Lecture Notes in Physics: Monographs

Download as pdf or txt
Download as pdf or txt
You are on page 1of 200

Lecture Notes in Physics

Monographs
Editorial Board
R. Beig, Wien, Austria
B.-G. Englert, Ismaning, Germany
U. Frisch, Nice, France
P. Hanggi, Augsburg, Germany
K. Hepp, Zurich, Switzerland
W. Hillebrandt, Garching, Germany
D. Imboden, Zurich, Switzerland
R. L. Jaffe, Cambridge, MA, USA
R. Lipowsky, Golm, Germany
H. v. Lohneysen, Karlsruhe, Germany
I. Ojima, Kyoto, Japan
D. Sornette, Nice, France, and Los Angeles, CA, USA
S. Theisen, Golm, Germany
W. Weise, Trento, Italy, and Garching, Germany
J. Wess, Munchen, Germany
J. Zittartz, Koln, Germany
Managing Editor
W. Beiglbock
c/o Springer-Verlag, Physics Editorial Department II
Tiergartenstrasse 17, 69121 Heidelberg, Germany

3
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Tokyo

The Editorial Policy for Monographs


The series Lecture Notes in Physics reports new developments in physical research and
teaching - quickly, informally, and at a high level. The type of material considered for publication in the monograph Series includes monographs presenting original research or new
angles in a classical field. The timeliness of a manuscript is more important than its form,
which may be preliminary or tentative. Manuscripts should be reasonably self-contained.
They will often present not only results of the author(s) but also related work by other
people and will provide sufficient motivation, examples, and applications.
The manuscripts or a detailed description thereof should be submitted either to one of
the series editors or to the managing editor. The proposal is then carefully refereed. A
final decision concerning publication can often only be made on the basis of the complete
manuscript, but otherwise the editors will try to make a preliminary decision as definite
as they can on the basis of the available information.
Manuscripts should be no less than 100 and preferably no more than 400 pages in length.
Final manuscripts should be in English. They should include a table of contents and an
informative introduction accessible also to readers not particularly familiar with the topic
treated. Authors are free to use the material in other publications. However, if extensive use
is made elsewhere, the publisher should be informed. Authors receive jointly 30 complimentary copies of their book. They are entitled to purchase further copies of their book at
a reduced rate. No reprints of individual contributions can be supplied. No royalty is paid
on Lecture Notes in Physics volumes. Commitment to publish is made by letter of interest
rather than by signing a formal contract. Springer-Verlag secures the copyright for each
volume.

The Production Process


The books are hardbound, and quality paper appropriate to the needs of the author(s) is
used. Publication time is about ten weeks. More than twenty years of experience guarantee authors the best possible service. To reach the goal of rapid publication at a low price
the technique of photographic reproduction from a camera-ready manuscript was chosen.
This process shifts the main responsibility for the technical quality considerably from the
publisher to the author. We therefore urge all authors to observe very carefully our guidelines for the preparation of camera-ready manuscripts, which we will supply on request.
This applies especially to the quality of figures and halftones submitted for publication.
Figures should be submitted as originals or glossy prints, as very often Xerox copies are not
suitable for reproduction. For the same reason, any writing within figures should not be
smaller than 2.5 mm. It might be useful to look at some of the volumes already published or,
especially if some atypical text is planned, to write to the Physics Editorial Department of
Springer-Verlag direct. This avoids mistakes and time-consuming correspondence during
the production period.
As a special service, we offer free of charge LATEX and TEX macro packages to format the
text according to Springer-Verlags quality requirements. We strongly recommend authors
to make use of this offer, as the result will be a book of considerably improved technical
quality.
For further information please contact Springer-Verlag, Physics Editorial Department II,
Tiergartenstrasse 17, D-69121 Heidelberg, Germany.

Series homepage -- http://www.springer.de/phys/books/lnpm

Heike Emmerich

The Diffuse
Interface Approach
in Materials Science
Thermodynamic Concepts
and Applications
of Phase-Field Models

13

Author
Heike Emmerich
Lehrstuhl fur Thermodynamik
Fachbereich Chemietechnik
Universitat Dortmund
Emil-Figge-Strasse 70
44227 Dortmund, Germany

Cover picture: Columnar dendrites: The shading figure denotes the variation of temperature in the melt. Typically columnar dendrites grow at the container walls as
well as at the macroscopic liquid-solid interface, whereas so called equiaxed crystals
develop in the inner regions of the molten phase. (by H. Emmerich)
Cataloging-in-Publication Data applied for
A catalog record for this book is available from the Library of Congress.
Bibliographic information published by Die Deutsche Bibliothek
Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data is available in the Internet at http://dnb.ddb.de
ISSN 0940-7677 (Lecture Notes in Physics. Monographs)
ISBN 3-540-00416-5 Springer-Verlag Berlin Heidelberg New York
This work is subject to copyright. All rights are reserved, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and
storage in data banks. Duplication of this publication or parts thereof is permitted only
under the provisions of the German Copyright Law of September 9, 1965, in its current
version, and permission for use must always be obtained from Springer-Verlag. Violations
are liable for prosecution under the German Copyright Law.
Springer-Verlag Berlin Heidelberg New York
a member of BertelsmannSpringer Science+Business Media GmbH
http://www.springer.de
c Springer-Verlag Berlin Heidelberg 2003

Printed in Germany
The use of general descriptive names, registered names, trademarks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
Typesetting by the authors/editors
Camera-data conversion by Steingraeber Satztechnik GmbH Heidelberg
Cover design: design & production, Heidelberg
Printed on acid-free paper

55/3141/du - 5 4 3 2 1 0

Acknowledgements

At this point I would like to thank Prof. Dr. Schreiber for giving me the
opportunity to teach at Chemnitz and to integrate my work in a university
chairs. I am also thankful for his detailed comments on the manuscript helping me to improve its readability in several points.
Moreover I would like to thank Prof. Dr. Kassner for introducing me to the
eld of simulating growth phenomena in materials science. Prof. Dr. M
ullerKrumbhaar, Dr. Brener, Dr. Misbah and Dr. Ihle gave valuable guidance,
which I am thankful for.
I am also grateful to Prof. Dr. Fulde and Prof. Dr. Rost for the stimulation
and the opportunities which arose from being a Distinguished Postdoctoral
Fellow at the Max Planck Institute for the Physics of Complex Phenomena
at Dresden.
Finally I would like to thank Prof. Dr. Sadowski for oering the quite
challenging new task to continue in the eld of materials science by bridging
the gap from the mesoscale to the atomar scale rising a respective simulation
group at Dortmund university.
Dortmund, November 2002

Heike Emmerich

Contents

1.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Structure and Scope of This Work . . . . . . . . . . . . . . . . . . . . . . . .

1
4

2.

What Is an Interface? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.

Equilibrium Thermodynamics of Multiphase Systems:


Thermodynamic Potentials and Phase Diagrams . . . . . . . . . .
3.1 Calculating Phase Diagrams
from Energy Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Abstracted Phase Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Constructing the Gibbs Free Energies . . . . . . . . . . . . . . .
3.2.2 Non-conserved Versus Conserved Variable Approach . .

21
23
26
28

4.

Thermodynamic Concepts of Phase-Field Modeling . . . . . . .


4.1 Derivation of Transport Equations . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Considering Conserved Quantities Only . . . . . . . . . . . . .
4.1.2 Extension to Non-conserved Quantities . . . . . . . . . . . . . .
4.2 Introducing the Phase-Field Variable . . . . . . . . . . . . . . . . . . . .
4.3 Thermodynamic Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 The Gradient Flow Method . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 Entropy Production in Terms of Transport Variables . .
4.4 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31
32
32
35
38
46
47
49
53

5.

Asymptotic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1 A Formal Mathematical Approach Towards Matching . . . . . . .
5.1.1 Illustration of Extended Domains . . . . . . . . . . . . . . . . . . .
5.1.2 Matching in the Context of Diuse Interface Modeling
5.2 Asymptotic Matching for Thin Film Epitaxial Growth . . . . . .
5.2.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 Sharp Interface Formulation . . . . . . . . . . . . . . . . . . . . . . .
5.2.3 The Diuse Interface Model Equations
and Their Asymptotic Analysis . . . . . . . . . . . . . . . . . . . .
5.3 A Generalized Approach Towards the Asymptotic Analysis
of Diuse Interface Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

59
59
63
67
68
69
70

19

72
78

VIII

Contents

5.3.1 Role of the Governing Thermodynamic Potential . . . . .


5.3.2 The Expansion Procedure in Detail . . . . . . . . . . . . . . . . .
5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.

78
81
92
93

Application of Diuse Interface Modeling


to Hydrodynamically Driven Growth . . . . . . . . . . . . . . . . . . . . .
6.1 Diuse Interface Modeling
for Hydrodynamically Inuenced Growth . . . . . . . . . . . . . . . . . .
6.2 The Selection Problem of Dendritic Growth Revisited . . . . . . .
6.3 Comparison to Experimental Data . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

99
104
119
127
129

7.

Application to Epitaxial Growth Involving Elasticity . . . . . .


7.1 Elastic Driving Forces within the Diuse Interface Approach .
7.2 Comparing Simulations and Experiments . . . . . . . . . . . . . . . . . .
7.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

131
132
134
140

8.

Conclusions and Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

97

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
A. Numerical Issues of Diuse Interface Modeling . . . . . . . . . . .
A.1 Relationship Between Physical Variables
and Numerical Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Computational Universality of Phase Field Models . . . . . . . . . .
A.3 Selected State-of-the-Art Numerical Approaches . . . . . . . . . . . .
A.3.1 2D Adaptive Mesh Renement Computation . . . . . . . . .
A.3.2 3D Simulations Including Fluid Flow . . . . . . . . . . . . . . .

145
146
151
154
155
161

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

1. Introduction

Many inhomogeneous systems involve domains of well-dened phases separated by a distinct interface. If they are driven out of equilibrium one phase
will grow at the cost of the other. Examples are phase separation by spinodal decomposition or nucleation and subsequent growth of the nucleus in
the nourishing phase [139]. Another example which has often been discussed
as a paradigmatic problem is that of dendritic solidication [29, 64, 79, 199].
The phenomenological description of these phenomena involves the denition of a precisely located interfacial surface on which boundary conditions
are imposed. One of those boundary conditions typically yields a normal velocity at which the interface is moving. This is the so-called sharp interface
approach, adopted both in analytical and numerical studies for a variety of
contexts involving a moving boundary. The origin of such a description is often transparent, being obtained by symmetry arguments and common sense.
Nevertheless the properties of sharp interface models can be quite subtle as in
the case for dendritic growth. This is strongly coupled to the question of how
to view the interfacial surface. Already when introducing the notion of a surface quantity Gibbs implicitly entertained the idea of a diuse interface [126]:
any density of an extensive quantity (e.g., the mass density) between two coexisting phases varies smoothly from its value in one phase to its value in the
other. The existence of a transition zone, though microscopically of atomic extent, underlies this denition of surface quantities as given by Gibbs. In phase
transition phenomena, this notion has been employed in the spirit of Landau
and Khalatnikov [271], who were the rst to introduce an additional parameter to label the dierent phases in their theory on the absorption of liquid
helium. Essentially diuse interface modeling, as it appeared subsequently
in the literature in the context of phase transition phenomena [54, 143], is
connected to such an additional order parameter. Clearly such models have
advanced numerical treatment as well as understanding of interfacial growth
phenomena since.
Even though quite a young approach to tackle such problems, diuse interface models have been employed by dierent groups in quite dierent spirits.
One might even be tempted to say that a variety of philosophies accompanying diuse interface modeling have already emerged. One way to view this
method to model interfacial growth is to understand it as a numerical tech-

Heike Emmerich (Ed.): LNP m73, pp. 15, 2003.


c Springer-Verlag Berlin Heidelberg 2003


1. Introduction

nique, which helps to overcome the necessity of solving for the precise location
of the interfacial surface explicitly in each time step of a numerical simulation.
This can be achieved by the introduction of one or several additional phaseeld variables. They are the key element to the resulting phase-eld modeling
approach for studying systems out of equilibrium. In such an approach the
phase-eld variables are continuous elds which are functions of space r and
time t. They are introduced to describe the dierent relevant phases. Typically these elds vary slowly in bulk regions and rapidly, on length scales of
the order of the correlation length , near interfaces. is also a measure for
the nite thickness of the interface. The free energy functional A determines
the phase behavior. Together with the equations of motion this yields a complete description of the evolution of the system. In other contexts, such as
critical dynamics [41, 139, 146], the elds are order parameters distinguishing
the dierent phases. In a binary alloy, for example, the local concentration or
sub-lattice concentration can be described by such elds. The ideas involved
in this approach have a long history, referring back to van der Waals [244].
On this background the materials science community associates the use
of continuum eld models in particular with the work of Cahn and collaborators [7, 54, 55]. Within their contribution to the eld, phase-eld models are
more but just a trick to overcome numerical diculties. Rather they are
rigorously derived based on the variational principles of irreversible thermodynamics as founded by Onsager [227]. Then ensuring thermodynamic consistency of the model equations can serve as a justication of a phase-eld
model. In this sense phase-eld models can also be formulated for problems,
for which sharp interface equations are not yet available. Consequently it
might be their analysis which yields a formerly unknown sharp interface formulation and helps to clarify the physics in the interfacial region.
One has to contrast this procedure to a very established second way to
validate a phase-eld model. This second approach assumes, that a given
sharp interface formulation of the growth problem is the correct description of
the physics under consideration. On the basis of this assumption, a phase-eld
model can be justied by simply showing that it is asymptotic to the correct
sharp interface description, i.e., that the latter arises as the sharp interface
limit of the phase-eld model when the interface width is taken to zero.
Obviously this procedure works only for cases, in which a well established
set of continuum equations describing the dynamics in the sharp interface
formulation does exist. Moreover, employed in this way phase-eld models
do not seem to be much help to elucidate the physics of the interfacial region
beyond what is captured within the sharp interface model equations.
However, the latter is only partially true and leads to a third philosophy
appearing in the phase-eld community lately. It is rooted in the understanding of the interfacial surface to be nite in the sense of Gibbs denoted
above: If one assumes a phase-eld model to be thermodynamically consistent
and to describe a physical situation, for which an established sharp interface

1. Introduction

formulation exists, as well, then, certainly, in the sharp interface limit the
phase-eld model should correspond precisely to that sharp interface formulation. However, keeping in mind that the interface can be understood to
be of nite width, not only the sharp interface limit of a phase-eld model
is a meaningful physical limit, but also the so-called thin interface limit introduced by Karma and Rappel [168, 169]. To clarify the dierence between
the sharp interface limit and this thin interface limit here I will consider
the growth of a dendrite with tip radius R into an undercooled melt [128].
Under more general circumstances, R might be representative of a typical
macroscopic length scale such as the container size. For dendritic solidication at large undercoolings the growth is rapid and the radius of curvature
of the dendritic tip is relatively small. As a consequence eects of capillary
action and kinetics on the local interfacial temperature can be signicant.
In this regime, sharp interface limits of the phase-eld equations have been
performed [49, 60, 115, 116, 121, 213], which assume that the dimensionless
interfacial temperature u is of the order of the small parameter /R. Contributions from capillary eects and kinetics can be regarded to be of the same
order. In this limit one also considers to be small compared to the capillary
length lc , which presents a stringent resolution requirement for a numerical
computation that aspires to describe this limiting case. At low undercoolings,
on the other hand, dendrites grow more slowly and have a larger radius of
curvature, so that it is reasonable to model capillary eects and kinetics as
small corrections. Karma and Rappel refer to the corresponding analysis as
the thin interface limit. For this thin interface limit one assumes  R but
allows lc . Almgren [11] has described this analysis as isothermal asymptotics, since to leading-order in /R the temperature is isothermal throughout
the interfacial region with u = O(/R).
Now again one interest in employing such an isothermal asymptotics or
thin interface limit can be understood to be of numerical origin: it can serve
to legitimate a choice of model parameters which ensures better1 numerical
performance. On the other hand, isothermal asymptotics can also be used to
obtain rst order generalizations of the well known Gibbs-Thompson relation,
which usually yields the temperature value locally at the interface. In turn
such a generalization can facilitate subsequent stability analysis of the model.
Thus momentarily diuse interface modeling is a eld, in which numerical eorts as well as an intense focus on thermodynamic backgrounds and
asymptotic behavior of the models drives the development to turn this approach into a more and more powerful technique. Within the present work I
will elucidate this symbiotic framework in more detail with respect to specic
growth phenomena. In particular I will show that it is this symbiosis which
opens up new perspectives to gain further understanding about interfacial
growth problems, if one extends the paradigmatic, purely diusion limited
1

The meaning of better as used in the above context will become elucidated further
in the numerical appendix of this book.

1. Introduction

dendritic growth problem step by step to additional physical mechanisms


such as hydrodynamics or elasticity. One might wonder if in the end this
approach can provide a framework to tackle the behavior of still more complicated systems, e.g. systems with an inherent multi-scale nature due to an
internal structure, such as liquid crystals or polymer solutions, as well.

1.1 Structure and Scope of This Work


To elucidate the above concepts arising in the context of diuse interface
modeling I will rst within Chaps. 3 and 4 present the thermodynamic
background underlying that kind of modeling. It allows me to derive the
respective model equations on the basis of variational principles.
I will then proceed to introduce the mathematical formulation which
allows me to establish their correspondence to the sharp interface models,
i.e. the method of asymptotic analysis.
Regardless of whether one intends to employ a diuse interface model
(a) merely as a computational trick or (b) to derive new information of the
physics in the interfacial region, it is an important tool to deal with diuse
interface models: In case (a) the analysis might impose restrictions onto some
of the numerical parameters of the model. Moreover it serves as validation of
the model. In case (b) the method of asymptotic analysis is an instrument to
obtain expressions, which directly yield the velocity and the equilibrium conditions of the interface. In this case ensuring thermodynamic consistency of
the models serves as their validation. Usually this has to be guaranteed in addition to the variational principles and the asymptotic formalisms involved in
the constitution of the models. Only recently a generalized asymptotic analysis which inherently implies thermodynamic consistency has been introduced
by Elder et al. [102]. The way it was introduced it works for a restricted class
of interfacial growth problems, only. In this way it is described in detail in
Sect. 5.3.
In Chap. 6 and Chap. 7 the asymptotic analysis is taken up and employed
to address very recent issues involved with interfacial growth in the case the
latter is inuenced not only by diuse transport in the bulk phases but by
hydrodynamic transport and elastic driving forces as well.
Finally, in Chap. 8 this text concludes with a summary of the trends and
the perspective which open up in this young scientic eld of diuse interface
modeling in materials science.
To close the circle of discussion an appendix to the overall text provides
an overview on the numerical, implementation related issues involved.
The general scope of this book is to provide material, the study of which
will enable the reader to apply the diuse interface approach to the interfacial
growth phenomena he is interested in. Along this line the second chapter can
be understood as an introduction to the notion of interfacial growth and the
success achieved in the past modeling it based on a sharp interface approach.

