JOURNAL OF COMPUTATIONAL PHYSICS 125, 293–312 (1996)
ARTICLE NO. 0095
Computation of Three Dimensional Dendrites with Finite Elements
ALFRED SCHMIDT*
Albert-Ludwigs-Universität Freiburg, Institut für Angewandte Mathematik, Hermann-Herder-Straße 10, D–79104 Freiburg i. Br., Germany
Received August 10, 1994; revised September 25, 1995
Starting from an initial seed crystal inside an undercooled liquid,
the solid phase begins to grow rapidly and develops unstable
growth patterns. Some growth directions are preferred because of
anisotropic parameters in the physical model. This results in the
development of dendrites. The physical model includes the heat
equation for both the liquid and solid phases; the Gibbs–Thomson
law couples velocity and curvature of the interface and the tempera-
ture. We describe a numerical method that enables us to compute
dendritic growth of crystals in two and three space dimensions. The
method consists of two coupled finite element algorithms. The first
one solves the heat equation; the other operates on a discretization
of the free boundary and computes the evolution of this moving
interface. The two methods work with totally independent grids.
By using timedependent, locally refined and coarsened adaptive
meshes in both methods, we are able to reach a spatial resolution
necessary to compute dendritic growth in two and three space
dimensions. Q 1996 Academic Press, Inc.
Solidification of a pure substance is a well-known phe-
nomenon, but nevertheless not completely understood, ei-
ther physically or mathematically. Strongly varying situa-
tions and interface geometries may develop, depending on
physical parameters and initial and boundary conditions.
We want to examine the unstable situation of a dendrite
growing in an undercooled melt.
The physical modelling of the solidification is mostly
complete. The article [G], for example, treats the deriva-
tion of models with anisotropic parameters. Mathematical
treatment of the modelling equations is a very progressive
field—known results mostly cover the classical Stefan
problem, which corresponds to a simplified model.
For a number of years, numerical simulation of crystal
growth and dendrites has been attempted. In the beginning,
only two dimensional problems were solved. Now, it is
possible to compute truly three dimensional solutions, us-
ing high performance computers and sophisticated algo-
rithms.
Many different numerical methods have been published,
based on several different formulations of the free bound-
ary problem. Nearly all showed only two dimensional re-
sults; a fairly incomplete list of references is [A, BJGLS,
* Partially supported by SFB 256, Universität Bonn.
293
CL, Ch, D2, FL, Har, HK, K1, Me, NPV1, NPV2, RT,
SGWMK, SS, Sm, St, SLN, TH, WMS]. Several people
are working on algorithms for 3D problems, but the only
published three dimensional results that we know of are
those of Kobayashi [K2] and a few results of Harnisch
[Har].
In this article, a numerical method is presented for the
computation of the dendritic growth of crystals from an
undercooled melt in a container of two or three dimen-
sions. This method consists of two finite element algo-
rithms, which are coupled through integrals over the free
boundary. One algorithm solves the heat equation on a
tetrahedral (resp. triangular) subdivision of the container.
The other operates on a triangular (resp. polygonal) dis-
cretization of the free boundary. It computes the evolution
of this moving interface. The two methods work with to-
tally independent grids. In each time step, both meshes
are locally refined and coarsened, based on a posteriori
error estimates of the discrete solutions. The location of
the interface does not enter the refinement decision di-
rectly, only through these estimates. As the free boundary
triangulation is deformed during the evolution in time,
special care has to be taken to prevent the development
of obtuse angles. Thus we have to improve the grid. Only
the use of timedependent locally refined and coarsened
adaptive meshes in both methods enables us to reach a
spatial resolution which is necessary to compute growing
dendrites in two and three space dimensions. A flow
diagram of the complete algorithm is shown in Fig. 0.1.
The physical model which describes the solidification is
presented in Section 1. Section 2 develops the two finite
element algorithms for the temperature and the free
boundary evolution. Some numerical details such as error
estimation and mesh adaptation and the coupling of the
two algorithms are described in Section 3, followed by
numerical results in Section 4.
We use the following notations and definitions through-
out this article: Lp(V) and H m,p(V) denote the usual Leb-
esgue and Sobolev spaces; see [Wl] about these spaces on
Lipschitz manifolds. (?, ?) is the L2 scalar product, and k?, ?l
the Euclidian scalar product in Rn. = G denotes the covariant
(tangential) derivative and DG the Laplace–Beltrami opera-
tor on a manifold G.
0021-9991/96 $18.00
Copyright  1996 by Academic Press, Inc.
All rights of reproduction in any form reserved.
294 ALFRED SCHMIDT

FIG. 0.1. Flow diagram of the adaptive algorithm.

1. THE PHYSICAL MODEL Vl(t) , V with Vs(t) > Vl(t) 5 B, V 5 Vs(t) < Vl(t) be
the parts of V containing the solid and liquid phases of
Solidification (crystallization) and melting of a pure heat the substance at time t. The free boundary between solid
conducting substance is described by the following physical and liquid phases will be denoted by G(t) :5 Vs(t) > Vl(t).
model. It incorporates the temperature and the distribution The principal situation in two dimensions is shown in
of solid and liquid material in a container. Similar equa- Fig. 1.1.
tions hold in a model for crystal growth in an isothermal The physical constants for the different phases are heat
solution, where chemical diffusion instead of heat diffusion conductivity, ks , kl , specific heat, cs , cl , and density, rs ,
is the controlling factor. Both models are described in rl . In each phase the heat equation for the (absolute)
detail in [G, L]. temperature T holds,
Let n 5 1 or n 5 2 be the dimension of the interface
and n 1 1 the space dimension of the domain V , Rn11 ­T ­T
5 Ds DT in Vs , 5 Dl DT in Vl , (1.1)
containing the substance. For t $ 0 let Vs(t) , V and ­t ­t
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 295

remainder of this paper, we will consider only the case

b . 0.
In three dimensions, it may be better to use an aniso-
tropic surface tension to model the anisotropy. This results
in some anisotropic curvature replacing the mean curva-
ture CG . Use of this anisotropic curvature leads to weak
formulations, similar to those used in Sections 2.1 and 3.3.
The overall numerical method does not change; results
will be presented in a forthcoming publication.
With a given initial temperature distribution T0 and an
FIG. 1.1. Solid phase Vs , liquid phase Vl , and interface G.
interface G0 , we pose the following boundary and initial
value conditions:
where the diffusion constants Ds , Dl are defined by the
T (x, t) 5 T0(x, t) on ­V, t $ 0,
physical constants as Di :5 ki /(ci ri), i 5 s, l. At the free
boundary G, its velocity and the temperature fulfil an equa- T (x, 0) 5 T0(x, 0) in V,
tion known as the Stefan condition:
G(0) 5 G0 .

