = 0.004, A = 0.4, k = 4, T0 = −1 and the domain Ω =


[−4, 4]2 .

Figure 10. Anisotropic Evolution of an Initially Circular Solid Phase

Using this parameters we did indeed find a transition from the initial circle to
a fourfold dendrite. As in the last example, there was nearly no indication for
secondary dendrites.
Choosing higher undercooling, we expect smaller temperatures inside of Ω, and
such a faster growth of the solid phase. In the following experiment, we observe the
evolution in case of different undercooling. We also switched to a sixfold anisotropy,
fixing their parameters to be δ = 0.01, A = 0.4 and k = 6. Figures 11 to 14 illustrate
the results obtained with the boundary temperatures T0 = −1 and T0 = −2,
respectively. For both examples we choose the domain Ω = B4 (0) and the initial
phase boundary Γ(0) := ∂B0.1 (0).
Figures 11 and 12 show the numerical evolution of the phase boundary Γ(t)
and the graph of Θ in case of T0 = −1. As in the above examples, we observe
approximately constant temperatures inside the solid and steep gradients ∇Θ in
the liquid phase near the phase boundary, while the graph of Θ flattens again far
away from the interface Γ(t).
Reducing the boundary temperature the growth rate increases. In order to avoid
an unstable evolution, we have to decrease the timestep size as well as the grid size
in regions where |∇Θ| is big. This is done by decreasing the prescribed tolerance
in the above described self adaptive methods.
24 MICHAEL FRIED

Figure 11. Sixfold Anisotropy: Evolution of the Phase Bound-


ary with T0 = −1

Figure 12. Sixfold Anisotropy: Graph of the Temperature with


T0 = −1

The evolution of the solid phase obtained with the boundary temperature T0 =
−2 shows some topological changes. Different fingers touch each other and merge.
In between this fingers notches are developing, while the temperature gradient near
the phase boundary becomes steeper.
The last example illustrates the situation where multiple seed crystals separated
by an undercooled liquid evolve and eventually merge during their growth. We
found the topological change to be dependent on the the grid size of the underlying
triangulation. This is demonstrated by the evolutions, which are shown in Figure
15. The only difference between the two calculations was the prescribed tolerance
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 25

Figure 13. Sixfold Anisotropy: Evolution of the Phase Boundary


with T0 = −2

Figure 14. Sixfold Anisotropy: Graph of the Temperature with


T0 = −2, right: zoom to the solid phase

for the error. While in both cases we found the evolution to be stable, prescribing
a smaller tolerance leads to a situation without topological changes as we see in the
right picture of Figure 15.
Both initially given crystals where of the form of equation (4.5), but the left one
was rotated by 45◦ . The parameters of the anisotropy were δ = 0.006, A = 0.4 and
k = 4.

References
[1] E. Bänsch, Local mesh refinement in 2 and 3 dimensions, IMPACT Comput. Sci. Eng.,
(1991), pp. 181–191.
[2] G. Caginalp, Length scales in phase transition models: Phase field, Cahn-Hillard and blow-
up problems, in Pattern formation: symmetry methods and applications, J. Chadam and
26 MICHAEL FRIED

Figure 15. Topological Changes: Dependency on the Grid size

other, eds., Fields Institute for Research in Mathematical Sciences, Waterloo, Canada, 1996,
American Mathematical Society, Providence, Rhode Island, pp. 67–83.
[3] Y.-G. Chen, Y. Giga, and S. Goto, Uniqueness and existence of viscosity solutions of
generalized mean curvature flow equations, J. Differ. Geom., 33 (1991), pp. 749–786.
[4] D. L. Chopp, Computing minimal surfaces via level set curvature flow, Journal of Compu-
tational Physics, 106 (1993), pp. 77–91.
[5] M. G. Crandall and P. L. Lions, Convergent difference schemes for nonlinear parabolic
equations and mean curvature motion, Numer. Math., 75 (1996), pp. 17–41.
[6] K. Deckelnick, Error analysis for a difference scheme approximating mean curvature flow,
preprint, Mathematische Fakultät Freiburg, 1998.
[7] K. Deckelnick and G. Dziuk, Convergence of a finite element method for non-parametric
mean curvature flow, Numerische Mathematik, 72 (1995), pp. 197–222.
[8] , Discrete anisotropic curvature flow of graphs, preprint 34–96, Mathematische
Fakultät Freiburg, 1996. Submitted to RAIRO.
[9] , Convergence of numerical schemes for the approximation of level set solutions to the
mean curvature flow, Tech. Rep. 17/00, Mathematische Fakultät Freiburg, 2000.
[10] , A fully discrete numerical scheme for weighted mean curvature flow, preprint 30–00,
Mathematische Fakultät Freiburg, 2000.
[11] W. Dörfler, A convergent adaptive algorithm for Poisson’s equation, SIAM J. Numer.
Anal., 33 (1996), pp. 1106–1124.
[12] G. Dziuk, Convergence of a semi–discrete scheme for the curve shortening flow, Mathemat-
ical Models and Methods in Applied Sciences, 4 (1994), pp. 589–606.
[13] C. M. Elliott, Approximation of curvature dependent interface motion, Tech. Rep. 96/21,
Centre for Mathematical Analysis and Its Applications, University of Sussex, Falmer,
Brighton BN1 9QH, UK, 1996.
[14] K. Eriksson and C. Johnson, Adaptive finite element methods for parabolic problems i: A
linear model problem, SIAM J. Numer. Anal., 28 (1991), pp. 43–77.
[15] L. C. Evans and J. Spruck, Motion of level sets by mean curvature I, J. Differ. Geom., 33
(1991), pp. 635–681.
[16] , Motion of level sets by mean curvature II, Trans. Amer. Math. Soc., 330 (1992),
pp. 321–333.
[17] , Motion of level sets by mean curvature III, J. Geom. Ana., 2 (1992), pp. 121–150.
[18] , Motion of level sets by mean curvature IV, J. Geom. Ana., 5 (1995), pp. 79–116.
[19] M. Fried, Niveauflächen zur Berechnung zweidimensionaler Dendrite, PhD thesis, Institut
für Angewandte Mathematik, Universität Freiburg, April 1999.
January 8, 2001: FINITE ELEMENT SIMULATION OF DENDRITIC GROWTH 27