1.1 Structure and Scope of This Work

Such a sharp interface formulation is the frame of reference for the discussion
of the diuse interface approach throughout this text. Thus the purpose of
Chap. 2 is to demonstrate, that even though phenomenological in nature
for the sharp interface approach it is possible to establish a thorough relation
between theory and experiment. In that sense it is regarded a solid basis it
is justied to build upon.
The following Chaps. 35 develop the thermodynamic background as well
as the mathematical tools required if one intends to engage into diuse interface modeling oneself. To keep them stringent, they remain deliberately
formal. However, they start from a level that any student of physics, mathematics or an engineering science could follow. For that purpose Appendix A of
Chap. 4 puts together some basics of the calculus of variations as required to
follow the derivation of model equations in Chap. 4 and succeeding chapters.
Moreover, Appendix B and Appendix C relate the CahnHilliard equation
respective the AllenCahn equation, as fundamental model equations summarizing the concept of diuse interface modeling, to some concrete physical
phenomena they allow us to describe. Within Chap. 5 the formal theory of
matching is placed into the context of physics by elucidating its application
for the example of thin lm epitaxial growth (Sect. 5.2).
Chapters 6 and 7 are the ones, in which all the formalism discussed before
is applied to two examples, namely
1. hydrodynamically inuenced dendritic growth and
2. elastically inuenced epitaxial growth.
These two applications serve as examples for phenomena, where the extension of sharp interface models to grasp all features of the growth process is
no longer clear. In that sense they are examples, for which the achievement
of diuse interface modeling is obvious at rst sight: It allows to model and
investigate these phenomena, for which a sharp interface approach would
fail. Moreover, in these cases it is the analysis of the diuse interface models,
which reveals new physical insight, which is not clear how to obtain otherwise. Chapters 6 and 7 are also meant to serve as guidance for the reader to
construct a diuse interface model for the application he might have in mind.
Within the concluding Chap. 8 the discussion of achievements of diuse
interface modeling as examplied through Chaps. 6 and 7 is continued. In
particular these achievements are put together with perspectives, thus bridging the gap from: What has been learned? to What can be learned?
An appendix to the book summarizes numerical issues, which apply to the
implementation of diuse interface models. In particular, some of the latest
high performance implementations are described. For the reader interested
in his own application this provides information for the step from model to
simulation. Moreover it gives him a feeling what simulations are possible from
point of view of computational capacity.

2. What Is an Interface?
Interfaces in Materials Science and Beyond

If we think of an interface in materials science, then obviously this term


applies to a boundary between two dierent phases of solid matter. In this
sense we might distinguish:

liquid - liquid,
liquid - solid,
solid - solid,
solid - vapor and
liquid - vapor interfaces.

It is relatively easy to agree on a common view for uid interfaces, for which
any physical quantity has a well dened smooth average prole. The situation
is somewhat more complicated for solid interfaces, for which the bulk solid
is modulated with the lattice periodicity. As a consequence one may view
interfaces between two phases, one of them solid, in two ways:
1. First, viewed on an atomic scale, the position of every atom is specied.
This leads to a description in terms of a terracestepkink model.
2. Second, viewed on a coarse grained scale, every property is averaged
over a nite volume in such a way as to have constant averages in the
bulk. These coarse grained quantities are then subject of a continuum
description in much the same way as uid interfaces.
Throughout this book I will choose the second approach. Then the notion of
surface excess becomes essential to quantify extensive variables in the interface region. This term dates back to the work of Gibbs. For its explanation
I start with a planar interface between two phases 1 and 2 located near the
plane x = 0 (see Fig. 2.1). An arbitrary extensive quantity has a prole (x),
where is the density per unit volume. In Gibbs description, one deliberately
ignores details of that prole. Rather one denes an integrated surface excess
s through the following construction: First choose an arbitrary dividing surface x = somewhere in the interfacial region as indicated in Fig. 2.1a. Next
extrapolate the bulk values 1 and 2 up to . Then the surface excess s is
dened by

s =

x2

x1

(x)dx 2 (x2 ) 1 ( x1 ) ,

where x1 and x2 are well inside each phase (Fig. 2.1a).


Heike Emmerich (Ed.): LNP m73, pp. 717, 2003.
c Springer-Verlag Berlin Heidelberg 2003


(2.1)

2. What Is an Interface?
(a)
x2

x2

(b)

x
phase 2

phase 2

phase 1

phase 1

x1

x1

Fig. 2.1. Denition of surface excess s (a) for a surface of zero volume, i.e. a sharp
interface, as well as (b) for a nite surface, i.e. a diuse interface.

It is obvious that the value of s is dependent on the precise choice of . If


changes by , then s changes by
s = (2 1 ) .

(2.2)

Only for quantities such that 1 = 2 the surface excess is dened unambiguously. In general any result referring to the surface must be invariant upon a
change of .
Equation (2.1) employs a single dividing surface, i.e. a surface of zero
volume. It is often more convenient to use two distinct dividing surfaces 1
and 2 . Then a denition of surface excess analogous to the one above results
in the following formulation:
 x2
(x)dx 2 (x2 2 ) 1 (1 x1 ) .
(2.3)
s =
x1

The underlying construction is depicted in Fig. 2.1b. For this case a change
of 1 and 2 alters the surface excess s in the following manner:
s = 2 2 1 1 .

(2.4)

In denition (2.4) the interface has a nite thickness ls = 2 1 , i.e. it is


diuse. In this sense it is possible to quantify extensive variables in an interfacial region based on Gibbs denition of surface excess for both, sharp as
well as diuse interfaces. Moreover, both interface types allow us to establish
thermodynamic relations between those quantities dealing with mechanical or
chemical equilibrium. The latter governs transfer of matter from one phase to
the other and thus is essential for interfacial growth. Classically formulating

2. What Is an Interface?

chemical equilibrium results in the well known GibbsDuhem relation or the


GibbsThomson relation for liquidliquid respective liquidsolid interfaces.
For a minimal model describing the growth of a phase-separating interface
only two additional physical requirements have to be met:
1. First, at the interface itself we would also require conservation of mass
and energy to be satised.
2. Second, within the bulk at least the transport equation for the physical
eld governing the growth process would have to be taken into account.
Indeed for dendritic growth, which can be regarded as a paradigmatic problem of interfacial growth ever since Kepler [178], the GibbsThomson relation
together with an interfacial boundary condition for the conservation of mass
and energy (Stefan condition) as well as an diusion equation for heat transport in the bulk is sucient to describe the growth process. Thus the minimal
model for dendritic growth reads:
Bulk

T 
= D 2 T  .
t

(2.5)




T  = TM
1 ,
L

(2.6)

Lv n = [(kT  )s (kT  )l ] n .

(2.7)

Interface
GibbsThomson relation

Stefan condition

Here equal diusion constants D for the transport of heat in the liquid and in
the solid phase have been assumed. Moreover, k = Dcp where cp the specic
heat per unit volume, which is assumed to be equal in both phases as well.

T  is the temperature eld, which governs the growth process. TM
denotes
the crystallization temperature of the planar interface, L the latent heat per
unit volume of solid, the anisotropic liquidsolid surface tension, and the
local curvature of the interface1 . One assumes a fourfold symmetry2 of the
crystal, so that the surface tension can be written as
= o a() = o (1 cos 4) ,
1
2

(2.8)

The prime refers to the fact that at this point the temperature eld is given in
dimensional units.
This multiple symmetry is a reminiscence of the underlying crystalline structure:
Even though Im only concerned with continuum descriptions of interfaces, the
crystalline grid results in an orientation dependence of surface tension along the
continuous interface also on the coarse grained scale. The precise dependence on
orientation can be obtained via a Wul construction [300]. According to such a
construction an underlying f cc-grid gives rise to a fourfold symmetry.

10

2. What Is an Interface?

where is the anisotropy factor and the angle between the normal to the
interface and the direction of propagation (cos = n ex ). The left-hand side
of formula (2.7) constitutes the latent heat release per unit volume of solid.
Its right-hand side accounts for the total energy ux away from the interface
to both sides of the interface due to heat conduction.
Equations (2.5)(2.7) constitute a sharp interface model for dendritic
growth. Even though simple in nature it is in good agreement with experimental data for single dendrites grown into transparent materials3 [128]. In
this sense it can be understood as an accepted, a validated model.
Extending the model only slightly allows us to study directional solidication within a sharp interface formulation. This is particularly convenient
for a validation through comparison to experiments, since directional solidication experiments can be carried out in a thin-lm geometry for samples as
thin as 12 m. Using for example CBr4 + 8 % C2 Cl6 with diusion length
50 m [5] the relevant dynamics of the experiment takes place in the 2D
plane. An experimental set-up underlying this type of directional solidication experiments is depicted in Fig. 2.2. The set-up specied here was used
by Akamatsu and Faivre to study the doublondendrite transition in directional solidication [5]. Pictures of the doublon morphology as obtained in
their experiments are displayed in Fig. 2.3. Simulations by Ihle revealed their
relation to a minimal sharp interface model of directional solidication, which
in terms of a dimensionless diusion eld u explained further in (2.15)(2.17)
reads:
Bulk

u = D2 u,
t

(2.9)

Interface
GibbsThomson relation
Vt
kin ()vn ,
lT

(2.10)

((1 K)uInt + K)v n = DuInt n .

(2.11)

u|Int = 1 d()
Stefan condition

Here = (y) := 0 denotes the x-position of the interface and d() = d0 (1


cos 4), where d0 is the capillary length which is related to the surface
tension of (2.8) via

cp TM
d0
.
(2.12)
2
L
Thus it inherits the fourfold symmetry of :
3

Systematic deviations from the theory at small undercoolings will be discussed


in detail in Chap. 6.

2. What Is an Interface?


d()

() +

2 ()
2

11


cp TM
.
2
L

(2.13)

The fourth term on the right hand side of the GibbsThomson relation,
i.e. kin ()vn , denotes deviations from local thermodynamic equilibrium. It
becomes relevant, if the interface is growing rapidly and the idea of an instantaneous relaxation of the physical elds to the position of the interface does
not hold any longer. Just as capillary eects kinetic eects can be anisotropic:
kin () = kin0 (1kin4 cos 4), where the kinetic anisotropy factor kin4 does
not necessarily have to be equal to the crystalline anisotropy factor . Moreover,
ml C
,
(2.14)
lT :=
G
where ml is the so-called liquidus slope, C the miscibility gap and G the
temperature gradient. K denotes a so-called distribution coecient.
(a)

(b)

video-tape
recorder
microscope

z
hot oven

70mm

cold oven

8mm

x
glass
plates

growth y
front

x
hot edge

11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
liquid
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
solid
00000000000
11111111111
00000000000
11111111111
000000
111111
00000000000
11111111111
000000
111111
000000
111111
00000000000
11111111111
000000
111111
000000
111111
00000000000
11111111111
000000
111111
000000
111111
00000000000
11111111111
000000
111111
000000
111111
00000000000
11111111111
000000
111111
000000
111111
11111111111
00000000000
000000
111111
000000
111111
11111111111
00000000000

y
growth front
solid-solid
transition
spacers

cold edge

Fig. 2.2. This experimental set-up was used by Akamatsu and Faivre for their
experiments on the directional solidication of CBr4 + 8 % C2 Cl6 thin-lms [5].
(a) Principle experimental set-up. (b) More specic sketch of the sample itself. x is
the direction, in which the sample is pulled with constant velocity V .

The picture underlying that model is that of a growth process driven by a


solute eld rather than a temperature eld. This implies that the experiments
are no longer done with pure materials but rather with alloys. Moreover,
within a directional solidication experiment the sample containing the alloy
is moved with constant velocity V within an externally opposed temperature
gradient such that the position of the interface is kept xed. This geometry is
depicted in Fig. 2.2 on the right. If the heat conductivity of liquid and solid
phase are approximately equal, then the temperature prole between the hot
edge and the cold edge (see Fig. 2.2) can well be assumed as:

12

2. What Is an Interface?

T = T0 + Gx .

(2.15)

By an appropriate choice of the temperature T0 it is possible to keep the


position of the interface at x = 0. The diusive transport equation, which
remains to be solved, is that of the solute eld C. Formulating solute diusion
one has to take into account, that the value of the solute eld depends on
temperature. Figure 2.4 depicts the dependence of C on T . Here liquidus
line, the slope of which appears as ml within (2.14), solidus line with slope
ms and the miscibility gap C are displayed. The distribution coecient K
appearing within (2.11) can now be understood as the ratio
K=

C
,
C + C

(2.16)

where C denotes the concentration of the solute eld at T = T0 (see


Fig. 2.4). Moreover, the governing eld u, in terms of which model equations (2.9)(2.11) are written, reads
u(x, y, t) =

C(x, y, t) C
.
C

(2.17)

Apart from the experiments by Akamatsu and Faivre quantitative validation of sharp interface modeling for directional solidication was obtained
e.g. through experiments by Bechhoefer et al. [25, 258] as well as Trivedi et
al. [275]. These experiments were concerned with the so called limit of absolute stability [223]. The limit of absolute stability is a maximal velocity of a
directional solidifying front in a parameter regime, where it is stable whatever the value of the thermal gradient is. Theoretically it was predicted to
be d0DK [223]. Indeed this was observed in the experiments by Bechhoefer et
al. on the liquid-to-nematic transition of 8CB [25, 258] as well as by Trivedi
et al. on distillated CBr4 [275]. Other evidence for a validation of sharp interface models from directional solidication experiments centered around the
issues of shape and velocity selection, e.g. [27, 90, 112, 249, 260, 274].
To summarize the experimental evidence at this point, one may state that
there is a long tradition in sharp interface modeling of solidifying phase
boundaries and relating these models to experiments. The same is true
for solidvapor or solidliquid interfaces subject of epitaxial growth (see,
e.g., [288] and references within). Thus it seems justied to construct such
kind of models simply from formulating the governing equations for transport in the bulk phases coupled to mass and energy conservation (Stefan
condition) as well as the requirement of local thermodynamic equilibrium
(GibbsThomson condition) at the interface.
From a mathematical point of view these problems constitute a class of socalled moving boundary problems. Usually analytical treatment of these type
of problems is very restricted. Thus it seems natural to search for adequate
numerical treatment. However, such numerical treatment proofs dicult, as
well.

2. What Is an Interface?

13

Fig. 2.3. The doublon morphology as obtained in thin-lm directional solidication


with CBr4 + 8 % C2 Cl6 by Akamatsu and Faivre. Picture taken from [5].

14

2. What Is an Interface?

x=(TT0 )/G

Liquid

Tm

Liquidus
Line

T0

x=0

Solid
C

Solidus
Line

Fig. 2.4. Part of the phase diagram of a binary alloy. Liquidus and solidus curve
are depicted just as well as the miscibility gap C and the alloy melting temperature Tm . Via an appropriate choice of the temperature T0 it is possible to keep the
position of the interface at x = 0 (see (2.15)). C = C S (T0 ) refers to the concentration of solute in the solid phase at T = T0 . Due to (2.15) the temperature scale
given on the left vertical axis of the diagram corresponds directly to the in growth
direction (i.e. the x-coordinate) of the sample as indicated by the second vertical
axis of the diagram to the right.

Ideally one would like to simulate the eect of the boundary by treating
it explicitly, i.e. with no smearing of information at the interface resulting
in numerical diusion. For xed boundaries, even if very complex in shape,
techniques for such an explicit treatment include blockstructured domain
decomposition [256], overset meshes [156, 264] or unstructured boundary
conforming curvilinear grids [286] to discretize the domain. For moving
boundaries, on the other hand, xed grid techniques are employed, in which
the computations are performed on the xed grid while at the same time the
interface is tracked explicitly as independent curve. These approaches constitute mixed EulerianLagrangian methods. Examples of such approaches
are the immersed boundary technique [157, 233, 280, 285], cut-cell type approaches [230, 241, 257, 278, 279], the immersed interface method [206] and
the ctitious domain methods [132].
The complications arising from such a mixed EulerianLagrangian method
are twofold:
1. First, it implies two distinct schemes of discretization, namely that for
the moving interface opposed to the xed grid employed for the transport equations. However, using the GibbsThomson relation properly as

2. What Is an Interface?

15

boundary condition for the heat transport eld requires to interpolate between these two discretization schemes and thus brings about unavoidable
interpolation errors. Moreover, the Stefan condition requires to determine
normal gradients away from the interface onto the xed grid, which again
involves interpolations. Thus ensuring numerical stability and adequate
accuracy gets expensive in terms of CPU time.
2. Second, for EulerianLagrangian schemes the number of discretization
points of the scheme varies from time step to time step. As a result
these approaches are dicult to parallelize eciently. However, having
in mind applications to real phenomena posing open questions such as 3D
systems including more than just diusive transport dynamics, ecient
implementation becomes essential to carry out the parameter studies
necessary to gain new physical insight.
Thus if it were only for numerics, a diuse interface approach in materials
science would seem to be a natural advance to treat interfacial dynamics numerically, since it overcame item 1 as well as item 2 of the enumeration above.
In this sense one would employ diuse interface models along the lines of the
rst philosophy I pointed out in the introductory chapter, i.e. merely as a
computational trick. Other philosophies motivating diuse interface theory
also from the point of view of new perspectives with respect to the modeling
itself have been discussed within the introductory chapter as well. Their impact will become clearer in Chaps. 6 and 7 of this book, being concerned with
phenomena, for which particularly the diuse interface approach provides a
way to gain new insight. Moreover I will summarize within the nal chapter
of the book in which way new perspectives for modeling as well as numerics
open up new chances for gaining insight in evolving phase boundaries.
In this context phase boundaries are more than the interfaces known from
materials science. Talking of interfaces in a broader sense one could include
evolving boundaries from e.g. combustion, image processing, computer vision,
control theory, seismology and computer aided design, as well. For some of
these problems the moving interface seems virtual and emerges only after
some sort of transformation. For example for the problem of shape-fromshading in computer vision, one has to map the image segmentation problem
onto a moving interface version of active contours. However, for all of these
examples the key ingredient to construct a diuse interface model is to rethink
the Lagrangian geometric perspective and replace it with an Eulerian, partial
dierential equation. Having accomplished that task, the gain is advanced
numerical treatment as well as new perspectives for modeling just as in the
case of materials science.
In many aspects the diuse interface approach resembles features of the
level set method. Nevertheless there are clear dierences, as well. To conclude
this chapter I will therefore contrast the two methods for applications in
materials science:

16

2. What Is an Interface?

The level set method was rst introduced by Osher and Sethian [228].
Since then level set algorithms have successfully been applied to a wide variety
of problems [66, 70, 114, 148, 215, 266, 302]. The fundamental idea behind
the level set method is the representation of the interface ij by the help of
a zero level set function (x, t), i.e.:
ij = {x : (x, t) = 0} .