LVG 5 Ds cs
­T u Vs
­ns
1 Dlcl
­T u Vl
­nl
F
5: 2 Di ci
­T
­nG
G
G
on G(t). Using a dimensionless temperature u :5 (cs 1 cl)(T 2 TM)/
2L, the system of equations (1.1)–(1.3) transforms to
(1.2)
­u
5 Ds,l Du in Vs,l , t . 0,
Here, ns and nl denote the outer normal vectors to the ­t

F G
domains Vs and Vl , and we take nG :5 ns 5 2nl to be the
normal of the free boundary G. VG is the velocity of the 22 ­u
VG 5 D i ci on G,
free boundary G in direction of nG , and L is the latent heat cs 1 cl ­nG G

per unit volume in the solid phase. Finally, [?]G denotes
the jump at the free boundary. The left hand side of (1.2)
describes the rate at which heat is generated by solidifica-
tion at the free boundary. The right hand side describes
the heat transport into the solid and liquid phase.
Additionally, a thermodynamical condition holds at the
free boundary:
u 5 2«C CG 2 «VVG on G,
which is a more ‘‘usual’’ formulation of our problem. Here
we used «C (nG) 5 c(nG)(cs 1 cl)TM /2L2 and « V(nG) 5 b(nG)
(cs 1 cl)TM /2L2. The initial and boundary values transform
in a similar way. With u0(x, t) :5 (cs 1 cl)(T0(x, t) 2 TM)/
2L we get

T 5 TM 1 2 S c
L
b
CG 2 VG
L
D on G. (1.3) u(x, t) 5 u0(x, t) on ­V, t $ 0,
u(x, 0) 5 u0(x, 0) in V, (1.4)
Here, TM is the melting temperature of the substance, and
G(0) 5 G0 .
c 5 c(nG) is the surface tension between the solid and
liquid phase (usually depending on the direction of nG). CG
is the mean curvature of the free boundary G (sum of the In the remainder of this paper, we want to work with the
principal curvatures). The sign of CG is taken in such a way simplified case, where the physical constants in the solid
that the mean curvature for a convex solid phase, Vs , is and liquid phases are equal:
positive. The coefficient b 5 b(nG) depends on nG in the
general case, too. Equation (1.3) is known as the Gibbs– ks 5 kl , cs 5 cl , rs 5 rl , Ds 5 Dl 5: D0 .
Thomson law. If c or b depends on the direction of the
normal nG then the coefficients are called ‘‘anisotropic,’’ This assumption applies not only to academic examples,
else ‘‘isotropic.’’ The term cCG /L describes the influence there exist real materials, such as Succinonitil, which obey
of surface tension, which stabilizes the motion and makes to the same physical constants in the solid and liquid phase
dendritic growth possible. With b 5 0, equation (1.3) de- (see [SLN, Table IV]). On the other hand, our method
scribes a situation in local thermal equilibrium (see [L] for may be adapted to the case with different constants in
details), while the nonequilibrium situation with a moving both phases. Altogether, we now get the following system
interface is modeled with b . 0 (compare [G]). In the of equations:
296 ALFRED SCHMIDT
Problem 1.1. Find a temperature u and a moving free
boundary G solving the equations
­u
2 D0 Du 5 0 in Vs < Vl , t . 0, (1.5)
­t
D0 F G
­u
­nG G
1 VG 5 0 on G, t . 0, (1.6)
u 1 «C CG 1 «VVG 5 0 on G, t . 0, (1.7)
together with the boundary and initial value conditions
(1.4).
2. FORMULATION OF THE NUMERICAL METHODS
In this section, we describe the main parts of a numerical
method for the solution of Problem 1.1 on a time interval
[0, Tmax]. We derive separate finite element methods for
the free boundary G and the temperature u. Numerical
details of both methods are treated in Section 3, and in
Section 3.4 we combine these two methods to get a method
for the solution of the whole problem.
The nonlinear time dependent Problem 1.1 will be dis-
cretized into n 1 2 decoupled scalar linear equations in
each time step, one for the temperature, and one for each
coordinate of the free boundary.
Wherever possible, we treat the two cases n 5 1 and
n 5 2 in common, so the term ‘‘surface’’ means ‘‘curve’’
as well if n 5 1.
2.1. Finite Element Method for the Free Boundary
For the derivation of our finite element method for the
interface G(t), we suppose that the temperature u(x, t) on
the free boundary is given for all times t. Then the free
boundary is described by equation (1.7) and the initial
value G0 .
Problem 2.1. Find a family G(t), t [ [0,Tmax], of smooth
surfaces, which solves the equation
«V (nG(t)(x))VG(t)(x) 1 «C (nG(t)(x))CG(t)(x)
5 2u(x, t) for all x [ G(t), t [ (0, Tmax), (2.1)
G(0) 5 G0 ,
where VG is the scalar velocity in the direction of the surface
normal nG , and CG is the scalar mean curvature of the
surface.
For this moving free boundary, we follow the formula-
tion of the algorithm for the mean curvature evolution
of surfaces from [D1]. We describe the moving surface
G(t) as a family of n-manifolds, parametrized by finitely
many (local) charts Fi(t): Vi R Rn11, i [ I, where the
parameter domains Vi , Rn are independent of t. Now the
determination of G(t) is equivalent to the determination of
parametrizations Fi(t), i [ I, and Eq. (2.1) can be
replaced by
«V(nG n Fi)VG n Fi 1 «C (nG n Fi)CG n Fi 5 2u n Fi
in Vi 3 (0, Tmax), i [ I.
After multiplication of this scalar equation by the unit
normal nG of G(t), we get
(«V (nG)VG nG) n Fi 1 («C (nG)CG nG) n Fi 5 2(unG) n Fi
in Vi 3 (0, Tmax), i [ I.
We demand that the parametrization is chosen in a way
that its time derivative (velocity) has no tangential compo-
nent: (­/­t)Fi 5 (VG nG) n Fi for all i [ I. This is possible
at least for short time intervals, because the distance
function is regular in a neighbourhood of a smooth
surface (with small t, for each point P of G(t 1 t), there
is a unique point P̃ of G(t) which moves with normal
velocity to P).
Using the fact that the curvature vector of a surface is
normal to the surface and equal to the Laplacian of the
identity on it,
2DG idG(x) 5 CG nG for all x [ G
with idG(Fi ( y)) 5 Fi ( y) for all y [ Vi , i [ I, we get the
following equation for the charts:
­
«V (nG n Fi) Fi 2 «C (nG n Fi) DG(t)Fi
­t (2.2)
5 2(unG) n Fi in Vi 3 (0, Tmax), i [ I.
With discrete time values 0 5 t0 , t1 , t2 , ... # Tmax
and time steps tm :5 tm11 2 tm , using an implicit Euler
discretization of the time derivative, we derive semidiscrete
equations for the charts Fim, m 5 0, 1, ..., approximating
Fi(tm) and manifolds Gm11 which approximate G(tm11):
F0i 5 Fi (0),
Fm
i
11
2 Fim
«V (nGm) 2 «C (nGm) DGm Fm 11
tm i
5 2unGm in Vi , i [ I, m 5 1, 2, ...,
Gm11 :5 <F i[I
m11
i (Vi).
This discretization is only implicit in the sense that the
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 297

Laplace–Beltrami operator DGm is applied to the unknown In [Wl, Chap. 4.2], the Sobolev space H 1(G) is defined for
parametrization Fm i
11
, but DGm is still defined by the known a Lipschitz manifold G. This definition applies here, as our
surface G . We can now leave the local parametrizations
m
discretization Ghm is Lipschitz continuous, and we get
of the manifolds over domains Vi , Rn and use a global W hm , H 1(Ghm).
parametrization of the next manifold Gm11 over the previ- The choice PT̂ 5 Pk , yielding same order polynomials
ous manifold Gm. We define Fm11: Gm R Gm11 by Fm11 for the discretization and the finite element functions, is
(Fim( y)) :5 Fm
i
11
( y) for y [ Vi . This definition is indepen- natural in the sense that we can expect optimal error esti-
dent of the choice of a local chart. After division by «C , mates; see [Ci, Chapter 4.4].
we get the following semidiscrete equation: Analogous to Problem 2.2, we now look for a weak
solution in the finite element space W hm on Ghm . This gives
«V (nGm(x)) Fm11 2 x an approximation Fm h
11
to the parametrization Fm11 and
2 DGm Fm11 defines a discrete manifold Gm 11
approximating Gm11.
«C (nGm(x)) tm h

1 Problem 2.3. Let G0h be a given triangulation of G0 . For

52 u(x)nGm(x) on Gm, m 5 0, 1, 2, ..., find a finite element solution Fm 11
[
«C (nGm(x)) m n11
h
Wh(Gh ) of the equation
Gm11 :5 Fm11(Gm).