[20] M. Gage, Evolving plane curves by curvature in relative geometries, Duke Math. J., 72
(1993), pp. 441–466.
[21] Y. Giga, S. Goto, and H. Ishii, Global existence of weak solutions for interface equations
coupled with diffusion eqautions, Siam J. Math. Anal., 23 (1992), pp. 821–835.
[22] M. E. Glicksman and S. P. Marsh, The dendrite, in Handbook of Crystal Growth, D. T. J.
Hurle, ed., vol. 1, North Holland, Amsterdam – London – New York – Tokyo, 1993.
[23] M. E. Gurtin, Thermomechanics of evolving phase boundaries in the plane, Clarendon Press,
Oxford, 1993.
[24] M. E. Gurtin and H. M. Soner, Some remarks on the Stefan problem with surface structure,
Quarterly of Applied Mathematics, 52 (1992), pp. 291–303.
[25] G. Huisken, Non–parametric mean curvature evolution with boundary conditions, Journal
of Differential Equations, 77 (1989), pp. 369–378.
[26] J. S. Langer, Instabilities and pattern formation in crystal growth, Reviews of Modern
Physics, 52 (1980), pp. 1–28.
[27] G. Lieberman, The first initial–boundary value problem for quasilinear second order para-
bolic equations, Ann. Sci. Norm. Sup. Pisa Ser., IV 8 ??? (1986), pp. 347–387.
[28] S. Osher and J. A. Sethian, Fronts propagating with curvature–dependent speed: Algo-
rithms based on Hamilton–Jacobi formulations, J. Comput. Phys., 79 (1988), pp. 12–49.
[29] M. Paolini, An efficient algorithm for computing anisotropic evolution by mean curvature.,
in Proceedings of the international conference on curvature flows and related topics held in
Levico, Italy, June 27-July 2nd, 1994., A. Damlamian, ed., Tokyo: Gakkotosho. GAKUTO
Int. Ser., Math. Sci. Appl. 5, 1995, pp. 199–213.
[30] M. Paolini and C. Verdi, Asymptotic and numerical analyses of the mean curvature flow
with a space–dependent relaxation parameter, Asymptotic Anal., 5 (1992), pp. 553–574.
[31] A. Schmidt, Die Berechnung dreidimensionaler Dendriten mit Finiten Elementen, PhD
thesis, Universität Freiburg, 1993.
[32] , Computation of three dimensional dendrites with finite elements, Journal of Compu-
tational Physics, 125 (1996), pp. 293–312.
[33] A. Schmidt and K. G. Siebert, Albert: An adaptive hierarchical finite element toolbox.
Lecture Notes, Freiburg, 1997.
[34] J. A. Sethian, Level Set Methods, Cambridge University Press, 1996.
[35] J. A. Sethian and J. Strain, Crystal growth and dendrite solidification, J. Comp. Phys.,
98 (1992), pp. 231–253.
[36] J. Strain, Fast tree–based redistancing for level set computations. Preprint, Submitted to
J. Comp. Phys., 1998.
[37] M. Sussman, P. Smereka, and S. J. Osher, A level set method for computing solutions to
incompressible two–phase flow, Journal of Computational Physics, 114 (1994), pp. 146–159.
[38] A. Veeser, Globale Existenz von Graphen unter inhomogenen Krümmungsfluß bei Dirichlet–
Randbedingungen, Master’s thesis, Institut für Angewandte Mathematik, Universität
Freiburg, December 1993.
[39] A. Visintin, Models of phase transitions, vol. 28 of Progress in Nonlinear Differential Equa-
tions and Their Applications, Birkhäuser, Boston, 1996.
[40] N. J. Walkington, Algorithms for computing motion by mean curvature, SIAM J. Nu-
mer. Anal., 33 (1996), pp. 2215–2238.

Institut für Angewandte Mathematik, Universität Freiburg, Hermann–Herder–Str.


10, D–79104 Freiburg, Germany

You might also like