(2.18)

Given a velocity eld v, one can analyze the dynamics of ij by relating


it to the motion of the zero level set of . The partial dierential equation
describing the temporal evolution of such that the level sets move with v
is

+ v = 0 .
(2.19)
t

Since the normal vector n can be written in terms of as n = ||


and since
further v = vn n, (2.19) is equivalent to the level set equation:

+ vn || = 0 .
t

(2.20)

The dicult issue when employing the level set approach is the derivation of
the velocity function v and its extension to the complete domain of the level
set function as required to solve (2.20). The extension of v to the complete
domain can be achieved via the so-called Fast Marching Method [253], which
results in a rst-order accurate solution only, if linear triangular elements are
used. Higher order elements enable higher order solutions, however at a much
higher computational price.
Moreover the boundary conditions have to be evaluated at the boundary
ij , now given by ij = {x : (x, t) = 0}. ij does not necessarily coincide
with the grid points of the underlying Eulerian discretization scheme. As a
consequence the boundary conditions cannot be resolved on exactly this grid.
In [71] this problem is treated via interpolations, which lead to a smearing
of information over the interface and limit the capability of this method to
take into account precise kinetics of an interface. In [72] an alternative way
is chosen, i.e. the problem is met by employing a nite element mesh with
linear triangular elements, whose nodes are placed exactly on the boundary. Obviously there is a price to pay for this as well, namely tedious mesh
adaption.
The basic dierence between the level set method and the phase-eld
method is that the latter method depends on a small parameter for the
interface thickness. Thus the interface is not a sharp interface as in the level
set case, and boundary conditions have to be fullled only asymptotically for
vanishing interface thickness. Obviously this restricts the operation of this
method to interface thicknesses proportional to the grid size. As a result it
used to be hard to employ the phase-eld method in a quantitative manner,
since it proved to be computationally demanding to choose the interface

2. What Is an Interface?

17

thickness small enough to resolve the desired sharp interface limit. In the
context of dendritic growth eorts to overcome this diculty resulted in
the so-called thin interface analysis [168]. This thin interface analysis also
improved the capability of the phase-eld method to model kinetic eects in
detail considerably. I will discuss it in more detail in Chap. 5.
For further information on the level set method the interested reader is
referred to [254], which is a comprehensive and self-contained introduction to
that eld.

3. Equilibrium Thermodynamics of Multiphase


Systems: Thermodynamic Potentials
and Phase Diagrams

A rst step when entering the eld of thermodynamics certainly is concerned


with equilibrium thermodynamics, i.e. with gaining an understanding about
how a system can be characterized by macroscopic state variables and potentials and how those can be organized into phase diagrams1 for systems at
equilibrium. Thermodynamic equilibrium is reached, if the state of a system
remains constant, unless variation of subsidiary conditions triggers a change
of governing variables. Thus a denition of thermodynamic equilibrium always comes along with a set of well dened rules (the subsidiary conditions),
which tell us, which (depending) variables can change freely2 and which variables are the state governing ones, which will have to be kept constant. A
thermodynamic state is described fully by a minimal set of governing variables, which are independent of each other. The number of those independent
governing variables is system dependent.
To describe a thermodynamic equilibrium situation mathematically, the
notion of a thermodynamic potential is essential, since it allows us to dene
the equilibrium state as the one, for which the governing thermodynamic
potential becomes extremal. The basic thermodynamic functions to consider
within this context are the Helmholtz free energy A and the Gibbs free energy G. One considers them as thermodynamic potentials, if they appear as
functions of their natural variables. For the Gibbs free energy G the natural
variables are the pressure P , the temperature T , and the molar number n of
a system. For the Helmholtz free energy A, on the other hand, the systems
volume V replaces the pressure P in the set of natural variables. With these
respective dependencies G and A allow for a complete description of a ther1

A phase diagram depicts the two phases solid and liquid in dependence on the
three variables temperature, concentration and pressure - thus a full phase diagram should be three dimensional. However, since approximately constant pressure is assumed, here the diagram turns out to be a 2D plot just as in Fig. 3.1.
Within the diagram equilibrium phases are denoted by S (solid) and L (liquid). Moreover, in the particular diagram given by Fig. 3.1, there is exactly one
third domain in the diagram referring to coexistence of liquid and solid phase
(S+L). The curves separating the coexistence domain from the solid and liquid
domain, respectively, are the solidus and liquidus curve introduced in Fig. 2.4.
and in fact will change during the evolution of the system towards thermodynamic equilibrium

Heike Emmerich (Ed.): LNP m73, pp. 1930, 2003.


c Springer-Verlag Berlin Heidelberg 2003


20

3. Equilibrium Thermodynamics of Multiphase Systems

T1

_m

TA

T2

S+L

_m
TB

Fig. 3.1. The typical phase diagram of a binary mixture with Gibbs free energies
of both phases as given in Fig. 3.2 for T 1 and Fig. 3.3 for T 2 .

Fig. 3.2. Typical curves depicting the molar Gibbs free energy for the solid phase
S
L
G and the liquid phase G of a binary mixture at a specic temperature T 1 .
Logarithmic contributions due to the mixture entropy are neglected. On the C-axis
the relative portion of species B is plotted. Here the liquid phase is the stable one
for all mixture ratios (compare to Fig. 3.3).

3.1 Calculating Phase Diagrams from Energy Functionals

21

modynamic system in the sense that the physical variables for all well dened
thermodynamic states can be determined from them via extremal principles.
In the following I will rst concentrate on the Gibbs free energy G and its
meaning for simple two-phase systems. This concept of G is an essential
background to describe the dynamics of phase separating interfaces later. I
will then specify some relevant Gibbs energies with respect to special classes
of model systems and also explain special phase diagram approaches to those
systems. This will allow me to elucidate the discrepancy of a locally nonconserving-variable-approach and a locally conserving-variable-approach
with respect to the concept of thermodynamic consistency as one of the key
ideas for the remainder of this work.
The following concepts can be found in more detail within elementary
thermodynamic textbooks as e.g. [4, 184]. The short-cut introduction presented here follows mainly [2].

3.1 Calculating Phase Diagrams


from Energy Functionals
As mentioned in the previous section, the equilibrium conguration of a thermodynamic system is characterized by an extremum of the relevant thermodynamic potential. For binary mixtures, for which two phases S (solid) and
L (liquid) can exist, the dierential of the Gibbs free energy for each of the
two phases is given by:
S
L
dG = dG + dG ,
(3.1)
where

dG = V dP S dT +

k, dnk,

{L, S} . (3.2)

k{A,B}

In the above equations P denotes the pressure of the system and T its temperature. These are the state governing variables. Thus they are not allowed
to depend on the phase . V and S are the volume and the entropy of
phase , respectively. nk, is the molar number of component k {A, B} in
phase . The chemical potentials are dened by

k, =

G
G
=
,
nk,
C k,

(3.3)

with the molar Gibbs free energy G and a relative portion of C k, of component k in phase . At this point surface eects at the phase-separating
interface, which will be an important ingredient of succeeding chapters dealing with the dynamics of the systems, are neglected. Assuming approximately
constant pressure, the P -dependence in (3.2) can be omitted, as well.

22

3. Equilibrium Thermodynamics of Multiphase Systems

_
C

_
C

Fig. 3.3. Same as in Fig. 3.2, however at a dierent temperature T 2 , which is


smaller than the temperature T 1 chosen for Fig. 3.2. As a result the energy curve
displays an equilibrium conguration at which both phases coexist.
L/S

are displayed as a function of the relative


In Fig. 3.2 typical forms of G
portion of species B in the respective phases for a xed temperature T 1 . As
proven in [142] the molar Gibbs free energy always takes a local maximum
for mixings 0 and 1. Between these two points the graph is a convex curve.
S
L
In Fig. 3.2 G is larger than G for all mixture ratios. Thus the liquid phase
is the stable phase for arbitrary concentrations. Reducing the temperature,
energetic congurations will be changed in favor of the solid phase. At a
m
certain temperature T A < T 1 , the melting temperature of A, both curves
will coincide at a concentration C = 0. Reducing the temperature further
to T 2 results in an intersection of curves and thus a coexistence of phases
(Fig. 3.3). The equilibrium conguration can be obtained from the following
consideration [294]: The identities

A, = G (C B, ) C B,
B,

G (C B, )
C B,

G (C B, )
= G (C B, ) + (1 C B, )
C B,

(3.4)
(3.5)

reveal that the chemical potential for a concentration C B, is determined


by the value of the curve tangential to G in C B, . On the other hand,
because of conservation of species as implied by (3.1) and (3.2), one obtains
the equilibrium condition
k,L = k,S ,
(3.6)

3.2 Abstracted Phase Diagrams


L

23
S

which implies that at equilibrium the two curves tangential to G and G ,


respectively, result in identical values for C B = 0 and C B = 1. This condition
L
S
is fullled by a doubletangent curve, which joint points with G and G
S
L
yield the equilibrium concentrations in liquid (C ) as well as solid phase (C )
(see Fig. 3.3).
Further reduction of the temperature results in changes of this value towards higher concentrations. Finally, if the melting temperature of component B is reached, the solid phase is stable for all mixture ratios. If one
plots the equilibrium concentration values versus temperature, one obtains
the phase diagram of two completely miscible species (Fig. 3.1). There is
a two-phase state S+L, for which the system minimizes the value of G
through the coexistence of solid and liquid phase. The relative amount of the
I
S
L
I
two phases at equilibrium is given by the ratio (C C )/(C C ), where
I
C is the overall fraction of species B in the mixture. A phase diagram of
this kind is for example found for CuNi-mixtures [198].
In general the exact form of a phase diagram depends on the two functions
S
L
G and G . If, for example, both of them take a minimum at nearly identical
L
mixture ratios, then the curve for G displays a stronger curvature than the
S
one for G . As a result, a further increase in temperature3 results in an
intersection of the two curves at a concentration close to the two minima. A
phase diagram dierent from the one depicted in Fig. 3.1 in the sense that
two distinct (but not disjunct) coexistence regimes arise, is the result. Such
a kind of phase diagram can for example be found for CsK-mixtures [198].
For a K concentration of 50.5 % solidus and liquidus slopes coincide. At this
point both curves are at a minimum (Fig. 3.4). The real phase diagram
displayed in Fig. 3.4 depicts the phase boundary of a CsK2 -phase (lower right
corner of the gure), as well.
This method to calculate the phase diagram from energy functionals
is not restricted to a system of two components, i.e. a binary mixture.
By introduction of suitable potentials even the complex phase diagrams of
multicomponentmultiphase systems can be calculated as well as validated
from experimental ndings [62].

3.2 Abstracted Phase Diagrams


If one focuses ones main interest on the physical processes underlying the
dynamics of two-phase systems and limits ones attention to certain mixture
ratios, then simplied phase diagrams can be developed. They can be understood as abstractions of basic types of phase diagrams as, for example, the
one depicted previously in Fig. 3.1.
3

The starting point for this consideration is a temperature, at which the system
is solid for all concentrations.

24

3. Equilibrium Thermodynamics of Multiphase Systems

Fig. 3.4. Real phase diagram of a CsK-mixture [198]. Compared to the phase
diagram depicted in Fig. 3.1 this displays a second, qualitatively dierent type
of diagram found for two-phase systems containing two distinct but not disjunct
coexistence regimes.

m
TA

1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
L
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
mL
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
mS
0000000000000000000000000000000000000000
_ 1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
T 1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000
1111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000
1111111111
S
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000
1111111111
S+L
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
1111111111
0000000000
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
1111111111
0000000000
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111

Fig. 3.5. Reduced phase diagram for a binary mixture of dominant component
A and only small portions of B (i.e. weakly contaminated melts). The coexistence
domain is hatched. It is bounded by the solidus curve with slope mS and the liquidus curve with slope mL . These two curves intersect at the melting temperature
m
TA.

3.2 Abstracted Phase Diagrams

25

There are many questions concerned with systems, which involve a dominant phase A mixed with only a small amount of a second species B. The
prototype is a weakly contaminated melt. In this case it is important to
know the precise conguration of the phase diagram only for a small interval
of concentration values. Then it is justied to expand solidus and liquidus
m
lines around the melting point T A , so that the two-phase domain is bounded
by solidus and liquidus curve with constant slopes mL and mS , respectively. Here mL and mS are both larger than 0 (Fig. 3.5). Reduced phase
diagrams of this kind have extensively been used for theoretical and analytical investigations of solidication phenomena in binary mixtures in the
past [28, 53, 65, 80, 81, 82, 125, 162, 163, 164, 199, 202].
If for a phase diagram as depicted by Fig. 3.1 the interesting concentration
of species B turns out to be in the range of C = 0.25 to C = 0.75, Fig. 3.5
is no longer a suitable abstraction. Rather liquidus and solidus curve can
be assumed to be approximately parallel, i.e. m = mL = mS . This type of
reduced phase diagram is depicted in Fig. 3.6. It is characterized by a miscibility gap, which is independent of the concentrations and implies invariance
of translation along the solidus and the liquidus curve. As a result, deriving
the energetic potentials to obtain dynamic equations subsequently is particularly simple. At rst sight, a severe short-coming of this type of diagram
seems to be the absence of a well dened melting temperature4 . However, this
does not have any consequences for the derivation of the evolution equations
of the system, since the melting temperature5 is only relevant to determine
the capillary length. In Sect. A.1 I will show that the capillary length is still
well dened for systems represented by Fig. 3.6 and that it can be obtained
in an alternative manner. Thus the absence of a precise melting point does
not interfere with the derivation of model equations, which are meant to display the relevant physical eects. In this sense the reduced phase diagram
Fig. 3.6 itself can be understood as comprising all information necessary to
derive the dynamic equations of the system. Because of its convenient behavior with respect to this point, I will present the mathematical formalism
of the succeeding chapters based on it as representative for the underlying
thermodynamics. Usually the formalism can be extended to cover systems
described by a reduced phase diagram as in Fig. 3.5, as well.
The phase diagram of Fig. 3.6 is characterized by four parameters:
MC , MT , C0 and T0 . They can be transformed into the two dimensionless
variables
C C0
T T0
C
;
T
.
(3.7)
MC
MT
T and C allow us to construct a parameter-free model phase diagram. Now
the miscibility gap as well as the temperature dierence between liquidus
and solidus line is 1, moreover mS = mL = 1. In the following section I will
4
5

Since there is no intersection of liquidus and solidus curve.


if it is available as in Fig. 3.5

26

3. Equilibrium Thermodynamics of Multiphase Systems

C0

1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000
1111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
000
111
0
T1
M
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
000
111
10
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
1111
0000
L
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
00C10
11
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
M
00
11
0
1
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
000
111
0000000000000000000000000000000000000000
_ 1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
T 1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
00000
10
10111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
S+L 1
00000
11111
0
00000000
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
1111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
S
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
0000000000000000000000000000000000000000
1111111111111111111111111111111111111111
T0

Fig. 3.6. Reduced phase diagram with constant miscibility gap MC . The temperature dierence MT of solidus and liquidus curve is constant. Therefore both curves
display the same slope m. The reference point (C 0 , T 0 ) can be chosen arbitrarily.

discuss the derivation of the Gibbs free energy in the liquid as well as the
solid phase for such a parameter-free diagram.
3.2.1 Constructing the Gibbs Free Energies
Close to thermodynamic equilibrium GL and GS can be expanded as:
GL (C, T ) = A1 (C A2 )2 + A3 (C A4 )2 + A5 CT
GS (C, T ) = B1 (C B2 )2 + B3 (T B4 )2 + B5 CT + B6 ,

(3.8)
(3.9)

with expansion parameters Ai , Bi . For an arbitrary temperature T the equilibrium concentration in liquid and solid is given by
C L (T ) = 1 T

C S (T ) = T .

(3.10)

The equilibrium conditions following from the doubletangent construction


at GL and GS (see (3.4)(3.6)) read


GL (C, T ) 
GS (C, T ) 
=
(3.11)


C
C
C=C L
C=C S

GS (C, T ) 
GL (C L , T )
= GS (C S , T ) +
(C L C S ) .(3.12)

C
S
C=C

3.2 Abstracted Phase Diagrams

27

(a)
8
T=0

4
G

(b)

T= 1
4
G

(c)

8
T= 2

4
G

Fig. 3.7. For three dierent temperatures the Gibbs free energy of liquid phase
(solid line) and solid phase (dashed line) are plotted: (a) T = 0, (b) T = 1 and (c)
T = 2. The energies follow from (3.21). The respective equilibrium concentrations
following from the doubletangent construction are indicated.

28

3. Equilibrium Thermodynamics of Multiphase Systems

They have to be fullled for arbitrary T and thus at any order of T , so that
the Ai and Bi obey the following relations:
A1 (1 A2 )
2

A3 A24

= B1 B2 ,
= B1 B2 (B2 2) + B3 B42 + B6 ,
= B5 2B1 ,

A1 (1 A2 ) +
A5 2A1
2A1 (A2 1) 2A3 A4 + A5 = 2B1 (B2 1) 2B3 B42 + B5 ,
A1 + A3 A5
= B1 + B3 B5 .