E «V Fm (x) 2 x
E
11
h
Thus, we have converted the nonlinear differential equa- ch 1 =Fm 11
=ch
m
Gh «C tm m
Gh
h

tion (2.2) into a sequence of linear equations.

Multiplication with a test function and integration by
parts leads to the following weak formulation, which is
52 Em
Gh
1
«C
unch for all ch [ Wh(Ghm),
(2.3)

well suited for a finite element approximation.

Gm
h
11
:5 Fm
h
11
(Ghm).
Problem 2.2. For m 5 0, 1, 2, ..., find a parametrization
Fm11 [ H 1(Gm)n11 which solves
For each time step this leads to a decoupled system of
n 1 1 linear systems of equations for the components of
EG m
«V (nGm) Fm11 2 x
«C (nGm) tm
c1 E Gm
=Gm Fm11 =Gmc the parametrization. The matrices of the n 1 1 linear sys-
tems are all equal, but right hand sides differ between
the components.
52 EGm
1
«C (nGm)
unGmc for all c [ H 1(Gm).
2.2. Finite Element Method for the Heat Equation
Triangulation of the Manifold and Finite Elements. For We now suppose that the manifolds G(t) and their veloci-
a finite element discretization of Problem 2.2, we first ties and curvatures are given and the temperature u(x, t)
need a discretization of the manifold Gm. We are able is unknown. Together with initial and boundary conditions,
to construct a discretization G0h of G0 by a conforming Eqs. (1.5) and (1.6) determine the temperature.
(globally continuous) mesh of parametric n-simplices of
polynomial order k (i.e., curved simplices, if k . 1), Problem 2.4. Find a temperature u which solves the
based on triangulations of the parameter domains Vi , equations
i [ I. Each simplex interpolates G0 at a set of Lagrange
nodes corresponding to the polynomial degree k. This ut 2 D0 Du 5 0 in (V\G(t)) 3 (0, Tmax),

FG
discretization can be constructed by polynomial Lagrange
interpolation of local charts; see [KW, N] for details. ­u
D0 5 2VG on G(t),
For m . 0, a discretization Ghm of Gm is generated from ­n
Gmh
21
, as we will show later. u 5 2«C CG 2 «VVG on G(t),
We use isoparametric finite elements over Gh . Let PT̂
be a finite dimensional function space on the n dimensional u 5 u0 on (V 3 h0j) < (­V 3 [0, Tmax]).
unit simplex T̂ satisfying Pk , PT̂ (usually, we take PT̂ 5
Pk). For every T , Ghm let FT : T̂ R T be the (polynomial) Using an implicit Euler discretization of the time deriva-
parametrization. Using this notation, we define the isopar- tive with 0 5 t0 , t1 , ... # Tmax and time steps tm :5 tm11 2
ametric finite element space tm , we derive equations for temperatures um approximating
u(tm). We set u0 :5 u0(t0), and for m 5 0, 1, 2, ..., let um11
W hm :5 hvh [ C 0(Ghm, R) : (vh n FT) [ PT̂ for all T [ Ghmj. be the solution of
298 ALFRED SCHMIDT

u m11 2 u m C 0(V 3 [0, Tmax]), and discrete initial values u 0h [ V 0h , find

2 D0 Du m11 5 0 in V\G(tm11), (2.4)
tm for each m 5 0, 1, 2, ... finite element temperatures
[ Vm
F G
11 11
um
h h which solve
­u m11
D0 5 2VG on G(tm11), (2.5)
­n
E 2 u hm
E E
11
um
h 1 m11
m11
5 2«C CG 2 «VVG on G(tm11), wh 1 D0 =u m 11
=wh 1 u h wh
u (2.6) V tm V
h
G(tm11) «V
(2.7)
u m11 5 u0 on ­V.
52 E «C
G(tm11) «V
CG(tm11)wh for all wh [ V m 11
h,0 ,

In the weak formulation, we combine (2.4)–(2.6). Equation

(2.4) leads to
with boundary values u m
h
11
5 P m11(u0(?, tm11)) on ­V.
As mentioned above, the discretizations for the free
EV
u m11 2 u m
tm
w 1 D0 E V
=u m11 =w boundary evolution and the heat equation are combined
in Section 3.4 to a numerical method for the solution of
5 E G(tm11)
VGw for all w [ H 10(V),
Problem 1.1.

as integration by parts produces the left hand side of (2.5)
on G(tm11). Using Eq. (2.6), which gives a relationship be-
tween the velocity VG , the curvature CG and the tempera-
ture at the free boundary, we get for all w [ H 10(V)
3. NUMERICAL ALGORITHMS
In this section, some of the numerical methods which
we use to discretize and solve the problem are described
in detail. We look at the algorithms which apply to the
heat equation, the moving free boundary, and the computa-

E V
u m11 2 u m
tm
w 1 D0 EV
=u m11 =w 1 EG(tm11)
1 m11
«V
u w
tion of its curvature, and last we describe the coupling of
all parts.

52 E
G(tm11)
«C
«V
CGw.
The integral over G(tm11) on the left-hand side leads to a
more stable discretization, as we use an implicit discretiza-
tion for the temperature on the free boundary and 1/«V is
positive. For simplicity, we assume that V is a polygonal
domain, so that we do not have to worry about an approxi-
mation of the boundary. Let TV be a subdivision of V into
(n 1 1)-simplices (triangles or tetrahedra). This triangula-
tion will be chosen different for different time steps, so we
use the notation T Vm for the triangulation at time tm . We
define finite element spaces V hm , V h,0
m
on the triangulation
m
T V by setting
V hm :5 hvh [ C 0(V) : vh u T [ Pk for all T [ T
and
m
V h,0 :5 hvh [ Vh : vh u­V 5 0j.
Finally, let P m denote the Lagrange interpolation operator
corresponding to V hm . This leads to the following discreti-
zation:
Problem 2.5. Given smooth surfaces G(tm), m 5 0,
1, ..., with mean curvature CG(tm) , boundary values u 0 [
3.1. Heat Equation
The solution of the heat equation varies rapidly in the
liquid phase Vl near the interface. We need a fine grid
there, to be able to get a good approximation of the
temperature near the interface. But, especially in the
three dimensional case, it is nearly impossible and much
too costly to use a globally fine mesh with an h-size
small enough to resolve the temperature with enough
exactness near the interface to compute dendrites. Fortu-
nately, the temperature is nearly constant in the solid
phase Vs and is also smooth in the liquid phase far away
from the interface, so that a relatively coarse grid is
sufficient to obtain a good approximation there. Due to
this observation, we may use a grid with spatially varying
fineness; this holds the memory requirements and compu-
m
V j
tation time low. As the free boundary moves in time,
the grid has to vary in time, too, and the highly refined
region must move accordingly.
These demands can be met by a selfadaptive time-
dependent strategy. It uses an a posteriori error estimator
that produces information, where we need a finer grid
or where we may use a coarser one, and ensures the
quality of the numerical solution. The mesh is not built
totally new in each time step; we locally modify the
mesh from the last time step to generate the mesh for
the actual one.
The grid modification algorithms for the local refinement
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 299

and coarsening of tetrahedral (resp. triangular) meshes are hT (uh)

S O E h UF­n­u GU D
taken from [B1, B2]. They are based on the bisectioning

5
2 1/2
of simplices. Using estimates hT(u hm) for the local error
iu(tm) 2 u hmiT on a simplex T, in each time step the adaptive
ch D20 E T
h4 uDuh u 2 1
D20
2 E[­T>V̊ E
3 h

E
method selects simplices for refinement or coarsening.
if T > G 5 B, and else
Given a triangulation of V consisting of N simplices and
an upper bound d and a lower bound d for the total error,
the adaptive method selects simplices T with hT(uh) ,
d / ÏN for coarsening and simplices with hT(uh) . d / ÏN
:5
ch SE G>T
h2 S «C CG 1 uh
«V
D E 2
1 D20
T
h3 uDuh u 2

O UF GU D
2 1/2
for refinement. This leads to an equidistribution of the
error over all elements. Due to the local situation (adja- 1
D20
2 E[­T>V̊
EE
h2
­uh
­nE
.
cency of elements marked for coarsening, for example),
not all simplices that are selected for coarsening may be
coarsened (compare [B2, He]). On the other hand, all Here, E , ­T > V̊ denote the non-boundary faces (resp.
simplices marked for refinement have to be refined, so that edges) of a simplex T, and nE the corresponding unit nor-
the local error can really be reduced where the estimator mal. The constant ch depends on the geometry (size of the
indicates that it is necessary. domain, etc.) and regularity of the mesh. The lower h
The whole adaptive strategy can be summarized as exponents near G account for the loss of regularity of the
follows: temperature at the free boundary. We can prove the a
posteriori error estimate
—Starting from an initial coarse triangulation of the

SO D
(time-independent) domain V, we first create a triangula- 1/2
tion for the initial value u 0h by adaptive refinement. Simpli- iu 2 uh iL2(V) # h(uh) :5 hT (uh)2
ces are selected for refinement based on local error esti- T[TV

mates for an approximation of the initial temperature, such

that the estimated error is finally below a given bound and
is equidistributed over all simplices of the grid.
—In every time step, starting with the mesh from the
previous time step, the mesh gets adaptively refined and
coarsened in such a way that the estimated error is beneath
the given bound d , not much smaller than d, and nearly
equidistributed over all simplices. We use the semi-implicit
strategy from [B3] for the mesh adaptation and tempera-
ture computation:
We first solve the heat equation on the previous time
step’s grid T Vm; this gives a solution ũ m
h
11
[ V hm . The grid
m m
T V is modified into T V , based on the error estimates
hT (ũ m
h
11
), T [ T Vm . We solve the heat equation on this
new grid again to calculate the new temperature u m h
11
[ triangles (or as a piecewise linear or quadratic curve in
m11
Vh .
The basis of an adaptive process is given by estimation
of the error. We derive an L2-error estimate for the temper-
ature, analogous to the methods by [EJ]. We obtain an a
posteriori error estimate for the solution of the associate
elliptic problem (Eq. (2.7) without the temporal differen-
tial quotient), assuming a smooth interface G(t) and regu-
larity of u for all t [ [0, T],
u(?, t) [ H 1,y(V) > H 2,2(Vs < Vl),
with a corresponding a priori estimate (see [CR] for esti-
mates for the anisotropic parabolic problem). Let
under the above regularity assumption. The values hT (uh)
are used by the adaptive method to drive the refinement
and coarsening of the mesh.
3.2. Free Boundary
The algorithms for the approximation of the free bound-
ary and its movement are mostly analogous to the heat
equation. Differences are necessary because of the use of
parametric elements and the changes of geometry in time.
For the triangulation of G, we use a regular simplicial
mesh; the free surface is modelled as a closed polyhedron
consisting of piecewise linear or quadratic parametrized
the case n 5 1). The Lagrange degrees of freedom are
located at the vertices of the mesh plus, for quadratic
polynomials, the midpoints of the edges. To each of these
points belong n 1 1 degrees of freedom for the spatial
coordinates of the point.
The shape of the free surface will vary strongly during
the evolution. Usually we start with a small sphere (or a
circle). Depending on the chosen parameters, the crystal
grows more rapidly at some places and ‘‘fingers’’ begin to
grow, which may develop branches afterwards. So highly
ramified spatial structures can develop from a small part
of the initial surface.
As the grid points of the discretization are mainly moved
into the direction of the surface normal, the distances of
the vertices vary from one time step to the other. The
300 ALFRED SCHMIDT
FIG. 3.1. Degenerating grid (n 5 1).
distances get smaller at locally concave parts of the solid
phase and grow where the interface is convex.
In the three dimensional case (n 5 2), there may evolve
parts of the surface where the two principal curvatures
differ greatly (one positive and the other negative, for
example). The grid gets strongly deformed after some time.
The largest angles of some triangles may tend to 1808,
which would enlargen the errors in the numerical method.
We have to prevent the mesh from degenerating too much
and take special care of those triangles. Figures 3.1 and
3.2 show two examples of degenerate grids.
The discretization of the surface has to be adapted in
time, to be able to resolve the evolving structures. This is
done mainly by local refinement and coarsening, driven
by an a posteriori error estimate, just as described above
for the heat equation.
Obtuse angles which develop during deformation of the
mesh cannot be prevented by these methods alone, because
of the possible distortion of the mesh during the evolution.
To avoid a degenerate grid, we additionaly provide grid
modification procedures which are based mainly on angle-
dependent refinement and displacement of the vertices on
the surface.
Altogether, the algorithm is an adaptive h–r-method for
the discretization and movement of the curved free
boundary.
Error Estimation. Similar to the heat equation case, a
simple error estimate for the parametrization of the
surface can be derived. For each simplex T [ Gh let
nT (x), x [ T, be the unit normal of the surface. For
x [ ­T let n (x) denote the outer unit normal to T at
x in the tangential plane at x. It holds that n (x) ' nT (x),
and for surfaces (n 5 2) we have n 5 t ` nT with a
unit tangent vector t(x) to ­T. Let fh be the vector
valued parametrization of the surface over itself, so fh 5
id on T (for curves, this is equivalent to an arc length
parametrization), and let ­fh /­n be the component of
the covariant derivative of fh in direction n. In this
situation, the methods from [EJ] lead to a local L2 -error
estimate on a triangle T , Gh of the type
SO E UF GU SD
2
­fh
E
2
u
hT (fh) :5 ch h3 1 h4
E,­T
E ­n T «C
1 E
T
h4(Dfh 2 Ph Dfh)2 D1/2
.
We can replace u[­fh /­n]u by u[6n]u or by u[nT ]u, which is
the jump of surface normals at the edge. Since fh 5 id on
T, all these three terms give equal values. Using the last
one, and neglecting the integral which includes Dfh , we
define a local error indicator
SO E S DD
E
2 1/2
u
hT :5 ch h3u[nT ]u 2 1 h4 ,
E,­T
E T «C
where the special parametrization fh does not appear ex-
plicitly any more. Based on these values hT , we are able
to select regions of Gh where the triangulation should be
refined or may be coarsened, just as described in the last
section.
Adaptive h-Method. Refinement and coarsening of the
grid is done by a slightly modified version of a standard
triangle bisection algorithm [B1, Mi]. We use a combina-
tion of the ‘‘newest vertex’’ and ‘‘longest edge’’ refine-
ment strategies described in [Mi], which ensures that the
edge opposite the largest angle of a triangle gets bisected,
if this angle is larger than a given bound. Triangles are
refined not only where the error estimator tells us, but
additionally, when the largest angle of the triangles is
larger than a given bound for the largest angle allowed
in the triangulation. The selection of the refinement
edge mentioned above assures that the obtuse angle will
be divided.
Adaptive r-Method: Node Displacement on the Sur-
face. We cannot totally prevent obtuse angles and degen-
erate triangles with local refinement alone. To improve
the grid, we use another method: a displacement of the
grid nodes on the (discrete) surface, which is the relocation-
part of the adaptive h–r-method.
The coordinates of the vertices (and midpoints of the
edges) are changed in a way that makes the triangulation
better. We use a single-stepping algorithm that moves only
one vertex at a time, and loops through all nodes of the
grid. Depending on the quality criterion for the grid, the
direction of node movement is chosen.
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH
FIG. 3.2. Degenerate surface grid (n 5 2); part of a triangulation where triangles with very large angles developed.
For each node P of the triangulation, we define a func-
tional F(P) that measures the local mesh quality, and try
to maximize F by changing the coordinates of P. To
this aim, we calculate the gradient (d/dP)F(P) at the
old position P of the node and search the maximum of
F in the direction of the gradient, by a bisectioning line
search algorithm. As we want to vary the nodes only
on the surface, not in the whole space, we have to
project all points from the descent line onto the discrete
FIG. 3.3. Unit 5-polygon D5 and parametrization of a neighbourhood
of P.
301
surface during the maximization process. During the
iterations of the node movement algorithm, we have to
remember the old positions of all nodes, because we
need them for the projection onto the discrete surface
defined by these values.
We use a functional F, which works on the angles of
the triangulation; it minimizes the difference between
the N angles around a vertex P. It is based on the
conformal energy (see [Hu] for definition and results)
of a parametrization F : DN R {T , Gh : T > P ? B}
of the neighbourhood of P over the unit N-polygon DN
(compare Fig. 3.3). We set
F(P) :5 2 E DN
uFu ` Fvu 2 EDN
u=Fu 2.
F is maximal (5 0) if F is a conformal mapping; in this
case all angles around P are equal. Straightforward compu-
tation gives
302 ALFRED SCHMIDT