(3.13)
(3.14)
(3.15)
(3.16)
(3.17)

These can be transformed to


A5
+ B1
2
A5
A3 =
+ B3
2

A1 =

B6 =

B5
2
B5
2

2B1 B2
,
(3.18)
A5 + 2B1 B5
2B1 + 2B3 B4 + A5 B5
A4 = 1 +
, (3.19)
A5 + 2B3 B5
A2 = 1 +

2B12 B22
(2B1 + 2B3 B4 + A5 B5 )2
+
A5 + 2B1 B5
2(A5 + 2B3 B5 )
B2
) B3 B42 .
+2B1 B2 (1
2

(3.20)

A5 , B1 , B2 , B3 , B4 and B5 remain parameters one can choose freely. If


the energy term does not contain any CT coupling (i.e. A5 = B5 = 0),
one obtains B1 = B2 = B3 = B4 = 1 and thus particularly simple energy
functionals in solid and liquid. They are visualized for three dierent values
of T in Fig. 3.7. Formally they are given by:
GL (C, T ) = (C 2)2 + T 2
GS (C, T ) = (C 1)2 + (T + 1)2 + 1 .

(3.21)
(3.22)

3.2.2 Non-conserved Versus Conserved Variable Approach


In succeeding chapters of this text one of the key issues is how to derive
a thermodynamically consistent phase-eld model for the dynamics of the
phase-separating interface based on a governing thermodynamic potential.
A key element to accomplish this task is to substitute the dimensionless
temperature T in (3.21) and (3.22) through the dimensionless energy density
Q. The necessity to do so is twofold:
1. First, the formalism to derive an evolution equation from the governing
potential is dierent depending on whether a variable is a conserved
one or a non-conserved one. This will become clearer in Chap. 4, where
this formalism is explained. As a consequence it is quite essential to
distinguish strictly between locally conserving and locally non-conserving
variables of the system (see (4.42) and its explanations).

3.2 Abstracted Phase Diagrams

29

2. Second, if evolution equations are derived from functionals, which contain the temperature as variable rather than the energy density, then the
dynamic behavior is characterized by the feature, that changes to temperature cause changes in energy, as well. The thermodynamic potentials
are no longer monotonically decreasing in time as could be proven by
Penrose and Fife [231, 232]. A transition from temperature to inner energy as governing variable overcomes this problem. Thus it opens a way
to derive thermodynamically consistent phaseeld models.
If one replaces T by Q, one has to pay attention to the fact, that the
doubletangent construction given by (3.11) and (3.12) is no longer suitable.
A more general formula has to be obtained. To do so, one minimizes the
energy functional with respect to the remaining independent system variables
by at the same time ensuring the necessary conservation conditions for the
functional. To obtain an adapted formula for the doubletangent construction
one considers a closed binary system with species A and B and with xed
number of particles NA and NB (N NA + NB ). Moreover, the total energy
E = E S + E L is xed. E S and E L are contributions to the energy E from
liquid and solid phase, respectively. In addition,
NS
n
C S
C L
S
Q
L
Q

:
:
:
:
:
:

total number of particles in solid phase


number of particles of species B in solid phase
concentration of species B in solid phase: C S n/N
NB n
concentration of species B in liquid phase: C L N
NS
S E S /NS
energy per particle in solid phase: Q
S EE S .
energy per particle in liquid phase: Q
N NS

Here variables C S , C L , ... have been introduced, which are identical with
the variables C S , C L , ... except for a normalization factor.
At this point the conguration of the system is determined by the values
S , as well as N , NB and E, which are external parameters.
of n, N S and Q
Minimizing
S ) NS G
S ) + (N N S )G
L (C L , Q
L ) (3.23)
S (C S , Q

G(N,
NB , E, n, N S , Q
S results in
with respect to n, N S and Q
GS (C S , QS )
GL (C L , QL )
=
C S
C L
GL (C L , QL )
GS (C S , QS )
=
QS
QL

(3.24)
(3.25)

L
L
L
GS (C S , QS )
L G (C , Q )

C
C S
C L
S
S
L
L
S
L
G (C , Q )
G (C , Q )
+ QS
QL
.
(3.26)
QS
QL

GS (C S , QS ) GL (C L , QL ) = C S

30

3. Equilibrium Thermodynamics of Multiphase Systems

These conditions represent a set of possible surfaces tangential to GL and


G , which allow us after xing one variable, e.g. QS to determine the
equilibrium conditions of the remaining variables QL , C L and C S .
S

Once the relation between T and Q in both phases is xed, the three
equations (3.24)(3.26) can be used to calculate the functions GL (C L , QL )
and GS (C S , QS ). For this it is convenient to transform the latent heat6 L
and the energy density Q to dimensionless variables as:
Q cp T 0
cp MT

L
cp MT

(3.27)

(here cp denotes the specic heat of the material). Afterwards the relation
between dimensionless energy density and dimensionless temperature reads7 :
 S
: in the solid phase
Q
T =
.
(3.28)
QL 1 : in the liquid phase
If the system is at thermodynamic equilibrium the relevant phase diagram
has to fulll
CS = CL 1

QS = QL 1

C S = QS . (3.29)

If one determines8 the parameters Ai and Bi such that (3.24)(3.26) and


(3.29) are satised, the thermodynamic functional reads
GL (C, Q) = (C 2)2 + Q2
GS (C, Q) = (C 1)2 + (Q + 1)2 .

(3.30)

These forms of the functional will be the basis for the derivation of model
equations governing the dynamics of a phase-separating interface in the following. The succeeding chapter will make use of it to develop the formalism
by which the potential is linked to the dynamics of the system in detail.
Essentially, the above concepts have to be extended to the theory of irreversible thermodynamic, which opens a way to derive the relevant transport
equations.

7
8

In the context of solidication latent heat refers to thermal energy which is released at the interface due to the phase change. Thus a transformation of the
inner energy of the system naturally involves a transformation of the latent heat,
as well.
Note that within this context the dimensionless latent heat is normalized to 1.
by proceeding just as in (3.8) and (3.9)

4. Thermodynamic Concepts
of Phase-Field Modeling

Thermodynamics provides functionals like the Gibbs free energy discussed in


detail in the previous chapter. They behave like Lyapunov functions changing
monotonically in time. In that sense, thermodynamics provides the basis for a
variational approach. If an extremal principle is applied to the rate at which
some total (free) energy decreases, then this yields the dynamic equations
governing the evolution of the physical system under consideration. The basic
theory of how to derive transport equations for irreversible thermodynamic
transport processes dates back to Onsager in 1931 [227]. Subsequently it was
rened and extended by dierent other authors [91, 140, 237]. The theory
is based on the assumption that the system under consideration is close to
an equilibrium condition, so that locally thermodynamic equilibrium can be
assumed. As a consequence the concepts of equilibrium thermodynamics can
be extended to these non-equilibrium conditions based on linearizations.
However, the Onsager theory is certainly only one step in the direction of
a diuse interface model for a moving, phase-separating surface. It gives an
idea of how transport equations can be obtained from the thermodynamic
functions of a system. On the other hand it does not contain any notion of
a phase-eld variable as introduced in Chap. 1. Moreover, kinetic principles
as an essential component of the physics of a moving interface are not part
of the theory.
Kinetic eects are more likely to be found in phenomenological microstructures evolution theories, which start with constitutive relations and
empirical data providing equations and principles for the dynamics. These
equations and principles, on the other hand, do not necessarily ensure thermodynamic consistency. Guaranteeing compatibility with the laws of thermodynamics yields additional restrictions, for instance on the elastic coecients
and chemical rate constants. In this sense, thermodynamic laws alone are not
sucient to derive the dynamics of microstructural evolution. Nevertheless
any comprehensive approach to model interfacial growth problems has to be
compatible with them.
This raises the question: What aspects does a diuse interface approach
to moving boundary problems have to fulll to qualify as comprehensive
approach for such problems? And: How does a thermodynamic foundation
help with respect to that point?

Heike Emmerich (Ed.): LNP m73, pp. 3158, 2003.


c Springer-Verlag Berlin Heidelberg 2003


32

4. Thermodynamic Concepts of Phase-Field Modeling

To address these questions the rst section of this chapter will be devoted
to the derivation of transport equations from an underlying thermodynamic
potential based on Onsagers theory. In the second section this approach is
extended to account for a phase-eld variable. The third section nally deals
with the question of thermodynamic consistency.

4.1 Derivation of Transport Equations


4.1.1 Considering Conserved Quantities Only
The starting point of the concepts developed in this section is the knowledge
of a thermodynamic potential of the system under consideration, e.g. the
Gibbs free energy or the entropy. I will refer to it as

P (X1 , ..., Xn )dV .
(4.1)
P(X1 , ..., Xn ) =
V

Here P is the respective density function. X1 , ..., Xn constitute the set of


relevant extensive variables normalized to a unit volume. Within this section
I assume Xi to be locally conserved. Because of the identity
dP (X1 , ..., Xn ) =

n
n


P
dXi
Fi dXi
Xi
i=1
i=1

(4.2)

the total density ow JP of P results from the density ows Ji of Xi as


JP =

n


Fi J i .

(4.3)

i=1

Here the conservation equations which have to be fullled by the extensive


variables can be written as
Xi
+ Ji = 0
t
Therefore the rate of the total production
equation
dP
P
dt

(i = 1, ..., n) .
dP
dt

per volume element obeys the

P
+ JP
t
n
Xi
(4.2),(4.3) 
=
(Fi
+ (Fi )Ji + Fi Ji )
t
i=1
=

(4.4)

n

i=1

(Fi )Ji .

(4.4)

(4.5)
(4.6)
(4.7)

4.1 Derivation of Transport Equations

33

((Fi ) are called anities of the potential.) Now assuming that the system
under consideration is a Markovian system and that the ows Jk depend only
on the actual values of the Fi as well as on the anities, one can write
Jk = Jk (F1 , ..., Fn , F1 , ..., Fn ) .

(4.8)

Since we are close to thermodynamic equilibrium it is justied to expand Jk


around its equilibrium value Jk = 0 and to neglect second and higher order
terms. The assumption that vanishing anities also imply vanishing Jk yields
a gradient expansion
n

Jk =
Lik Fi
(4.9)
i=1

with
Lik

Jk
|J =0 .
(Fi ) k

(4.10)

Lik are called Onsager coecients. Within the context of the following they
obey the Onsager reciprocity relations [227]
Lik = Lki .

(4.11)

Inserting relation (4.7) into continuity equation (4.4) yields the desired transport equations
n

Xk
=
(Lik Fi ) .
(4.12)
t
i=1
A second way to obtain this transport equation is based on the principle
of minimal energy dissipation postulated by Onsager as well [227]:

!
dV = extremum ,
(4.13)
V

where is the so-called dissipation function. Within the context of evolution


of conserved extensive quantities is dened as
2

n


.
Lik (Fi )(Fj ) = P .

(4.14)

i,k=1

When applying the variational principle one has to take into account, that
the interest is not to obtain a free extremum of (4.13). Rather (4.13) has to
be solved together with the condition
P = 2 ,

(4.15)

since only expressions satisfying the linear thermodynamic relations proposed


by Onsager qualify as solutions. Introducing the Lagrangian parameter one
obtains

34

4. Thermodynamic Concepts of Phase-Field Modeling

or

 
V

+ (2 P )dV = 0




P
(Fi )dV = 0 ,
+ 2

(Fi )
(Fi ) (Fi )

(4.16)

(4.17)

respectively. Since the latter equation has to be fullled for all virtual displacements (Fi ), (4.7) and (4.14) can be employed to formulate
(2+1)

n


Lik (Fi )(Fk )

n


Fi Ji = 2(2+1)P = 0 . (4.18)

i=1

i,k=1

Comparing this with (4.15) yields = 1. Finally inserting into (4.16) results
in the free variation equation

(P )dV = 0 .
(4.19)
V

Now (4.7) and (4.14) together yield


0=

 
n
V i=1

n
1 
Lik (Fi )(Fk )dV
2

(4.20)

i,k=1

 
n

Ji Fi +


1
(Ji )Fi
Lik (Fi )(Fk ) dV

2
i=1
i,k=1

n


V i=1

Ji Fi dS .

(4.21)

The variation over the surface integral vanishes, since the boundaries of integration are xed. Together with continuity equation (4.4) this yields

 
n
n

Xi
1 
Lik (Fi )(Fk ) dV = 0 ,
(4.22)

Fi +

2
V i=1 t
i,k=1

with Lagrangian density


L=

Xi
t Fi

1
2

i,k=1

Lik (Fi )(Fk ).

To fulll (4.22), the EulerLagrange equation given by


L
d L
d L
d L

=0
F
F
k
k
k
Fk
dx x
dy y
dz F
z

(4.23)

has to be satised. Here x, y and z are the relevant space variables. The
variation has to be carried out solely with respect to the Fi . The gradient

4.1 Derivation of Transport Equations

35

terms, which denote the driving forces of the system, are not taken into
account. For this case the EulerLagrange equation can be rewritten as
L
L
L
L

=0.
Fk
k
k
x F
y
z
Fk

F
x
y
z

(4.24)

This results directly in the desired transport equations:


n


Xk
=
(Lik Fi ) .
t
i=1

(4.25)

Obviously (4.25) is identical to (4.12), thus revealing the equivalence of the


two approaches.
4.1.2 Extension to Non-conserved Quantities
If the density function P of the last section does not only contain conserved
quantities {X1 , ..., Xm }, but also non-conserved quantities {Ym+1 , ..., Yn }, the
calculations of the previous section have to be extended. In this case one can
write
P = P (X1 , ..., Xm , Ym+1 , ..., Yn ) .
(4.26)
In direct analogy to the procedure of the previous section the extended forms
of (4.2), (4.8) and (4.9) can be obtained as
m
n


P
P
dP (X1 , ..., Xm , Ym+1 , ..., Yn ) =
dXi +
dYi (4.27)
X
Y
i
i
i=1
j=m+1

m


Fi dXi +

i=1

n


Gj dYj

(4.28)

j=m+1

Jk = Jk (F1 , ..., Fm , Gm+1 , ..., Gn , F1 , ..., Fm , Gm+1 , ..., Gn )(4.29)


m
n


Jk =
Lik Fi
Ljk Gj .
(4.30)
i=1

j=m+1

In (4.29) and (4.30) ow functions Jk can only be formulated for conserved


quantities, thus index k runs from 1 to m. Inserting the above relations into
the continuity equation results in the transport equations for the conserved
quantities of the system
m
n


Xk
(Lik Fi )+
(Ljk Gj )
=
t
i=1
j=m+1

(k = 1, ..., m) . (4.31)

We are left with looking for transport equations for the non-conserved quantities. Generally temporal changes of the Yj depend on all of the Fi and Gj .

36

4. Thermodynamic Concepts of Phase-Field Modeling

Since we assume that deviation from thermodynamic equilibrium is small,


it is sucient to consider only non-vanishing terms of leading-order in Fi
and Gi . At thermodynamic equilibrium P takes an extremal value, so that a
necessary condition reads
P
P
=
=0.
Xi
Yj

(4.32)

(The extremal principle underlying this formulation and its relation to the
EulerLagrange equation is elucidated further in Appendix A of this chapter.)
If these conditions are not fullled for all Yj , usually the system has to be
considered as driven out of equilibrium. As a consequence the Yj will display
variations in time. Therefore, within the framework of linear approximation,
we can assume
Yk
P

t
Yj

(j = m + 1, ..., n)

(4.33)

for all k {m + 1, ..., n}.


If the equilibrium state is perturbed by deviations of one conserved quanP
(as one can take from the continuity
tity it is not sucient to change X
i
equation). Rather the gradients of this quantity have to be varied as well. To
ensure spatial isotropy of temporal changes in the non-conserved quantities,
to leading-order the dependence on the conserved quantity has to satisfy
Yk
P
2
t
Xi

(i = 1, ..., m) .

(4.34)

Together (4.31), (4.33) and (4.34) result in


m
n


Xk
P
P
(Lik
)+
(Ljk
)
=
t
Xi
Yj
i=1
j=m+1

(k = 1, ..., m)

m
n


P
P
Yk
=
(ik
)
(jk
)
X
Y
t
i
j
i=1
j=m+1

(k = m + 1, ..., n) .

(4.35)

Here the non-diagonal elements Lij and ij (i = j) describe cross coupling


eects. In the framework of solidication these are known to result in the
Soret [177] or the Dufour [152] eect.

4.1 Derivation of Transport Equations

37

If one assumes the system to depend on one non-conserved quantity


and two conserved quantities C and Q only, as well as the energy functional
to depend explicitly on the spacedependent observables and their gradients
only, the impact of the cross coupling eects can be analyzed starting from
the quadratic approximation
P =

A3 2 A4
A 5 2 A6
A1 2 A2
+
()2 +
C +
(C)2 +
Q +
(Q)2 . (4.36)
2
2
2
2
2
2

Again a close-to-equilibrium condition is assumed. The dynamics resulting


from this reads as follows


C =
(4.37)
t
Q


11 (A2 2 A1 ) 12 2 (A3 A4 2 ) 13 2 (A5 A6 2 )

L21 2 (A1 A2 2 ) L22 2 (A3 A4 2 ) L23 2 (A5 A6 2 ) C .


L31 2 (A1 A2 2 ) L32 2 (A3 A4 2 ) L33 2 (A5 A6 2 )
Q
Employing the ansatz

0
C = C0 eikr+t
Q
Q0

(4.38)

a periodic perturbation of the equilibrium situation with growth rates in


terms of the wave vector k can be carried out. Assuming constant phenomenological coecients, (4.37) and (4.38) result in


 + 11 (A1 + A2 k 2 ) 12 (A3 k 2 + A4 k 4 )
13 (A5 k 2 + A6 k 4 ) 

det  L21 (A1 k 2 + A2 k 4 ) + L22 (A3 k 2 + A4 k 4 ) L23 (A5 k 2 + A6 k 4 )  = 0
 L31 (A1 k 2 + A2 k 4 )
L32 (A3 k 2 + A4 k 4 ) + L33 (A5 k 2 + A6 k 4 ) 
(4.39)
with k = |k|. If k represents a wave length which is large compared to the
systems inherent length scales1 , terms of fourth order in k and higher can
be neglected. Thus the above determinant yields:
0 = 3 + (11 (A1 + A2 k 2 ) + L22 A3 k 2 + L33 A5 k 2 ) 2
+(L22 A3 + L33 A5 )11 A1 k 2 ,
with three solutions

11 (A1 + A2 k 2 )
=
+ O(k 4 ) .