uQj 2 Pu 2 1 uQj11 2 Pu 2 Ch,1 :5 Ph(kCh , nhl)

F(P) 5 O1
N

j 51
2 2 cos(a)kQj 2 P, Qj11 2 Pl
2 sin(a) as the piecewise linear interpolate of the scalar product of
the discrete curvature vector with the approximate
normals.

2 u(Qj 2 P) ` (Qj11 2 P)u

2 and
Alternatively, we define a discrete curvature Ch,2 using
a discrete second fundamental form of Gh . For a smooth
manifold with parametrization F and normal n, the curva-

OS
ture is equal to the trace of the second fundamental form,
d N
1 2 cos(a)
F(P) 5 ((P 2 Qj11) 1 (P 2 Qj))
dP j 51 sin(a)
CG 5 Ogh ,
n
ij

S DD
ij
(Qj 2 P) ` (Qj11 2 P) i, j51
2 ` (Qj 2 Qj11) ,
u(Qj 2 P) ` (Qj11 2 P)u
where gi j 5 kDiF, DjFl, (g i j) 5 (gi j)21, and hi j 5 k2n,
Di Dj Fl 5 kDi n, DjFl. For each vertex P of the triangula-
where Qj , j 5 1, ..., N, are the neighbouring vertices to P
tion, we introduce a plane at P approximately tangent to
and a 5 2f/N.
Gh , project the triangles T around P onto that plane, and
We examined the node moving algorithm for some
denote the projected triangles by T̃. Now the neighbour-
model problems and compared the results to those ob-
hood of P, consisting of all adjacent triangles T, is the graph
tained with a simpler functional F̃, namely
of a function Fh , defined on the tangent plane. We set

F̃(P) :5 2 O uQ 2 Pu ,
j51
N
j
2 d
dP
O N
F̃(P) 5 2 (Qj 2 P);
j 51 (Di nh)h(P) :5 2
oT ]P eT̃ nh Di fP
,
oT ]P eT̃ fP
it minimizes the square sum of the lengths of all edges
around one vertex P. This functional is motivated by the oT ]P eT̃ Fh Dj fP
fact that the edge opposite an obtuse angle is much longer (DjFh)h(P) :5 2 ,
oT ]P eT̃ fP
than at least one of the other two edges of a triangle.
Comparison of the results obtained with both cost func-
tionals showed a smaller maximal angle in the triangula- where fP is the nodal basis function, and define a discrete
tions when using F (the first one). This justifies the larger second fundamental form at P:
amount of work that is necessary to compute this functional
and its derivative. hi j,h(P) :5 k(Di nh)h(P), (DjFh)h(P)l.

With a tangent plane at P, the metric form fulfils gi j (P) 5

3.3. Computation of the Curvature
di j , and we can easily define Ch,2 piecewise linear on all
The right-hand side of the heat equation (2.7) uses the triangles of Gh by
value of the mean curvature CG of the free boundary
G(tm11). There is no straightforward definition of a curva-
ture for Gmh
11
, as we approximate the free boundary G by Ch,2(P) :5 Oh
i51
n
ii,h(P).

a piecewise polynomial, globally Lipschitz-continuous dis-

crete free boundary Gm h
11
. In the sequel, we use the notation
Gh for the free boundary and drop the time step index. We
examined two different definitions of a discrete curvature
which fit well in the finite element context.
The first one is based on the identity CG n G 5 2DGF,
which is true for smooth manifolds with parametrization
F. We use a weak formulation and H21-projection to define
a vector-valued curvature Ch [ Wh(Gh)n11 by
(Ch , ch)L2(Gh) 5 (=GhFh , =Ghch)L2(Gh) for all ch [ Wh .
Using approximate normals nh at the grid vertices, the
discrete scalar mean curvature is now defined by
Applied to some model problems, the second definition
Ch,2 of a discrete curvature gives smaller errors between
the curvature of a smooth manifold and the discrete curva-
ture of its discretization.
3.4. Adaptive Algorithm for the Crystal Growth Problem
The adaptive methods for the heat equation and the
interface motion are combined, defining an algorithm for
the numerical solution of Problem 1.1.
In each time step, given the old values u hm and Ghm , we
first move the interface Gh to a new position G̃m
h
11
by solving
m
Eq. (2.3) with temperature u h on the right hand side:
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 303