2
(L22 A3 + L33 A5 )k

as e.g. the capillary length (i.e., k  1)

(4.40)

38

4. Thermodynamic Concepts of Phase-Field Modeling

According to this result, for k 1 the growth rate depends only on the
potential parameters Ai as well as the phenomenological coecients 11 , L22
and L33 , thus the diagonal elements. The dependence on non-diagonal elements is of higher order. Therefore it is justied to neglect the cross coupling
terms in (4.35).
Moreover one can deduce from the fact that (4.40) is independent of A4
and A6 that the dependence of the energy functional on the derivative of
the conserved quantities can be neglected as well. Thus the dynamics of the
locally conserved observables in (4.37) is reduced to a diusion equation2 .
Dening
Lk Lkk
;
k = kk
(4.41)
(4.35) can now be simplied to read
Xk
P
= (Lk
)
Xk
t
P
Yk
= (k
)
Yk
t

(k = 1, ..., m)

(4.42)

(k = m + 1, ..., n) .

Here the Lk are apart from a term resulting from the derivative of the
variation identical with the diusion constants Dk .
The diusion constants Dk usually depend on the Xk , too. E.g., for mass
transport in dilute binary mixtures the diusion constant is proportional to
the impurity concentration C. This can be obtained from setting the ow
density proportional to the product of a mobility assumed to be approximately constant and the concentration [185, 210, 295]. Within this limit the
thermodynamic potential contains a term proportional to C ln C. As a consequence the dynamics of the concentration eld is that of a diusion equation
with an eective constant diusion coecient.
Equations (4.35) and (4.42) depending on whether one intends to include cross coupling eects or not contain the basic variational principles
underlying the concept of diuse interface modeling developed in the following.

4.2 Introducing the Phase-Field Variable


The previous section gave a description of how transport equations of a thermodynamic quantity can be obtained by the variation of an appropriate thermodynamic potential with respect to that variable. The introductory chapter,
2

If one is explicitly interested in a study of cross coupling eects, terms of order


k4 have to be taken into account. As a consequence the diusion equation has
to be supplemented by 4 -terms, as well.

4.2 Introducing the Phase-Field Variable

39

on the other hand, gave a notion of a phase-eld variable as an essential


ingredient of diuse interface models. Thus if variational principles are to be
employed to obtain diuse interface models, the thermodynamic potential
has to be a function of the phase-eld variable, too. Accordingly one would
formulate an ansatz:

P (X1 , ..., Xm , Ym+1 , ..., Yn , )dV .
P(X1 , ..., Xm , Ym+1 , ..., Yn , ) =
V

(4.43)
The inclusion of into a thermodynamic potential as in (4.43) refers back
to work done by Fix [119], by Langer [200] and by Collins and Levine [74]
in the eighties. The basic idea seems to have been introduced by Landau
and Khalatnikov [271] in 1954 to formulate their theory for the absorption
of sound in liquid helium. Since then it has been used and developed further
by several authors (e.g. [33, 124, 263, 273]) in the context of phase transitions. In this context basic ideas were formulated by picturing the underlying
thermodynamic potential P to be the Gibbs free energy G with free energy
density G.
Extending the LandauKhalatnikov ansatz to describe non-equilibrium
growth phenomena is motivated by the success of introducing such an additional order parameter in the context of the above equilibrium thermodynamic phenomena3 . In the sense that for non-equilibrium systems the
-dependence of the Gibbs free energy cannot be obtained from microscopic
rst principle considerations, this introduction of remains phenomenological. Certainly this raises the question how to verify models based on a potential as given in (4.43) in a non-equilibrium system. Dierent approaches to
handle the question of verication have been pointed out in the introductory
chapter of this work. Here I will describe the basic ideas of how to construct
an appropriate -dependence of the energy functional based on work done
by Cahn and Hilliard [54, 55] and Allen and Cahn [7]. Further-reaching verication issues will be treated in succeeding sections.
In its generic form the CahnHilliard equation reads


C
= MC 2 f  (C) C 4 C .
t

(4.44)

As such it describes the evolution of a conserved concentration eld C during phase separation. MC has to be understood as a constant mobility, f is
a free energy. The term C 4 C arises as a penalty term from the extra
energetic cost associated with the transition region between the two phases.
3

Note that to derive the transport equations of a thermodynamic eld in the


context of a moving boundary system, there is an underlying assumption of instantaneous relaxation of elds to the respective interface position. In that sense
the assumption of being close to local thermodynamic equilibrium as necessary
for the calculations of the previous section, still holds for those elds. It is the
moving interface itself which turns the system into a system which is driven out
of equilibrium globally.

40

4. Thermodynamic Concepts of Phase-Field Modeling

The CahnHilliard equation applies for example to the phenomenon of spinodal decomposition. This is discussed in more detail in Appendix B of this
chapter.
The AllenCahn equation, on the other hand, describes the evolution of
a non-conserved order eld during anti-phase domain coarsening. Amongst
others things it has been used as a model equation for grain growth. A detailed
description of this example is given in Appendix C. In its generic form the
AllenCahn equation can be written as



= M 2 f  () .
t

(4.45)

Here denotes a local order parameter. It can be identied with the phaseeld variable appearing in the context of diuse interface modeling. One
would picture the latter to vary smoothly from a xed value in one phase to
another xed value in the other phase as pictured in Fig. 4.1. 4

Liquid

Interface

-1
Solid

x
Fig. 4.1. In the context of diuse interface modeling the phase variable varies
smoothly from a xed value in one phase to another xed value in the other phase.
Here = 1 is chosen to represent the solid phase, = 1 the liquid phase.

The CahnHilliard equation (4.44) and the AllenCahn (4.45) equation


dier from each other by the fact that the former satises an additional
conservation condition. The diuse interface models derived from a Cahn
Hilliard equation can be generalized to the viscous CahnHilliard equations.
4

I refrain from calling itself an order parameter to point out the dierence
between equilibrium phase transitions and the conditions of interfacial growth,
i.e. a non-equilibrium conditions in which the introduction of has to be motivated slightly dierently than in the context of the former. This is discussed in
some more detail in the remainder of this section.

4.2 Introducing the Phase-Field Variable

41

These contain the CahnHilliard equation and the AllenCahn equation in


respective limits. In that sense both equations can be treated equally on the
basis of one generalized diuse interface approach. The basic feature of such
a generalized description assumes that G is controlled by one contribution
originating from a potential V () which has two local minima at L and S .
L and S refer to the two separated phases, here liquid and solid. They
are also the two homogeneous state solutions the system can take. In the
following I will assume S to be 1 and L to be 1. V () is given by a
doublewell potential as depicted in Fig. 4.2.

V( )

V( )

V( V(
) + M( )
=1

=1

Fig. 4.2. The solid line visualizes the doublewell potential V () with the two
minima at = 1 and = 1 representing the homogeneous states S and L ,
respectively. The dashed line, on the other hand, displays how this potential is
unsymmetrized to favor one of the two phases and thereby trigger its growth at the
cost of the other. This eect is due to the contribution from M ().

Assuming that is an analytical function of space and time, V () can


be expanded into a Taylor series. Such an expansion up to fourth order in
yields a doublewell potential. Terms of uneven exponent vanish, since
V () represents a symmetric state, in which the two dierent phases are
energetically equal. As a consequence V () has to be invariant with respect
to the transformation . If the two minima of V () dene the energy
level 0, then V () reads
V () =

V0 2
( 1)2 ,
4

(4.46)

42

4. Thermodynamic Concepts of Phase-Field Modeling

with a scale factor V0 . Dropping the assumption that V () originates from


a Taylor expansion, and retaining only the requirement of two minima at
1 [23], one can dene V () section-wise as a construct of two parabolas [37,
38]. In that way it is possible to avoid a 3 inhomogeneity in the derived model
equations, which renders analytical treatment more dicult. For numerical
simulations, on the other hand, the instability at the crossing point of the
two parabolas is undesired, so that for those one usually uses a form of V ()
as given by (4.46).
To be able to consider inhomogeneous states as well, the energy density
functional has to depend also on the gradient of the phase-eld variable.
Lowest order terms which are compatible with invariance under rotation and
invariance under translation are 2 and ()2 . For the calculation of G
the relevant volume integrals of both terms proof to be identical except for a
surface integral. This surface integral, however, describes only boundary effects. It vanishes if the phase boundary is completely inside of the integration
volume and thus can be neglected. Therefore it is sucient to simply extend
(4.46) by ()2 and express G as:
G(, X, r) =

2
()2 + V () ,
2

(4.47)

where I have summarized the variables X1 , ..., Xm and Ym+1 , ..., Yn of (4.43)
as vector X and introduced the vector r to denote space coordinates.
The potential (4.47) still refers to an equilibrium condition. It is possible
to give a microscopic derivation for this potential as well. It can be obtained
on the basis of a lattice model for a two-component equilibrium thermodynamic system [48, 93, 118, 154]. Within such a model each of the lattice
squares x of a rectangular grid is occupied by a variable (x) {1; 1}. The
variable distinguishes between the two possible states. For the energy of a
given conguration the following ansatz is chosen

1
G=
(x) .
(4.48)
J(x x )(x)(x ) H
2 
x
x,x

The rst term denotes the interaction energy originating from the coupling
term J(x x ), the second term contains the energy contribution from external elds.
Often it is feasible to treat as a continuous variable. In this case it
is necessary to include a function W ((x)). W is a weight function which
determines the physically favored values of . With the introduction of W
one obtains the LandauGinzburgWilson functional as


1
G=
J(x x )(x)(x ) H
(x) +
W ((x)) .
(4.49)
2 
x
x
x,x

For W ((x)) a symmetric potential with second- and fourth-order terms is


the suitable form to describe a two-state system. For the case of a short-

4.2 Introducing the Phase-Field Variable

43

range interaction J(x x ) (e.g. nearest neighbor interaction) as well as vanishing external elds, (4.49) can be transformed into a continuity equation
analogous to (4.47) ( ) [48, 154]. Thus for thermodynamic equilibrium
congurations microscopic and macroscopic derivation yield the same ansatz
for G(, X, r). Furthermore the microscopic approach provides additional insight into the nature of diuse interface models. It yields an understanding,
that the nite interface thickness of a diuse interface model and the surface energy term originate from nite correlation lengths on a microscopic
scale. Moreover assuming nearest neighbor interaction and constant coupling constant J interface thickness and surface energy are proportional to
J.
For non-equilibrium conditions a coherent microscopic derivation is missing. From a macroscopic point of view, (4.47) can be extended to account for
non-equilibrium conditions by introducing a further energetic term M (, X)
such that
2
G(, X, r) = ()2 + V () + M (, X) .
(4.50)
2
The contribution M (, X) depends on the special physical properties of the
system under consideration, e.g. the underlying phase diagram. Together with
V () this additional term M results in a doublewell potential, which is no
longer symmetrical with respect to . This is depicted in Fig. 4.2.
The dynamic equations obtained from (4.50) owe the driving force, which
moves the interface, to the term M (, X). In this context the evolutionary
equation for the phase-eld variable can be interpreted as a mean-eld
approximation of the non-equilibrium interface dynamic.
Analytical as well as numerical treatment of model equations obtained
from (4.50) is simplied by the localizing of the minima of the unsymmetric
doublewell potential V () + M (, X) at 1. This yields
M (, X)
| =1 = 0 .

(4.51)

Equation (4.51) has to be valid for arbitrary X. One can ensure the latter
choosing the ansatz
M (, X) =

n


M,k ()MX,k (X) ,

(4.52)

k=1

where

M ()
| =1 = 0 .
(4.53)

In this way one avoids the undesired eect that within a pure liquid or a pure
solid, respectively, changes to the equilibrium value of the phase eld result
in emission of latent heat.
If one would classify the dynamics of interfacial growth we are interested
in here according to a scheme of dynamic universality classes proposed by

44

4. Thermodynamic Concepts of Phase-Field Modeling

Hohenberg and Halpering in [146] 5 it would correspond to their model C.


A classication as in [146] can be done based on a thermodynamic functional
as G(X, ) in (4.47). However, the specic functional employed in [146] is
the Helmholtz free energy functional, which I will denote by A(X, ) in the
following. Even without a precise expression for A(X, ) one can distinguish
between three dierent types of evolution equations obtained for the phaseeld variable in the following manner:
1. Model A states a purely relaxational evolution of . It can be written
as

A
= K
.
(4.54)
t

Here K is a positive coecient which might depend on and on the temperature. Equation (4.54) can be understood to model systems where an
internal order parameter expressed by characterizes a non-equilibrium
structure which eventually6 approaches an equilibrium value obtained by
minimizing the free energy function.
2. Model B applies to a system in which an order parameter represents an
equilibrium thermodynamic variable, e.g. the concentration. In this case
a generic evolution equation based on A reads
A

= (M
),

(4.55)

where M is a positive coecient related to diusivity. This expression


was introduced by Cahn as a model for spinodal decomposition in 1961
[57].
3. Model C is the most specic of the three generic models. It describes
the dynamics of a system for which one order parameter as given by
is insucient to describe the local state of the system. Additional elds
need to be incorporated into the coupled model equations as well. The
most popular model for this case involves the temperature T in addition
to [143] and thus applies to solidication of pure undercooled melts:

= 2 2 + g() f (T )
t

= 2 f .
t
t

(4.56)
(4.57)

Here f is a strictly increasing function of the temperature, g is a negative


doublewell function, , and are positive constants.
As stated above, the coupled set of equations (4.56) and (4.57) can be
viewed as one of the rst diuse interface or phase-eld models for interfacial
growth. Introduced in the same manner as by Halperin et al., it remains a
5
6

An earlier approach in similar direction can be found in [143].


Here eventually refers to the limit t .

4.2 Introducing the Phase-Field Variable

45

phenomenological approach. The reason is that even though (4.56) is obtained


formally by carrying out a variation according to (4.42), (4.57) is constructed
merely by common sense introducing the eects due to the moving interface
as source term
t by hand. Moreover, strict relaxational behavior of the
potential during the course of evolution is not ensured. I will come back to
this point in the discussion following (4.62).
In 1990 Penrose and Fife managed to place this model upon a formally
more rigorous foundation. To do so they expressed (4.56) and (4.57) in terms
of an entropy functional (see [231], p.50). This results in
2

S
=K
t

(4.58)

1
u
= (M ) .
t
T

(4.59)

Here K and M are positive constants similar to those above, and u is a


homogenous energy density. In [231] (p.51) it is given by
u f (T )() + e =

d ah (t,)
T
,
d T1

(4.60)

where and e are dened as the number of degrees of freedom per unit
volume and the potential energy, respectively. ah (t, ) is the homogenous
contribution to the Helmholtz free energy density. The entropy functional
dened by Penrose and Fife (see [231], p.49) reads

1
2
{s(u, ) ||
}dV .
(4.61)
S=
2
V
Here = /T . s represents the homogenous entropy density dened in the
usual sense via the thermodynamic relation
s=

ah (t, )
u

.
T
T

(4.62)

Thus applying standard thermodynamic relations to model C, Penrose and


Fife used homogeneous densities only.
The great dierence between model equations (4.56)(4.57) compared to
(4.58)(4.59) is that the latter are thermodynamically consistent, whereas the
former are not. As introduced in Chap. 1 thermodynamic consistency refers
to a strict relaxation of thermodynamic potentials during the evolution of
the respective system.
The fact that models obtained from a Helmholtz free energy (like model
equations (4.56) and (4.57)) are not thermodynamically consistent, originates from the point that changes in temperature trigger changes of the inner energy as well. Those changes would have to be integrated into the above
derivation. This happens naturally, if one transforms from the Helmholtz free

46

4. Thermodynamic Concepts of Phase-Field Modeling

energy as underlying potential to the entropy, as well as from the temperature


as independent variable to the inner energy density u.7
Remembering that in the beginning of this section the introduction of a
phase-eld variable in non-equilibrium growth processes could be motivated
only phenomenologically, renders the concept of thermodynamic consistency
an important backbone of phase-eld modeling. It places the usage of a phase
eld upon a solid theoretical foundation. In this sense diuse interface models
ensuring strict relaxational behavior of the potentials are more but a phenomenological description of an interfacial growth problem.
However, keeping in mind that the classical approach to interfacial growth
problems, the so-called sharp interface approach (see Chap. 1), is phenomenological in nature, too, employing a diuse interface approach, it might not
always be necessary to formulate the model to be thermodynamically consistent. Just as well it can be justied to work with a phenomenological model
as model C, if the analogy to an established sharp interface model can be
proven via an asymptotic matching expansion as described in Chap. 5. From
a numerical point of view this is often more ecient. In this sense within
the numerical appendix of this book the concept of universality of dierent
diuse interface models, thermodynamically consistent ones as well as phenomenological ones, will be addressed for the very classical phenomenon of
purely diusion limited growth. However, it has to be stressed again that the
methodology of thermodynamically consistent diuse interface modeling is
the one, which opens new perspectives if diuse interface modeling is applied
to more complicated physical systems. For such cases many important issues
are still researched. Chapter 6 will therefore address the topic of thermodynamic consistency again in the context of an example, which demonstrates,
how the methodology can be used to learn more about the physics of more
complicated interfacial growth phenomena. The remainder of this chapter
will merely summarize some approaches to ensure thermodynamic consistency developed so far.

4.3 Thermodynamic Consistency


As described in Sects. 4.1 and 4.2 the starting point to obtain the evolutionary equations of a diuse interface model is given by the variational principles (4.35) or (4.42), respectively. On the other hand, applying such methods to a somewhat phenomenologically8 introduced phase-eld variable, does
not necessarily yield a thermodynamically consistent dynamics. This implies
that strict relaxational behavior of the thermodynamic potentials during the
course of evolution is not necessarily ensured. The rst part of this section
7
8

just as it was done in (4.58) and (4.59)


The notion phenomenological refers to the respective discussions in Sects. 4.1
and 4.2.

4.3 Thermodynamic Consistency

47

describes a comprehensive approach to extend the variational methods of the


previous sections in such a way that they guarantee conformance with thermodynamics. This comprehensive approach assumes that interfacial motion
is driven either by surface diusion or by curvature. In that sense this approach is not suitable to model general interfacial growth problems. However,
it reveals two important aspects:
1. How kinetics can be integrated into a variational approach, and
2. how thermodynamic consistency can be guaranteed by a variational
method.
The approach ensuring the above is called gradient ow method. It was
developed intensively in the mathematical moving interface community
in the nineties [8, 63, 265, 268, 269]. An important concept, which arose
along with that development is the CahnHomann capillary vector to model
anisotropic surface properties in the context of curvature driven interfacial
motion [58, 59, 145, 267]. It is an essential ingredient of any diuse interface model for interfacial growth phenomena, which is meant to recover the
anisotropy of surface energetics.
As mentioned before, for the inclusion of further physical mechanisms or
elds (e.g. a hydrodynamic eld) a comprehensive approach is still missing.
Nevertheless by expressing local entropy production in terms of the transport quantities of derived evolutionary equations and identifying the forms
of these quantities which ensure it to be non-negative, consistent diuse interface models have been obtained for even more complex systems. One example
along that line is the approach by Anderson, McFadden and Wheeler to solidication taking into account hydrodynamic convection [15]. In this case the
analysis, in which the thermodynamically consistent diuse interface model
emerges, is case sensitive adapted to the individual physical situation. As an
example for such a case sensitive ansatz compared to the above comprehensive framework, the main steps of their model approach is summarized in
the second part of this section. The question of what remains to be done to
nd a general approach for any interfacial growth phenomenon, complex as
it might be, is left to Chaps. 6, 7 and 8.
4.3.1 The Gradient Flow Method
The basic idea underlying the gradient ow method is to guarantee consistency with the laws of thermodynamics by constructing a gradient ow which
decreases the energy as much as possible at a given time step. This gradient
ow is obtained from an inner product which yields a kinetic measure of the
time it would take for a system to evolve from one microstructure to another.
Mathematically its derivation reads

A 
= A, v .
(4.63)
t t=0

48

4. Thermodynamic Concepts of Phase-Field Modeling

Just as in the previous sections A can be understood to be the functional of a


free energy density. For the inner product A, v a norm has to be specied.
A will depend on the choice of the inner product, i.e.
A, v = , A, v .

(4.64)

The most important norms for the discussion of interfacial motion are the
L2 norm and the H 1 norm dened by:

f vdV ,
(4.65)
f, vL2 = f v =
V


f, vH 1 =

f v dV =

 2 1 
( ) f vdV ,

(4.66)

where f is a solution of the Poisson equation


2 f = f (r) .

(4.67)

Applying this formalism to interfacial motion triggered by surface diusion,


weighting of the inner products with a function M (r) allows me to take into
account anisotropic mobility constants. This modies the inner products as

fv
dV ,
(4.68)
f, vL2 ,M =
V M



f, vH 1 =
(M )1 f vdV .
(4.69)
V

In analogy to the discussion following (4.54) of Sect. 4.2 for the free energy
density an ansatz a((r)) = ah + 2 () can be chosen. It enters (4.64)
via the functional A = V adV . Here the dependence of 2 () on the
direction of is directly proportional to the orientational dependence of the
anisotropic interfacial energy density [268]. As in Sect. 4.2 ah does not need
to be specied any further. It is the choice of a specic norm which results in
a specic kind of dynamics. An underlying H 1 norm, for example, yields a
generalized anisotropic CahnHilliard equation (see (4.44))
C
= M (f  (C)  ) .
(4.70)
t
On the other hand, applying the L2 norm for a non-conserved parameter
generalizes the AllenCahn equation (see (4.45)) to anisotropic surface
physics yielding

= M (  f  ()) .
(4.71)
t
Other specications of A will result in still dierent dynamical behavior, e.g.
purely relaxational kinetics or the Mullins surface diusion equation [222]. In
any case the dynamics is obtained in a thermodynamically consistent formulation. In this sense and in the sense of inclusion of anisotropic surface physics,
even if only applicable to a very restricted set of physical situations, (4.64)
results in a more advanced variational approach than the model Ctype of
approach of the previous section.

4.3 Thermodynamic Consistency

49

4.3.2 Entropy Production in Terms of Transport Variables


Since the above gradient ow method does no longer hold, if one turns to
more complicated interfacial growth phenomena, one is left with the variational principles of the previous sections and has to ensure thermodynamic
consistency additionally by hand. One way to proceed is to follow the formalism of irreversible thermodynamics [91] and embed the variational principles
applying to diuse interface models into this framework. This procedure is
for example underlying the work, by which Penrose and Fife managed to
put model C onto a more fundamental foundation [231] (see Sect. 4.2). It
has been introduced to dierent growth conditions treated on the basis of
a diuse interface approach by several authors since [283, 289]. In a similar
manner one might start from transport equations obtained via balance law
formulations, express local entropy production in terms of the transport variables and identify the forms, which ensure it to be non-negative. To illustrate
this procedure here the basic steps to derive a diuse interface model for
solidication inuenced by hydrodynamic convection presented by Anderson
et al. [15] are summarized in the succeeding paragraphs:
Following [15] I assume that the total entropy S within a volume V (t) is
given by



1 2 2
s S () dV .
(4.72)
S=
2
V (t)
Here and s are density and entropy per unit mass, respectively. The rst
term in the integrand (i.e. s) is the classical entropy density. The second is
a non-classical term associated with spatial gradients of the phase eld. For
simplicity the gradient entropy coecient S is assumed to be a constant.
is a homogeneous function of degree unity. As introduced in Sect. 4.3.1, this
function allows for a general anisotropic surface energy of the solidliquid
interface.
For the total mass M, linear momentum P and internal energy E associated with the volume of the material the following ansatz is chosen:

M=
dV ,
(4.73)
V (t)

P =
udV,
(4.74)
V (t)



1
1 2 2
e + |u|2 + E
() dV.
2
2
V (t)


E =

(4.75)

Here u is the velocity, e the internal energy density per unit mass and E the
gradient energy coecient, which is assumed to be constant. The thermodynamic relations
p
e
de = T ds + 2 d +
d ,
(4.76)

e = T s p/ +
(4.77)

50

4. Thermodynamic Concepts of Phase-Field Modeling

are assumed to apply locally, where p is the thermodynamic pressure and


the chemical potential per unit mass.
The physical balance laws for mass, linear momentum and internal energy
are given by
dM
=0,
dt

dP
=
n
mdV ,
dt
V (t)


dE
qE n
dA =
n
m udA ,
+
dt
V (t)
V (t)

(4.78)
(4.79)
(4.80)

respectively. Here n
is the outward unit vector normal to V (t), m the stress
tensor9 , and qE the internal energy ux. Momentum balance given by (4.79)
requires that the total momentum ratechange of the materials volume results from forces acting on its boundary V (t). External forces such as gravity are neglected. The equation for energy balance (4.80) equates the rate of
change of the total internal energy of V (t) plus the energy ux through its
boundary to the rate of work of the forces at this boundary V (t).
In addition, entropy balance is given by


dS
qS n
dA =
s prod dV (t) ,
(4.81)
+
dt
V (t)
V (t)
where qS is the entropy ux and s prod the local rate of entropy production10 .
The second law of thermodynamics requires it to be non-negative.
To proceed, Anderson, McFadden and Wheeler rewrite the conservation
laws (4.78)(4.81) as dierential equations. These are used to express the
local entropy production in terms of the uxes m, qE , and qS , as well as
the material derivative of which is given by D/Dt =
t + (u ). Then
they identify forms for these quantities which ensure that the local entropy
production is non-negative. The uxes that result from this procedure involve
both classical and non-classical contributions. In addition, they obtain an
evolution equation for the phase eld.
Applying the Reynolds transport theorem [19] to the mass balance law
(4.78) yields the continuity equation in its conventional form

10

D
+ u = 0 .
Dt

(4.82)

Tensors and vectors are both denoted by bold print. The only exception is the
linear momentum P, which for clarity purposes is expressed with a vector symbol
to distinguish it from the scalar functions M, E and S expressed in calligraphic
letters as well.
Note that due to the evolving boundary V (t) is changing in time. For simplicity
I will drop the dependence on t in the following.

4.3 Thermodynamic Consistency

51

Again the material derivative Dt


= t
+ (u ) has been used. Similarly, the
linear momentum equation yields the expression

Du
= m,
Dt

(4.83)

where m has components mjk /xj (mjk denotes the components of m).
The derivation of the energy equation is more involved. It requires to cast
(4.80) into


 
d
De
1 2 2
Du

()dV = 0 .
+ u
(m u) + qE dV +
Dt
Dt
dT V 2 E
V
(4.84)
Further the identity


d
1 2 2
E QG dV ,
(4.85)
E ()dV =
dT V 2
V
which is proven in the appendix of [15], is employed, where


D
1
D

( ) u : + 2 u . (4.86)
QG =
Dt
Dt
2
Here refers to the outer tensor product and : to the double contraction
of the tensor product. denotes the CahnHomann capillary vector mentioned in the introductory part of this section. It follows that

Du
De
2
+ u
(m u) + qE + E
QG = 0 .
Dt
Dt

(4.87)

Conservation of linear momentum as given by (4.83) can be used to rewrite


this equation as

De
2
+ u ( m) (m u) + qE + E
QG = 0 .
Dt

(4.88)

With the identity


(m u) = ( m) u + m : u

(4.89)

this simplies to yield the energy equation as

De
2
QQ ,
+ qE = m : u E
Dt

(4.90)

where the double contraction of the tensor product is given by m : u =


mkj uj /xk (with summation over repeated indices implied). In direct analogy, entropy balance as given by (4.81) results in

Ds
+ qS = s prod + S2 QG .
Dt

(4.91)

52

4. Thermodynamic Concepts of Phase-Field Modeling

The thermodynamic relation between Ds/Dt and De/Dt follows from (4.76).
It is given by
Ds
p D
e D
De
=T
+ 2
+
.
(4.92)

Dt
Dt
Dt
Dt
At this point the continuity equations (4.82) and (4.86), as well as (4.90) and
(4.92) can be employed to express entropy production as

 

1
2
s prod =
m + F2 + p F 2 I : u
T
2


1
e D
+
2 ( )
2 F
Dt


2
qE
D
F
+ qS
T
T
Dt


1
D
2
( ) .

+ qE + E
(4.93)
Dt
T
Here 2F = 2E + T 2S.
The constitutive equations for the uxes and D/Dt can still be specied
ensuring positive s prod


F2 2
I F2 + ,
m
= p
(4.94)
2
e
D
= F2 ( )
,
(4.95)
M
Dt

1 2 D ,
qE
= k
(4.96)
E
T
Dt
1 2 D .
(4.97)
= k
qS
S
Dt
T
Here is the viscous stress tensor, which for a Newtonian uid is given by
= (u + uT ) + (u)I. and are coecients of viscosity, I is the unit
tensor, and M is a constant positive mobility coecient. Moreover, a constant
value for the thermal conductivity k is chosen corresponding k = T 2 k. The
resulting coupled set of equations is given in [15] by
D
Dt
Du

Dt
D
M
Dt
De

Dt

= u ,



F2 2
2
= p +
I F + ,
2
e
= F2 ( )
,

D
2
= [kT ] + E
( )
Dt



S2 2
+ p + I T S2 + : u .
2

(4.98)
(4.99)
(4.100)

(4.101)

4.4 Appendix

53

Equations (4.98)(4.101) constitute a diuse interface model for hydrodynamically inuenced solidication. It is a thermodynamically consistent
model. Here its derivation serves as an example for a second approach to
ensure thermodynamic consistency after the discussion of the gradient ow
method in Sect. 4.3.1. It is carried out by expressing entropy production
in terms of the transport quantities of derived evolutionary equations and
identifying the forms of these quantities that ensure that the local entropy
production cannot be negative. One can interpret the thermodynamic consistency of (4.98)(4.101) as a validation of this model. On the other hand,
analyzing such thermodynamically consistent model equations based on the
method of matched asymptotic expansion (see Chap. 5) might revise classical sharp interface equations and thus contribute to new insight into the
respective interfacial growth phenomena as I will demonstrate in Chap. 6.

4.4 Appendix
Appendix A: Excursion to Some Basic Calculus of Variation
To derive appropriate transport equations for non-conserved thermodynamic
variables, in Sect. 4.1.2 we started out from the consideration, that directly at
thermodynamic equilibrium the potential P should take an extremal value.
P is dened (see (4.1) and (4.26)) as

P(X1 , ..., Xm , Y1 , ..., Yn ) =
P (X1 , ..., Xm , Y1 , ..., Yn )dV ,
(4.102)
V

where P is the respective density function, the Xi with i = 1, ..., m conserved


variables and the Yj (j = 1, ..., n) non-conserved variables. Then requiring
.
P(X1 , ..., Xm , Y1 , ..., Yn ) = extremum ,

(4.103)

P
is formally a classical variational problem. We could rewrite the variation X
i
within (4.32) as


P
d
P(Xi + s )
,
(4.104)
:=
ds
Xi
s=0

where the variable is dened on V (just as the Xi and Yj ). If Xi minimizes


P
(maximizes) P, then X
= 0 for all such . The classical strategy for solving
i
the problem
!
P = minimum
(4.105)
consists in solving the respective EulerLagrange equations and then investigating whether a solution of the equations is a minimum of P or not. Assuming that the Xi depend solely on the time t, the corresponding Euler
Lagrange equations read:

54

4. Thermodynamic Concepts of Phase-Field Modeling

d
Pp (t, Xi (t), Xi (t)) Pn (t, Xi (t), Xi (t)) = 0 ,
dt

(4.106)

where Pn is a vector of the partial derivatives of P with respect to the Xi and


Pp an analogous vector of the partial derivatives of P with respect to the Xi ,
which are the derivatives of the Xi with respect to time. For a derivation of
the EulerLagrange equations the reader is referred to introductory courses
into classical mechanics as e.g. [134]. Here I will conclude this excursion on
variations with an example to elucidate the relation between an extremal
principle and the EulerLagrange equations somewhat further.
Example: The question I want to answer is how to minimize the arc
length of the graph of a function u : [a, b] IR, i.e. the length of a curve
(t, u(t)), which is a subset of IR2 , among all graphs with xed boundary values
u(a), u(b). This leads to the variational problem

a

1 + u (t)2 = minimum .

(4.107)

To compute the relevant EulerLagrange equation we rst have to evaluate


the partial derivative with respect to u as well as the partial derivative with

respect to u . They read Pu = 0 and Pp = u (t) 2 , respectively. Thus the
1+u (t)

Euler-Lagrange equation is given by


0=

u (t)
d
u (t)
u (t)2 u (t)

= 

3
dt 1 + u (t)2
1 + u (t)2
1 + u (t)2
= 

i.e.

u (t)
1 + u (t)2

3 ,

u (t) = 0 .

(4.108)
(4.109)

(4.110)

This implies that u is a straight line between a and b, which in this case is a
solution obvious at rst sight.
Appendix B: The CahnHilliard Equation
As mentioned in Sect. 4.3 the CahnHilliard equation is a model for the evolution of a conserved concentration eld C during phase separation. Phase
separation might for example describe the change of a binary alloy from a uniform mixed state to that of a spatially separated two-phase structure. Such
a change occurs, if the temperature of the system is reduced rapidly below
a critical temperature T . Comparing the Gibbs free energy G belonging to
a two-phase structure to that of a mixed state results in the phase diagram
depicted schematically in Fig. 4.3. The so-called spinodal and conodal

4.4 Appendix

55

Temperature

stable
concentrations

Conodal line

metastable
concentrations
unstable
concentrations
Spinodal line

Concentration

Fig. 4.3. Phase diagram indicating spinodal line (inner curve) and conodal line
(outer curve). The lines separate regimes of stable, unstable and metastable concentration for a system undergoing phase separation.

lines separate regimes of stable, unstable and metastable concentrations, respectively.


In Fig. 4.4 the temporal evolution of a binary alloy into a separated twophase structure starting from a mixed state is displayed. The red and green
coloring indicates, which of the two phases is dominant. Taking into account
the actual value of the dominant phase, the blurred phase transition and the
intensity of the coloring give an impression of the state of the separation.
Appendix C: The AllenCahn Equation
In contrast to the CahnHilliard equation, the AllenCahn equation describes
the evolution of a non-conserved order parameter eld. This applies for example to the growth of grains in a crystalline sample. If we consider the sample
to be composed of several grains, then surface motion does not necessarily
have to conserve the amount of material within each individual grain. The
total energy of the system would be the sum of the surface energies of the
grain interfaces plus the bulk energies of the phases, i.e.:


F =
(n (x))dA +
F dV .
(4.111)

xS

At grain junctions usually several equilibrium congurations are possible


depending on the shape of the grains. Assuming all grains to be regular
octagons, three dierent equilibrium congurations might arise (Fig. 4.5).

56

4. Thermodynamic Concepts of Phase-Field Modeling

Fig. 4.4. Phase separation of a binary alloy as modeled by the CahnHilliard


equation (4.44). The single pictures within the gure represent consecutive states
of the system during its temporal evolution from a mixed state (upper left picture)
towards a separated two-phase structure (lower right picture). Each snapshot depicts spatial domains for which the two colors red and blue indicate a dominance
of either of the two phases. The transition occurs smoothly via yellow and green
with respective shading.

(a)

(b)

(c)

(d)

Fig. 4.5. Equilibrium congurations at grain junctions picturing all grains as


regular octagons. (a) Geometry of the grain itself. (b)(d) The three possible congurations at the junction points for this underlying grain geometry. From point of
view of energetics other congurations as < are identical to the ones depicted in
(b)(d) due to the parity symmetry of the system.

4.4 Appendix

57

The number of equilibrium triple junctions is even larger if grains are assumed
to be regular hexagons. The six possible congurations belonging to this case
are depicted in Fig. 4.6.
To determine the temporal evolution of such interface congurations one
can employ the following variational formulation: The minimum of
F, v +

1
v, v
2t

(4.112)

occurs at v = F t. Moreover, the energy change F (Sv )F (S) for moving


an initial surface conguration S to a new surface conguration Sv is well
approximated by F, v. Thus one can obtain an iteration for the evolution
of the interface conguration by minimizing
F (Sv (0)) F (S(0)) +

1
v, v .
2t

(4.113)

Here S(0) is an arbitrary initial conguration. This procedure has to be


applied to the congurations S(n t) (n = 1, 2, ....) until convergence is
obtained.
Now given a line segment Si within a two-dimensional interface conguration S , the energy change for moving segment Si in normal direction
over a distance vi is given by vi (fi+ fi ). Here fi can be obtained from
geometrical considerations. This is illustrated in Fig. 4.7, where n = n(Si )
refers to the normal of line segment Si and li to its length. Formally one

(g)
(a)

(b)

(c)

(d)

(e)

(f)

Fig. 4.6. Equilibrium congurations at grain junctions picturing all grains as


regular hexagons. (a) Geometry of the grain itself. For this kind of underlying
grain geometry six dierent congurations at junction points are possible. They
are displayed in graphs (b)(g). Again other congurations as < can be mapped
onto the ones depicted here due to the parity symmetry of the system.

may write


fi (Si ) = i ( (ni ) nni (n)) / 1 (nni )2 .