E «V Fm (x) 2 x
E E E
11
1 0
=u 0h =w 1 u hwh
h
ch 1 =Fm 11
=ch D0
m
Gh «C tm m
Gh
h
V G0h «V

52 E m
Gh
1 m
«C
u h nch for all ch [ Wh(Ghm). 52 E V
ut,0wh 2 EG0h
«C
CG0 wh for all wh [ V 0h,0 ,
«V h
(3.1)

u 0h 5 P 0(u0(t0)) on ­V,
The triangulation of the interface is adaptively refined,
coarsened, and changed according to the adaptive h–r-
method from Section 3.2, which results in the new free with given data G0h , u0 , and ut,0 . Besides the generation of
boundary Gm 11
. The curvature C m 11
of Gm 11
is computed
h h h
T 0h , defining the temperature u 0h as the solution of Eq.
as described in Section 3.3. Now, the new temperature
11 11
(3.1) ensures that it is compatible with the discrete inter-
grid T m and the new temperature u m are computed
h h
face G0h , its curvature C G0h , and the quadrature formulas
by the semi-implicit adaptive strategy from Section 3.1 for
used.
the solution of Eq. (2.7) using integrals over Gm h
11
and
In Fig. 0.1, a flow diagram shows the whole adaptive
curvature C h on the right hand side, solving for all wh [
m11
11 algorithm, including the generation of an initial tempera-
Vmh,0
ture grid. This algorithm is easily supplemented by a higher
order time discretization (Richardson extrapolation is im-
E 2 u hm
E E
11
um
h 1 m11
wh 1 D0 =u m 11
=wh 1 u h wh plemented) and adaptive time step control (using the para-
V tm V
h m11
Gh «V bolic error estimates and methods from [EJ], for example).

E «C m11 Because of technical details (more intermediate solutions

52 C h wh . have to be computed and combined), the algorithm loses
m11
Gh «V
a bit of its simple structure in these cases, and we do not
11 present details here.
The temperature meshes T m h and the surface meshes
Gh are both adaptively generated using information only
m11

from the corresponding a posteriori error estimates (and

the interface motion); no direct relationship between
them exists.
All integrals over the free boundary are computed by
quadrature formulas. The coordinates of the quadrature
points on the (curved) surface are computed easily from
their position in the reference simplex using the local
parametrization of the discrete interface. For integrals
which involve uh or another function from the finite
element space on the (n 1 1) dimensional temperature
mesh, the positions of these quadrature points in the
temperature mesh (simplex number and barycentric coor-
dinates) can be found using local search operations,
when the positions of the surface grid vertices in the
temperature mesh are (approximately) known, for exam-
ple from the last time step. Such a local search is done
by running from one simplex to an adjacent one which
is closer to the point, until all barycentric coordinates
are nonnegative. Some global search operations are
needed only once before the first time step, but the
same technique can be used, and their number may be
minimized using neighbourhood information about the
4. NUMERICAL RESULTS
We implemented a two and a three dimensional version
of the adaptive algorithm, using piecewise linear and qua-
dratic finite elements for both the temperature and free
boundary discretization. First results of our method in two
space dimensions, showing different growth morphologies
(tip splitting and dendritic growth) as well as fourfold and
sixfold symmetric solutions, have already been presented
in [BS].
4.1. Convergence of the Numerical Method
We compare the solution of our numerical method
with the known exact solution in case of an isotropic
model problem. With given positive constants «, l,
R0 , and
2«(l 1 n)
R(t) :5 ÏR 20 1 2lt, U(t) :5 ,
R(t)

E
2
vertices on the surface. s e2(l /2)z «l(l 1 n)
The initial temperature grid T 0h is generated by an appli- T(s) :5 2le l / 2 dz, f (t) :5 U̇(t) 5 ,
1 zn R(t)3
cation of the adaptive finite element method to an elliptic
Problem 2.4 with a given time derivative ut,0 on the right
hand side. Thus, the initial temperature u 0h [ V 0h is defined where the temperature u and the free boundary G are
as the solution of given by
304 ALFRED SCHMIDT

x [ Rn11, uxu # R(t),

5
U(t)
u(x, t) :5
U(t) 1 T S Duxu
R(t)
x [ Rn11, uxu . R(t),

G(t) :5 hx [ Rn11, uxu 5 R(t)j,

u and G are solutions of the system of equations

u̇ 2 Du 5 f for t $ 0, x [ Rn11 \G(t),

FG
­u
­n
5 2VG(t) on G(t),

«VG(t) 1 «CG(t) 1 u 5 0 on G(t),

where VG(t) 5 Ṙ(t) 5 l /R(t) and CG(t) 5 n/R(t) are the

velocity and mean curvature of the free boundary.
First, we want to compare the estimated errors from the
solution of the initial value equation (3.1) with the exact
errors. For « 5 0.001, l 5 0.5, R0 5 0.5 and V 5 (24,
4)n11, Fig. 4.1 shows the error iu(?, 0) 2 u 0hiL2(V) and the
corresponding a posteriori estimate h(u 0h) for the three
dimensional model problems, with global resp. adaptive
local refinement of the temperature mesh. We see that the
error estimate is exact up to a constant. The estimated
error is much larger than the exact one, because we used
a coefficient ch 5 1 for the computation of the values
hT (uh). Adjustment of this coefficient leads to a rather
exact estimate. For the two dimensional problem, we get
FIG. 4.2. Temperature L2-error and number of unknowns.
similar results. Figure 4.2 compares the exact error and
the number of unknowns from global resp. adaptive local
refinement. Here, we see the large advantage of local over
global refinement. Using adaptive local refinement in three
dimensions, we need less than 10% of unknowns to reach
an error iu(?, 0) 2 u 0hiL2(V) P 0.05, for example, and this
factor gets larger when the error gets smaller. This advan-
tage will be even greater in the case of the ‘‘real’’ dendrite
growth problem with larger curvature and interface veloc-
ity, which directly affect the temperature and thus the error
near the free boundary. Without this advantage of the
adaptive method, we would not be able to compute three
dimensional dendrites.
Similar results are obtained for the a posteriori estimate
of the free boundary error iG 2 GhiL2 .
Figures 4.3 and 4.4 present the temperature and interface
errors over the time interval [0, 1], they indicate conver-
gence of the method. The computations in three dimen-
sions were done using piecewise quadratic elements for
both temperature and interface approximation and adap-
tive local refinement and coarsening. In Table 4.1, we show
the timestep t and the parameters d and d used in the
adaptive method for the temperature mesh as well as the
parameter d G used for refinement of the discrete interface.

4.2. Two Dimensional Results

The problem is solved in the domain V 5 (28, 8)2 with
diffusion constant D0 5 1 and undercooling boundary val-
FIG. 4.1. Exact and estimated temperature L2-error. ues u0 5 20.5 on ­V. We use anisotropic coefficients
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 305

FIG. 4.3. Temperature L2-error. FIG. 4.4. Interface L2-error.