Thus the expression to minimize is




vi2 li
1
.
vi (fi+ fi ) Fi li +
2t M (n(Si ))
i

(4.114)

(4.115)

58

4. Thermodynamic Concepts of Phase-Field Modeling

ni

ni+

z (ni+ ) csc

Si

z (n) cot
Fig. 4.7. Sketch of how the geometry at a junction point determines the contributions to the energy conguration of a line segment Si given by fi+ fi . Here the
corner labeled by ni+ is depicted. Thus the gure displays the geometric quantities
applying to the calculation of fi+ . In a similar manner fi can be obtained from
an analogous construction at the left grain corner of line segment Si . fi+ is determined from the line tension in direction of ni+ , i.e. (ni+ ), minus the line tension
(n) projected in direction of ni+ , i.e. (n)nni+ , normalized with respect to

1 (nni+ ) (see (4.114)).

Carrying out the minimization results in




fi+ fi
t .
vi = M (n(Si )) Fi
li

(4.116)

One recognizes its equivalence to the AllenCahn equation (4.45) by identifying vi in the nal expression (4.116) with
t in (4.45) and realizing that

=
F
.
This
in
turn
implies
that
the
AllenCahn
equation contains all
t
of the features of non-conserved grain growth discussed above.

5. Asymptotic Analysis

The previous chapter was devoted to a detailed discussion of the thermodynamic concepts underlying the formalism of diuse interface models. This
chapter, on the other hand, introduces a further method required to employ
diuse interface modeling to gain new insight into interfacial growth phenomena: This is the method of matched asymptotic expansion. Classically
this method is employed to establish the equivalence of sharp interface and
diuse interface models and thus is an important concept to validate diuse
interface models. I introduce the method of matched asymptotic expansion
in Sect. 5.1 within the general mathematical framework of matching to provide a comprehensive background for its application in the context of diuse
interface modeling. This section is succeeded by the discussion of a diuse
interface model for thin lm epitaxial growth, which serves as example for
the use of the asymptotic matching method purely for validation purposes.
On the other hand, as mentioned in the introductory chapter, an asymptotic analysis can also serve to nd or at least reconsider sharp interface
models of growth phenomena for cases in which the latter are not well established. A necessary requirement to employ diuse interface models for this
task is to ensure thermodynamic consistency of the diuse interface model.
Usually this is an additional step yielding further restrictions after the diuse
interface model has been established via a variational procedure as described
in Chap. 4. In this context an analysis ensuring thermodynamic consistency
automatically would be very desirable. Only lately such a generalized analysis
could be formulated for a restricted class of interfacial growth problems [102].
It is described in detail in Sect. 5.3. Finally, in the discussion section at the
end of this chapter, the methods presented here are placed into the overall
context of diuse interface modeling.

5.1 A Formal Mathematical Approach


Towards Matching
The following is a mathematical excursion about matching. It contains the
precise mathematical denition of matching and thereby provides the formal
framework of the method of matched asymptotic expansion. As indicated in
Heike Emmerich (Ed.): LNP m73, pp. 5996, 2003.
c Springer-Verlag Berlin Heidelberg 2003


60

5. Asymptotic Analysis

the introductory chapter within the context of diuse interface modeling only
recently one has considered the relevance of higher order matching conditions.
From point of view of the underlying mathematical theory obtaining relevant
matching conditions does not only depend on determining relevant orders of
the expansions, but also on the construction of extended domains of validity.
In the following this issue is discussed in its formal mathematical framework
and illustrated by hand of a simple example. The discussion of this formal
framework follows [31, 115, 197]. Its connection to its application within the
context of matched asymptotic expansion as necessary for the analysis of
diuse interface models is given in Sect. 5.1.2.
Here I start from the assumption that D = [0, 1], I = (0, 1 ) and y(x, )
are continuous on D I. Furthermore, I suppose that there are yk (x) such
that the outer expansion
y y0 (x) + y1 (x) + 2 y2 (x) +

(5.1)

is uniformly valid on [
x, 1], x
> 0, as 0+ . The inner variable reads
X=

x
.

(5.2)

Moreover, I assume Yk (X) to exist such that


y Y0 (X) + Y1 (X) + 2 Y2 (X) +

(5.3)

for some X
< 1/1 as 0+ . This is identical to
uniformly on [0, X]
the assumption that there is both an inner and an outer expansion for the
same function y(x, ). The function y(x, ) can be viewed as a solution of the
algebraic problem
f (x, y, ) = 0
(5.4)
or a boundary-value problem like
d2 y
dy
+ b(x) = f (x, )
+ a(x)
dx2
dx
y(0, ) = A
and
y(1, ) = B .

L [y]

with

x (0, 1) ,

(5.5)
(5.6)

For the outer limit in (5.1), x is xed. For the inner limit in (5.3), X is
xed.
denote the regions of uniform validity for the
x) and Di (X)
Now let Do (
outer and inner expansions, respectively:
x) = {(x, ) : x [
x, 1], I}
Do (

I} .
Di (X) = {(x, ) : x [0, X],

(5.7)
(5.8)

depend on the -independent, xed values x


x) and Di (X)

By denition Do (
The precise values proof to be irrelevant for the discussion of overlap
and X.

5.1 A Formal Mathematical Approach Towards Matching

61

regions and matching. Thus I denote these regions simply by Do and Di .


This underlying picture of Do and Di is important to understand the following extension theorem necessary to dene matching. Extension theorems
are theorems which extend the region of uniformity of asymptotic statements
like (5.1). An early theorem is due to Kaplun [161]:
Theorem 1 Let D = [0, 1], I = (0, 1 ) and y(x, ) be continuous on D I.
Also, let y0 (x) be some continuous function on (0, 1] such that
lim [y(x, ) y0 (x)] = 0

0+

(5.9)

uniformly on [
x, 1] for every x
> 0. Then there exists a function 0 < ()
1
such that
(5.10)
lim [y(x, ) y0 (x)] = 0
0+

uniformly on [(), 1].


(See [99] for further extension theorems.) Obviously there is a whole range
of functions satisfying this theorem. One simple example is
y(x, ) = x + ex/ +

y0 (x) = x .

(5.11)

x, 1].
The limit (5.9) implies y(x, ) y0 (x) + O(1) uniformly on [
Basically the eect of the Kaplun extension theorem is to extend the
o
region of uniform validity Do to a layer around this region, which I call D

in the following. To more carefully dene Do , intermediate variables need to


be introduced. Let () be any function with 0 < ()
1. Here I dene the
intermediate variable x by
x = ()x .

(5.12)

Then, in accordance to Theorem 1, one may state


lim

0+ ,x xed

[y(x , ) y0 (x )] = 0

(5.13)

x , 1] for all with = O(). Generally, when introducing


uniformly on x [
intermediate variables one views as satisfying

1, though to dene
o one usually sets equal to or 1:
D
o (
x ) = {(x, ) : x [
x (), 1], I} .
D

(5.14)

In (5.11), for example, there is an intermediate variable x with


y(x, ) y0 (x) = e

+ = O(1)

(5.15)

> 0 and
. Here one could for example
uniformly on [
x , 1] provided x
choose () = 1/2 in accordance to the theorem.

62

5. Asymptotic Analysis

In an analogous manner one can construct an extended domain of validity


i for the inner expansion (5.3) choosing as inner variable
D
X=

()x
.

(5.16)

For some (x, ) near (0, 0) the non-extended domains Do and Di do not
In
overlap, i.e. they do not intersect regardless of the choices of x
and X.
a similar fashion one can construct non-overlapping extended domains and
overlapping extended domains. If there is an overlapping extended domain,
there are functions i () and o () such that for any intermediate variable
x with i ()
()
o () inner and outer expansion are uniformly valid.
This implies that given any with i ()
()
o (), there is an independent interval I such that both
=0
[y(x , ) y0 (x )]

 x 

lim
=0
y(x , ) Y0

0+ ,x xed
lim

0+ ,x xed

(5.17)
(5.18)

are uniformly on x I , x > 0. Subtracting these expressions one obtains


a matching condition:

 x 

=0.
(5.19)
lim
y0 (x ) Y0

0+ ,x xed
Further, if y0 (0+ ) and Y0 () does exist, then
lim y0 (x) = lim Y0 (X) .

x0+

(5.20)

This is the so-called Prandtl matching condition. If (5.19) were satised, then
one would consider the leading-order outer expansion y0 (x) to match the
leading-order inner expansion Y0 (X) on an overlap domain
0 = {(x, ) : x = x I , i ()
()
o ()} .
D

(5.21)

At this stage it is worthwhile reecting a few points in detail. First, y0 (0+ )


or Y0 () may not exist at all. In this case inner and outer expansion cannot be matched to leading-order by use of the Prandtl matching condition.
However, it may still be possible to match the expansions by demonstrating
the existence of an overlap domain for which (5.19) is satised. Second, even
if matching condition (5.19) cannot be satised, there is still the possibility
that an order P of the outer expansion matches an order Q of the inner
expansion. That implies that there may be some overlap domain where

 P
Q
x 



n
n
=0.
(5.22)
yn (x )
Yn
lim

0+ ,x xed
n=0
n=0

5.1 A Formal Mathematical Approach Towards Matching

63

This allows me to dene matching:


x
Denition Choose x = ()
IR and let R be any non-negative integer. Consider outer and inner expansion dened in (5.1)- (5.3) to match
R (x ), if there are functions 1
to O(R ) on a common domain of validity D
and 2 with 1
2 and integers P, Q such that

lim

0+ ,x xed

lim

MP Q
=
R
 P

n=0

0+ ,x xed

x

Q
n yn (x ) n=0 n Yn
=0
R

(5.23)

for any function satisfying 1



2 and
R (x ) = {(x, ) : x = x, 1 ()
()
2 ()} .
D

(5.24)

5.1.1 Illustration of Extended Domains


To illustrate the formalism developed in the previous part of this section
here I consider the following simple example, where I will use the following
assumptions throughout the discussion: If 0 < (), x > 0
| log()|

e
n

n > 0

= Os (| log()|) e = Os (1)
x
| log()| ex/
n n > 0 .

(5.25)
(5.26)
(5.27)

Here = Os () denotes = O() and = O().


In particular I will consider matching of inner and outer expansion of the
function




1
x
x 
exp (1 + 1 4)
exp (1 1 4)
2
2
1 4
(5.28)
which is the solution of the singular initial value problem
y(x, ) =

y  + y  + y = 0 ,

(5.29)

1
.
(5.30)

Using (5.1) the rst two terms of the outer expansion can easily be determined
from (5.28). Fixing x and expanding in one nds


2
(5.31)
y = (1 + 2 + O(2 )) exx+O( ) ex/+x+O() ,
y(0, ) = 0 ,

y  (0, ) =

64

5. Asymptotic Analysis

from which I deduce


y0 (x) = ex

y1 (x) = (2 x)ex .

(5.32)

Similarly, to compute the inner expansion by the help of (5.3) one expresses
(5.28) in terms of X, xes X and expands in :


2
3
2
Y = (1 + 2 + O(2 )) eX X+O( ) eX+X+O( ) .
(5.33)
Thereby one obtains
Y0 (X) = 1 eX

Y1 (X) = (2 X) (2 + X)eX .

(5.34)

Before nding the overlap domains where outer and inner expansion
match to O(1) and O(), I will discuss how these expansions would arise
if one had not known the exact solution a priori.
By substituting the outer expansion into (5.29) one obtains the problems:
O(1) : y0 + y0 = 0
O() : y1 + y1 = y0 .

(5.35)
(5.36)

Assuming a0 and b0 to be constant the general solutions for (5.35) and (5.36)
are
y0 (x) = a0 ex
y1 (x) = (b0 a0 x)ex .

(5.37)
(5.38)

Clearly, a0 cannot be chosen such that y0 (x) satises both initial conditions.
Therefore, there must be a layer of extended validity at x = 0. In terms of Y
and X the initial value problem (5.29)(5.30) can be written as
Y  + Y  + Y = 0 ,
Y (0, ) = 0 ,

(5.39)

Y  (0, ) = 1 ,

(5.40)

from which one obtains the inner equations


O(1) : Y0 + Y0 = 0 ,
O() : Y1 + Y1 = Y0

Y0 (0) = 0 ,
,

Y0 (0) = 1

Y1 (0) = 0 ,

Y1 (0)

=0.

(5.41)
(5.42)

Their solutions are given in (5.34). In contrast to boundary value problems,


the unknown constants of integration to be determined from matching are
part of the outer solution. If one applied Prandtl matching to match y0 and
Y0 one would obtain
lim y0 (x) = a0 = 1 = lim Y0 (X)

x0+

and thus recover y0 (x) in (5.32).

(5.43)

5.1 A Formal Mathematical Approach Towards Matching

65

To nd an extended domain for the outer expansion one assumes ()


1
and searches for 1 () such that 1 ()
() implies (5.17) for the intermediate variable
x
(5.44)
x = > 0 .

Given (5.31), the limit is valid as long as ex /


1. To assure this, one can
choose 1 () = | log()|. Here the notation
= implies either
or
o is an extended domain for the outer expansion
= Os (). It follows that D
as long as satises
1,0 | log()|

= 1 .

(5.45)

Though for x I each denes a dierent region in the (x, )-plane, all
that really matters for the limit to vanish is that satises (5.45). So it is
common practice to consider the extended domain of a single term of the
outer expansion of y0 (x) to be given by (5.45).
To nd the extended domain for the rst-order outer expansion y0 (x) +
y1 (x) one assumes ()
1 and searches for an 1,1 () such that 1,1 ()

() implies
lim

0+ ,x

[y(x , ) y0 (x ) y1 (x )]
=0.

xed

(5.46)

Again from (5.31) one obtains that the above limit holds if satises (5.45),
i.e. the choice 1,1 = 1,0 works. If one continues this process of extending the
domain in an Rth order outer expansion to nd 1,R it is often the case that
1,R
1,R+1 . The reason is that adding further terms to the limit places
additional restrictions on . For this particular example the extended outer
domains at O(1) and O() turn out to be the same.
To nd an extended domain for the inner expansion to lowest order O(1)
one assumes
() and seeks an 2 () such that
2 () implies (5.18).
Again from (5.31) it is easy to verify that the extended domain for this inner
expansion is dened by

=
2,0 1 .
(5.47)
Finding the extended inner domain to O() is more delicate. In terms of
the intermediate variables one obtains
y(x , )

1 ex /

=
x x ex /


2
x /
+ O() .
+2 2e
+ O() + O

Moreover the intermediate variables yield the expression

(5.48)

66

5. Asymptotic Analysis

1 ex /

1
y(x , )

Y0 + Y1 =
=
x x ex / + 2 2ex / .

(5.49)
Subtracting these two expressions it becomes apparent that
lim

0+ ,x xed

[y(x , ) Y0 (x /) Y1 (x /)]
=0

(5.50)

provided 2 /
1. Thus the choice 2,1 = 1/2 ensures that the limit vanishes
and that the extended inner domain to O() reads

=
2,1 1/2 .

(5.51)

Here the extended domain to O() is smaller than the domain to O(1), i.e.
2,1
2,0 .
Based on the previous discussion it is obvious that the overlap domains
to O(1) and O() are, respectively:
1,0

2,0
1,1

2,1

(5.52)
(5.53)

or
| log()|

1
| log()|

1/2 .

(5.54)
(5.55)

Now three cases can be distinguished:


1. If satises these asymptotic relations, outer and inner expansion match
to O(1) and O(), respectively. In particular if satises (5.52) then
lim

0+ ,x xed

[y0 (x ) Y0 (x /)] = 0 .

(5.56)

2. If satises the more stringent requirement (5.53), then


[y0 (x ) + y1 (x ) Y0 (x /) Y1 (x /)]
=0.

xed
(5.57)
3. If the exact solution y was not known a priori then one would choose a0
in the incomplete outer solution y0 (x) = a0 ex and nd 1,0 , 2,0 such
that (5.56) is satised.
lim

0+ ,x

Remarks on Overlap Domains and Matching


1. General theorems showing the existence of overlap domains have not been
found [196]. In practice, the existence of overlap domains where inner and
outer solutions can be matched is done on a case by case basis.

5.1 A Formal Mathematical Approach Towards Matching

67

2. Prandtl matching corresponds to lowest-order matching with P = Q =


R = 0.
3. In some problems P and Q may not be known a priori. Moreover, P may
not equal Q.
4. Some expansions cannot be matched. The matching in (5.23) is dened
with respect to underlying gauge functions n () = n , n 0. Obviously,
some functions y may have more general outer expansions:

n ()yn (x) .
(5.58)
y(x, )
n0

Indeed, the inner variable could be dened in a more general way,


X = x/(), 0 <
1. Then the inner expansion may be expressed
with respect to dierent gauge functions. However, usually these sorts of
generalizations are not taken into account.
5.1.2 Matching in the Context of Diuse Interface Modeling
In the context of diuse interface modeling one employs an asymptotic analysis to establish the relationship between diuse interface and sharp interface
modeling equations. The basic idea underlying this asymptotic analysis is the
method of matched asymptotic expansion, whereby matching conditions are
employed to nd integration constants occurring in the inner expansion. In
this context an outer solution, valid away from the interface, is matched to an
inner solution, which is valid in the interfacial region [115]. The inner expansion near a point x0 on the interface is described via a stretched coordinate
, where n
is the local unit vector normal to the interface.
, with x = x0 + n
The low-order terms in an inner expansion of the solution can be obtained
by using the relations
2 =
and

1 2

+ O(1)
( ) +
2

vn
=
+ O(1) .
t

(5.59)

(5.60)

Here the dimensionless mean curvature


and the dimensionless interface
speed vn are assumed to be of order unity. In the inner region the phase-eld
and the transport eld are expanded in powers of ,
= 0 + 1 + O(2 ) ,

(5.61)

c = c0 + c1 + O(2 ) .