«C(n) 5 « C (1 1 AC cos(f0 1 kf)),
«V (n) 5 « V (1 1 AV cos(f0 1 kf)),
where f is the angle between the normal n and the x1-axis,
and k determines the symmetry. Parameters are chosen as
«C 5 0.001, «V 5 0.01, AC 5 AV 5 20.2, with (k 5 4)-fold
symmetry. The initial temperature u 0h is computed from
Eq. (3.1) using an interface G0 5 {x [ R2, uxu 5 R0} and
2 solution computed with the smallest mesh and time step
2u0V 2e2(u x uV / 2R0)
ut,0(x) 5 if uxu . R0 ,
y
R 20 2 E e2s
V /4 s
ds
ut,0(x) 5 0 if uxu # R0
with V 5 1.56 and R0 5 0.05, which is the time derivative of
the temperature field around a circular growing interface;
compare [F].
The same problem is solved with different time step
sizes and error tolerances d, d G for mesh adaptation to see
their influence on the numerical solution; see Table 4.2. All
computations were done with the Euler time discretization
described above and started with initial data taken at time
t 5 0.05 from the run with parameter set E2. Results are
presented in Fig. 4.5. We show the interface Gh at times
t 5 0.05, 0.15, 0.25, ..., 0.95, with the first curve being the
same for all runs, as explained above. The outer square
depicts the subdomain (24, 4)2. We see clearly that the
sizes (which should be more accurate than the others)
develops nearly no sidebranches, and branch size increases
with growing parameters t, d. The sidebranching instability
is amplified by coarser discretizations.
Next, we examine the behaviour of the dendrite tip ve-
locities, curvatures, and the associated Peclet numbers.
The Peclet number of a dendrite tip is defined by
rV
p :5 ,
2D0

TABLE 4.1
TABLE 4.2
Parameters for the Adaptive Method
Discretization Parameters for Curves E1–E6
Error curve no. t d d dG
Curve no. E1 E2 E3 E4 E5 E6
1 0.1 0.1 1.0 0.01
2 0.05 0.05 0.5 0.005 t 0.00025 0.0005 0.001 0.002 0.005 0.010
3 0.025 0.025 0.25 0.0025 d 0.1 0.1 0.2 0.4 0.6 0.8
4 0.0125 0.0125 0.125 0.00125 dG 0.25 0.25 0.25 0.50 0.70 1.00
306 ALFRED SCHMIDT

FIG. 4.5. Evolution of the interfaces, parameter sets E1–E6.

 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH
FIG. 4.6. Tip velocity V.
where r is the radius of curvature at a tip and V denotes
its velocity. In the case of zero surface tension, «C 5 0,
exact solutions—the Ivantsov parabolas—are known [I].
For a given undercooling uy at infinity, all parabolas x 5
2y2 /2r 1 Vt with the same Peclet number p 5 p(uy) give
self-similar solutions of the dendrite growth problem. Even
for non-zero surface tension, the dendrite tips are nearly
parabolic and behave in a similar way, and their Peclet
number should be nearly equal to the Ivantsov value p(uy).
The Peclet value for an undercooling uy 5 20.5 is
p(20.5) P 0.187.
In Figs. 4.6, 4.8, and 4.9, we present the tip velocities,
tip radii, and Peclet numbers from the six runs of our
program. At each time step, these numbers are mean values
from the four dendrite tips. Additionally, the curves were
FIG. 4.7. Original tip velocity curve E1.
307
FIG. 4.8. Tip radius r.
smoothed. This was done because temporal variations of
the tip velocity are quite large (see Fig. 4.7 for an original
curve). Parabola were least-squares fitted against the tips
to compute r. Tip velocity and radius both show conver-
gence. As a result, we obtain convergence for the Peclet
number, which additionally depends less on the discretiza-
tion parameters. Velocity increases and tip radius de-
creases with mesh refinement, while their product remains
nearly the same.
All the above examples show computations, where the
symmetry of the problem is the same as the symmetry of
the temperature mesh: based on the initial macro triangula-
FIG. 4.9. Peclet number p.
308 ALFRED SCHMIDT
FIG. 4.10. Interface with 6-fold symmetry.
tion of the domain, all edges of triangles are either hori-
zontal, vertical, or diagonal at a 458 angle. The next two
pictures show the simulation with 6-fold symmetry in the
anisotropy functions «C and «V , whereas the temperature
mesh is still built out of triangles with only 908 and 458
angles. Figure 4.10 shows the evolving interface, and Fig.
4.11 the temperature triangulation from the final time step.
As all six fingers grow quite similarly, the influence of the
mesh orientation is very small.
4.3. Three Dimensional Results
We present numerical results in V 5 (24, 4)3 with an
external undercooling of u0(x) 5 20.5 on ­V and a spheri-
cal initial interface G0 5 {x [ R3, uxu 5 0.1}. Following
[K2], we use the anisotropic coefficients
«C (n) 5 «V (n) 5 «(1 2 A(1 2 (n 41 1 n 42 1 n 43))), n [ S 2.
The preferred growth directions are vectors n whose value
«C (n) is relatively small. Using a parameter A . 0, the
eight vectors (61, 61, 61) point to the preferred growth
directions, and with A , 0, dendrites grow in the six direc-
tions of the coordinate axes. We used the values A 5 23.0
and « 5 0.001, which result in an interface which develops
six dendritic arms.
The initial temperature uh,0 was computed using Eq.
(3.1) with right hand side
2
2u0Ve2(u x uV / 2R0)
ut,0(x) 5 y if uxu . R0 ,
/V 2 eV / 2 e2s ds)
2 2
2R0uxu(e2V /4
ut,0(x) 5 0 else.
This choice of ut,0 gives a temperature field u0 around a
growing sphere (with isotropic coefficients, compare [F]),
analogous to the two dimensional case. Here, the radius
of the initial interface is R0 5 0.1, and V 5 V (uy) with
V(20.5) P 2.076.
In Fig. 4.12, we show an evolving interface after 0, 100,
200, 300, 400, and 500 time steps ranging over the time
interval [0, 0.5]. The computations were done with a time
discretization using Richardson extrapolation and piece-
wise quadratic ansatz functions for both the temperature
and interface approximation.
The next picture, Fig. 4.13, presents a part of the free
boundary discretization after 500 time steps, which consists
of 82,696 triangles, parametrized with piecewise quadratic
polynomials. Due to the mesh improvement algorithms
described in Section 3.2, the development of obtuse angles
is successfully prevented.
Figures 4.14 and 4.15 present the traces of the three
dimensional temperature mesh on the plane z 5 0 after
280 resp. 500 time steps. The adaptive mesh refinement
and coarsening, driven by the a posteriori error estimates,
lead to a highly refined grid just near the free boundary,
especially near the dendrite tips, where both velocity and
curvature of the interface are big, resulting in steep temper-
FIG. 4.11. Final temperature mesh.
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH 309

FIG. 4.12. Evolution of the free boundary after 0, 100, 200, 300, 400, 500 time steps.

FIG. 4.13. Detail of the free boundary discretization after 500 time FIG. 4.14. Trace of the temperature mesh after 280 time steps with
steps. the plane z 5 0.0.
310 ALFRED SCHMIDT

FIG. 4.17. One interface computed with symmetry.

FIG. 4.15. Trace of the temperature mesh after 500 time steps with
the plane z 5 0.0.

FIG. 4.16. Interface together with a slice through the temperature mesh after 500 time steps.
 FINITE ELEMENT COMPUTATION OF DENDRITE GROWTH
FIG. 4.18. Another interface computed with symmetry.
ature gradients. Inside the solid phase, the temperature
is nearly constant and the mesh gets recoarsened by the
adaptive method. Finally, the interface is shown in Fig.
4.16 together with a slice through the temperature grid
after 500 time steps. The free boundary nearly reached the
domain’s boundary.
The last two pictures (Figs. 4.17 and 4.18) show some
results from simulations with different parameter sets.
These computations were done in only one cube octant
with symmetry boundary conditions for all equations. They
used a maximum of 60,000 tetrahedra for temperature
discretization and needed two resp. one day(s) of a work-
station’s computing time.
ACKNOWLEDGMENTS
I thank Gerhard Dziuk as well as Eberhard Bänsch, Klaus Deckelnick,
and Kunibert G. Siebert for stimulating discussions, many ideas, and their
support. The heat equation parts of our programs are based on E. Bänsch’s
adaptive 2D and 3D heat equation solvers [B4, B5]. All three dimensional
pictures were created using the Grape visualization package [GR, RS],
and XGraph [Ha] was used for the diagrams.
REFERENCES
[A] R. Almgren, J. Comput. Phys. 106, 337 (1993).
[B1] E. Bänsch, IMPACT Comput. Sci. Eng. 3, 181 (1991).
[B2] E. Bänsch, J. Comput. Appl. Math. 36, 3 (1991).
[B3] E. Bänsch, in Adaptive Finite and Boundary Element Methods, ed-
311
ited by C. A. Brebbia and M. H. Aliabadi, Computational Mechanics
Publication (Southhampton, Boston, 1993), p. 47.
[B4] E. Bänsch, Adaptive 2D heat equation solver, Fortran program,
1988 (unpublished).
[B5] E. Bänsch, Adaptive 3D heat equation solver, Fortran program,
1990 (unpublished).
[BS] E. Bänsch, A. Schmidt, in Computational Crystal Growers Work-
shop, edited by J. E. Taylor, AMS Selected Lectures in Mathematics
(Amer. Math. Soc., Providence, 1992), p. 16.
[BJGLS] E. Ben-Jacob, N. Goldenfeld, J. S. Langer, and G. Schön, Phys-
ica D 12, 245 (1984).
[CL] G. Caginalp and J. T. Lin, IMA J. Appl. Math. 39, 51 (1987).
[CR] X. Chen and F. Reitich, J. Math. Anal. Appl. 164, 350 (1992).
[Ch] A. J. Chorin, J. Comput. Phys. 58, 472 (1985).
[Ci] P. G. Ciarlet, The Finite Element Method for Elliptic Problems,
(North-Holland, Amsterdam, 1980).
[D1] G. Dziuk, Numer. Math. 58, 603 (1991).
[D2] G. Dziuk, in Computational Crystal Growers Workshop, edited by
J. E. Taylor, AMS Selected Lectures in Mathematics (Amer. Math.
Soc., Providence, 1992), p. 34.
[EJ] K. Eriksson and C. Johnson, SIAM J. Numer. Anal. 28, 43 (1991).
[FL] G. J. Fix and J. T. Lin, Nonlinear Anal. Theory Methods Appl. 12,
811 (1988).
[F] G. C. Frank, Proc. Roy. Soc. A 201, 586 (1950).
[GR] M. Geiben and M. Rumpf, in Advances in Scientific Visualization,
edited by F. H. Post and A. J. Hin, Lecture Notes in Physics (Springer-
Verlag, Berlin/New York, 1993).
[G] M. E. Gurtin, Arch. Rational Mech. Anal. 100, 275 (1988).
[Ha] D. Harrison, XGraph, Version 11, UC Berkeley (1989).
[Har] A. Harnisch, Dissertation (Essen), Berichte aus der Metallurgie
(Shaker Verlag, Aachen, 1994).
[He] D. Hempel, IMPACT Comput. Sci. Eng. 5, 309 (1993).
[HK] R. H. W. Hoppe and R. Kornhuber, in Multigrid Methods, Lecture
Notes in Pure Applied Mathematics, Vol. 110 (Dekker, New York,
1988), p. 267.
[Hu] J. E. Hutchinson, Proc. Centre Math. Anal. Austral. Nat. Univ. 26,
140 (1991).
[I] G. P. Ivantsov, Dokl. Akad. Nauk SSSR 58, 567 (1947).
[KW] K. Kalik and W. Wendland, Computing 47, 255 (1992).
[K1] R. Kobayashi, Physica D 63, 410 (1993).
[K2] R. Kobayashi, Exper. Math. 3, 59 (1994).
[L] J. S. Langer, Rev. Mod. Phys. 52, 1 (1980).
[Me] D. I. Meiron, Phys. Rev. A 33, 2704 (1986).
[Mi] W. F. Mitchell, ACM Trans. Math. Software 15, 326 (1989).
[N] J. C. Nedelec, Comput. Meth. Appl. Mech. Engrg. 8, 61 (1976).
[NPV1] R. H. Nochetto, M. Paolini, and C. Verdi, Math. Comput. 57,
73 and supplement S1 (1991).
[NPV2] R. H. Nochetto, M. Paolini, and C. Verdi, SIAM J. Sci. Statist.
Comput. 12, 1207 (1991).
[RT] A. Roosen and J. E. Taylor, Research Report GCG36, The Geome-
try Center, University of Minnesota, 1992 (unpublished), to appear in
Interface Dynamics and Growth, MRS Sym. Proc. Ser. 237.
[RS] M. Rumpf, A. Schmidt, et al., Report 8, SFB 256, Bonn (1990).
[SGWMK] Y. Saito, G. Goldbeck-Wood, and H. Müller-Krumbhaar,
Phys. Rev. A 38, 2148 (1988).
[SS] J. A. Sethian and J. Strain, J. Comput. Phys. 98, 231 (1992).
[Sm] J. B. Smith, J. Comput. Phys. 39, 112 (1981).
312 ALFRED SCHMIDT

[St] J. Strain, in Computational Crystal Growers Workshop, edited by
J. E. Taylor, AMS Selected Lectures in Mathematics (Amer. Math.
Soc., Providence, 1992), p. 104.
[SLN] J. M. Sullivan, Jr., D. R. Lynch, and K. O’Neill, J. Comput. Physics
69, 81 (1987).
[TH] K.-H. Tacke and A. Harnisch, in Computational Modelling of Free
and Moving Boundary Problems, Vol. 2. Heat Transfer, edited by L. C.
Wrobel and C. A. Brebbia, Computational Mechanics Publication
(Southhampton, Boston, 1991), p. 165.
[WMS] A. A. Wheeler, B. T. Murray, and R. J. Schaefer, Physica D 66,
243 (1993).
[Wl] J. Wloka, Partielle Differentialgleichungen: Sobolevräume und
Randwertaufgaben (Teubner, Stuttgart, 1982), and Partial Differential
Equations (Cambridge Univ. Press, London/New York, 1987).