(5.62)

The resulting equations are solved order by order in , with far-eld boundary
conditions, which are obtained by matching to the outer solution. The outer

68

5. Asymptotic Analysis

solution of the phase-eld has (x) = 0 or (x) = 1 to all orders on either


side of the boundary ij . The transport eld has the expansion C(x) =
C(x)(0) + C (1) (x) + O(2 ). The limiting behavior near x0 is given by
(0)

(1)

C(x0 + n
) = C (x0 ) + [C (x0 ) +
(0)

(0)

C
(x0 )] + O(2 ) ,
n

(5.63)

(0)

where C+ and C denote the limits of C as 0+ and 0 , respectively, and so forth. The inner solution must match with this behavior
as [49, 50, 115]. In this way matching conditions yield the physics
of the interfacial region. Typically to dominant order the motion of the diffuse interface is governed by a Stefan problem. The next order approximation
brings in eects relating the temperature at the interface to its curvature and
normal velocity. Physically such eects arise from surface tension, which is
therefore an essential ingredient of diuse interface models. If corresponding
sharp interface models are well established those have to be recovered via
the asymptotic analysis. This might impose restrictions to the numerical parameters of the diuse interface model as will be discussed in the numerical
appendix of this book. On the other hand this analysis might also lead to
the reconsideration of sharp interface modeling equations (or even completely
new formulations of such models) as I will demonstrate in Chap. 6.

5.2 Asymptotic Matching


for Thin Film Epitaxial Growth
This section discusses an example [110], which displays how the above asymptotic analysis is employed to validate a diuse interface model for a free
boundary problem. The example chosen is that of thin lm epitaxial growth.
For this phenomenon a set of sharp interface equations constituting a well
established free boundary type model does exist. An essential ingredient of
this sharp interface model are terms to describe the attachment kinetics at
the interface. How to recover the precise kinetics via a diuse interface model
is not obvious at rst sight. This section discusses an approach, which derives
the diuse interface equations from two thermodynamic potentials, namely
the inner energy functional E and the entropy functional S. This way, when
writing down the free energy density function, an additional degree of freedom compared to models derived from one potential only is introduced. This
term serves to recover the precise attachment kinetics in a rst order analysis.
Thus this section provides essentially an example, in which matching is used
purely for the purpose of validation. The rst order analysis becomes necessary to recover the full physics of the sharp interface model, in particular
the kinetics involved. Since rst order analysis is based on a sharp interface
analysis, it is an example, which develops the matching analysis in the context of diuse interface modeling from its classical limit to the so-called thin

5.2 Asymptotic Matching for Thin Film Epitaxial Growth

69

interface limit. The latter has been introduced in the context of dendritic
growth by Karma and Rappel in [168].
5.2.1 Motivation
Growth of thin crystalline lms involves the physical transport of material to
a surface, where lm constituents are incorporated. Appropriate conditions
for such growth processes can be established by various kinds of epitaxy.
The ultimate goal of epitaxial growth is to control the structuring of a thin
lm on smaller and smaller scales. Modeling and simulating growth dynamics
and kinetics during epitaxy is a useful tool to predict the structures to be
expected under certain growth conditions. It contributes to an understanding of what physical mechanisms act on which scales and of what kind of
growth morphologies are triggered by their interaction. One of the challenges
to physicists and material scientists concerned with this task is the range of
length and time scales represented by problems of practical interest, e.g. the
growth of device layers [141]. Atomistic processes on the nanoscale as the
attachment of single atoms to a step or a defect can signicantly aect the
evolution of an overall morphology on a surfacesubsection several orders
of magnitude larger [21]. Moreover, depending on the epitaxial conditions,
the precise transport as well as the precise kinetic mechanisms might vary.
Growth modes might be as diverse as step ow growth, island growth or layerby-layer growth [47] with possible transitions from one to the other. A model
for epitaxial growth with potential for engineering applications would therefore (a) describe spatial and temporal scales of several orders in magnitude,
(b) provide for a straightforward extension to further transport elds and
kinetic quantities and (c) be capable of dealing with the topological changes
arising from the transition between growth modes.
Focusing just on the surface height prole of a lm above a surface, continuum equations in the form of partial dierential equations [235, 287, 288]
describe the evolution of surface morphologies at large scales. However, a
necessary assumption for their applicability is microscopic smoothness of the
surface, for the spatial derivatives to exist. Therefore they are unsuitable for
the description of roughening on the atomic scale, which is a big concern
in many device applications. On the other hand, particle based numerical
approaches as Kinetic Monte Carlo (KMC) algorithms can be employed to
model the atomistic processes in more detail and thereby constitute an alternative to continuum equations [73, 211, 216, 293]. A big advantage of these
models is their capability to quantify atomistic kinetic processes via attachment rates, which can be determined from rst principle calculations [242].
They have been applied widely to various issues of epitaxial growth in the
past [276] and constitute a comprehensive framework to describe homoepitaxial systems as well as very limited cases of heteroepitaxial growth [105].
However, they are not easily extended to include further growth mechanism
as, e.g., hydrodynamic transport in the bulk, elastic interaction between steps

70

5. Asymptotic Analysis

or entropic connement. Moreover, modeling systems of practical interest is


not feasible, since simulations are usually based on length and time scales of
single atoms and adatom hopping rates.
To meet challenges (a) to (c) formulated above and simulate surface
sectors of sizes of practical interest, continuum models allowing for fast computation on the one hand and for the incorporation of quantitative kinetic
information on the atomic scale on the other hand, constitute a promising modeling approach. Their common feature is the formulation of growth
problems as the one arising in epitaxy as a free-boundary problem either in
its sharp interface or its diuse interface formulation.
5.2.2 Sharp Interface Formulation
Mathematically the problem of epitaxial growth can be understood as the
solution of one (or several coupled1 ) transport equations for a variable in
two domains 1 and 2 , which are separated by a boundary ij dened by
(2.18):

+ u = k + f .
(5.64)
t
Equation (5.64) is the most general form of a diusionadvection equation, in
which k is a function of the vector of space coordinates x. u is the convective
speed, if hydrodynamic transport is considered in 1 and/or 2 . u may be
a function of . itself may be discontinous at the interface, i.e.:
[]ij =

(5.65)

and/or

] = .
(5.66)
n ij
The notation [g]ij = g + g denotes the dierence of the values taken by
the function g(x) in the limits x 0 and x 0, if the interface is assumed
to be at 0. f is a bodyforce term. The interface is subject to boundary
conditions of the general form:
[k

F (,

) = ij .
n

(5.67)

These boundary conditions ensure the conservation of mass and the conservation of energy at the interface during growth.
Equations (5.64)(5.67) can be adapted to constitute the two-sided sharp
interface model for epitaxial growth, i.e.:
1

For the sake of simplicity I will stick to the case of one transport eld in the
following. Extension to two or more is straightforward.

5.2 Asymptotic Matching for Thin Film Epitaxial Growth

= D2 ev + F
t

+
+ +
| = k ( eq )ij
D
n ij

| = k ( eq )ij
D
n ij
+

vn = D(

) .
n
n ij

71

(5.68)
(5.69)
(5.70)
(5.71)

Here vn is the normal velocity at the interface. The transport variable of


(5.64) is the density of adatoms above the surface denoted by . The dynamics of the adatoms on the surface is a diusionrelaxation process as described
by (5.68) and visualized in Fig. 5.1 with an external supply of atoms to the
surface quantied by F entering the equation, as well. Adatoms may evaporate after a time ev . Equations (5.69)(5.71) are boundary conditions of the
general form of (5.65) and (5.66), which ensure the conservation of energy,
(5.69) and (5.70), and the conservation of mass, (5.71), at the boundary ij .
D is the diusion constant involved in the process, the atomic area of the
atoms under consideration and k + and k are coecients to quantify the kinetic attachment to the boundary. They are dierent depending on whether
a boundary is reached from an upper (denoted by ) or lower side (denoted
by + ). This is due to the EhrlichSchwoebel eect [104], which describes an
energy barrier the adatom attaching from the upper side has to overcome.
Notice that this dierence results in two equations for the conservation of energy to be evaluated in + -direction respective -direction, namely (5.69)
and (5.70). The whole set of equations dates back to Burton, Cabrera and
Franck in 1956 [47]. Obviously a boundary can have very dierent kinds of
geometries depending on the epitaxial growth mode of the substrate. Pictured in Fig. 5.1 are boundaries during step ow and island growth. Note
that in the case of epitaxial growth anisotropy due to the underlying crystal
structure of the material is contained in eq = eq
0 (1 cos4). Detailed
studies of the inuence of crystalline anisotropy in epitaxial growth are given
in [106, 109].
The set of equations (5.68)(5.71) constitutes a modeling approach which
can be understood as a direct translation of the basic physical mechanisms of
epitaxy into mathematical equations. The physical picture consists of transport in the bulk, see (5.68), and attachment kinetics at the interface, see
(5.69)(5.71). Going back to the modeling challenges (a) to (c) formulated
in Sect. 5.2.1 it is important to notice that challenge (a), namely the description of scales of several orders in magnitude, can be achieved due to the
fact that these continuum equations themselves can be employed to simulate
structures of the order of 104
A, while at the same time information on the
atomic scale enters the equations via the attachment coecients k and k + .

72

5. Asymptotic Analysis
(a)
F
ev

(b)
F

ij
D

ev
+

ij

Fig. 5.1. Visualization of kinetics and dynamics during epitaxial growth: Adatoms
reach the surface with a rate that is proportional to F . They diuse on the surface
and may desorb from it according to the desorption parameter ev . Reaching a
boundary ij (e.g. a step (a) or an island (b)) they are incorporated with kinetic
attachment rates depending on their position with respect to the surface, i.e. from
up or down. This results in growth, i.e. movement of the boundary.

These can be determined on the basis of rst-principle calculations.2 Also


(b), namely the extension of the model to further transport elds and kinetic
mechanisms, can directly be addressed in this formulation: a second (diusive) transport eld would yield a further diusion equation to be coupled
to (5.68). In the same way hydrodynamic transport could be included in the
model via a coupling to the NavierStokes equations. Inserting other physical
mechanisms as diusion along steps or elastic step interaction into the model
will change the boundary conditions in two ways: On the one hand it results
in a change of the chemical potential at the steps and thereby in a dierent
expression for eq in (5.69) and (5.70). On the other hand it enters (5.71) in
gradient form to take into account the changed mass balance.
5.2.3 The Diuse Interface Model Equations
and Their Asymptotic Analysis
The basic idea underlying the phase-eld type of modeling epitaxial growth
is to replace boundary conditions (5.69)(5.71) of the sharp interface formulation by an additional dierential equation for the order parameter
introduced in Sect. 4.2. Here I assume to vary smoothly over the boundaries ij such that dierent layers of the substrate are labeled by dierent
integer numbers (see Fig. 5.2). 3 This convention has been employed to model
epitaxial growth by means of a phase-eld before by Karma and Plapp [170]
as well as by Liu and Metiu [207]. However, in their contributions equal
2
3

Obviously there is a lower limit to the structures to be resolved which stems


from the assumption of a continuum description.
The variation has to take place on a scale much less than the structuring of the
substrate.

5.2 Asymptotic Matching for Thin Film Epitaxial Growth

73

attachment kinetics at the interface had to be assumed. That way, the problem was directly mapped to that of dendritic growth with equal diusion
constants in both phases. No discontinuities causing fundamental modelingrelated questions arose. Here, on the other hand, I propose a diuse interface
model which handles the case of unequal attachment kinetics from both sides
of the interface. This model is thermodynamically consistent. It is derived
from conservation laws for the energy as well as the entropy. Thereby an additional degree of freedom is obtained. This is one important new feature of
the model necessary to recover the full two-sided growth problem described
by equations (5.68)(5.71) in the thin interface limit. The second essential
new ingredient for the thin interface analysis to work out correctly is the
which varies smoothly
introduction of a continuous attachment function ,
in the interfacial region, see (5.82) of the following detailed derivation of the
model.

interfacial region

0
x

Fig. 5.2. The phase-eld function as it varies over several layers of an epitaxial
substrate.

As discussed above the following derivation of model equations is based


on conservation laws for energy as well as entropy. These may be written in
the form

dE
+
JE n
dA = 0
(5.72)
dt
ij


dS
JS n
dA =
s P dV ,
(5.73)
+
dt
ij
V
where JE and JS represent the ux of energy and entropy through the bound . s P represents the entropy proary ij , which has an outward unit normal n
duction term. Following [110] for the inner energy functional E and the inner
entropy functional S an ansatz of the form

74

5. Asymptotic Analysis

(e(, ) + ||2 )dV


2
V

(s(, ) + ||2 )dV


S=
2
V

E =

(5.74)
(5.75)

is chosen. The term 2 ||2 denotes the contribution of the diuse interface,
where is a gradient coecient, which can be assumed constant. e and s are
the internal energy and entropy densities, respectively. Their relation to the
free energy density is given by
f
and

e(, ) = f (, ) + s(, ) .

s(, ) =

Thermodynamic consistency requires positive entropy production. Governing


equations, which guarantee positive local energy production are obtained by
choosing

t
k

JS = .

t
JE = k

(5.76)
(5.77)

with a non-classical ux
t associated with the interfacial region [289].
Entropy production is then given by
s P =

||2 +
[( + )2
].

(5.78)

The corresponding phase-eld equation is chosen such that the latter expres2
sion takes the form ||2 + (
t ) with > 0. Thus considering a free
energy of the form
f (, ) =

VE + VS

( eq )
k() ln( eq ) +
d() +
2
eq

(5.79)

the following set of governing equations can be derived for the case of epitaxial
growth:

f
= ( + )2
t

V
eq 
E + VS 
d ()
k () ,
= ( + )2
2
eq

= D2 + G()
+ F + ev .
t
t

(5.80)
(5.81)

Here 1 > 0 denotes a mobility coecient, d() = 2 (1 )2 a doublewell


potential function with a barrier height given by VE + VS . ( + ) > 0

5.2 Asymptotic Matching for Thin Film Epitaxial Growth

75

serves as gradient energy coecient and k() = 2 (3 2). The source term
G()
t describes the inuence of movement of the boundary ij on the
evolution of the transport eld , with G() = 2 d () 2 + 1. Within
G() the function in its suitably rescaled form is given by:
D 12X
D 12X
(1 ) + +
() .
=
k
l
k
l

(5.82)

The rescaling of the model equations simplies the asymptotic analysis by


which the sharp interface equations (5.68)(5.71) are to be recovered. It is
this asymptotic analysis, which serves as validation for the model at this
point. I will turn to it next.
Equation (5.80) and (5.81) can be made dimensionless by employing the
macroscopic length scale X , the time scale X
D and a scale for the governing
eld set to unity. Moreover, l denes the interface width, i.e. the scale on
which the phase eld varies from one integer value to another. lc takes the
role of a capillary length, which is proportional to surface tension , i.e.:
lc =

,
kB T

(5.83)

where kB refers to the Boltzmann constant and T to the temperature. With


abbreviations eq = + eq and V eq = VE + eq VS these terms allow me to
dene the following dimensionless parameters:
=

l
X

eq D
l
, =
6l lc
6 lc
Vs

= eq , = eq .
V

,
(5.84)

Based on those dimensionless parameters equations (5.80) and (5.81) can


be formulated as:
2

(1 + c) 

= 2 (1 + c)2
d () ck  () ,
2
t

(5.85)

c

2 c + G()
c ,
=D
(5.86)
t
t
2 d () + 1.

with the rescaled G()


= 2
2
To validate the above model equations one employs the method of matched
asymptotic expansion (see Sect. 5.1.2) with relations (5.59) and (5.60) for
the low-order terms in the inner expansion. To keep the succeeding analysis

2
clear for (
) the form is used, for
the form and t denotes t
.
Applying the method of matched asymptotic expansion to the dimensionless
model equations (5.85)(5.86) yields a leading-order phase-eld equation:
1
(0)
d ((0) ) = 0 .
2

(5.87)

76

5. Asymptotic Analysis

The leading-order equation for the transport eld reads:


c
(0) = 0 .
D

(5.88)

Integrating and matching the outer solution twice in succession gives that
(0)
(0)
c(0) = C+ (x0 ) = C (x0 ) is constant. The rst-order phase-eld equation
follows as:
1
(1)
(0)
vn +
) + c(0) k  ((0) ) ,
g  ((0) (1) ) = (
2

(5.89)

and a solvability condition for this equation is given by


c(0) = d0 +

vn
,
v0

(5.90)

where v0 corresponds to the velocity of a straight step. This can be identied


as a modied GibbsThomson boundary condition. With 0 it can be
understood as an equilibrium concentration corresponding to eq in equations
(5.69) and (5.70). Thus in this 0th order of the expansion the concentrations
right at the interface, denoted by + and in (5.69) and (5.70), would
simply be eq . The dierence to this equilibrium value is of kinetic origin and
retrieved in the next order of the expansion.
For the eld variable a rst-order equation reads:
(0)
c
(1) .
vn = D

(5.91)

By integrating this equation over the interval < < and using
the matching conditions, the leading-order massconservation condition is
recovered in the dimensionless form
(0)

(0)

C .

C+ D
vn = D
n
n

(5.92)

Thus the boundary condition (5.71) of the sharp interface formulation is


retrieved. It still remains to be proven that the boundary equations (5.69)
and (5.70) can be obtained in this order of the expansion, as well.
Following the procedure in [110], the leading- and rst-order phase-eld
equations are obtained as
1
(0)
d ((0) ) = 0 ,
2

(5.93)

1
(1)
d ((0) )(1) = r1 ,
2

(5.94)

where
(0)

r1 = c(1) +

(1)  (0)
(0)
c d ( ) ( vn +
) + c(1) k  ((0) ) .
2

(5.95)

5.2 Asymptotic Matching for Thin Film Epitaxial Growth

77

The rst two terms of r1 can be simplied using (5.93) to yield:


(1)  (0)

c d ( ) = c(1) g  ((0) ) ,
(5.96)
2
2
where is dened as + . The solvability condition for (5.94) can then be
written in the form
(0)

c(1) +


0=

+
2

(0)
r1 d

(0)

c(1) k  ((0) ) d

1
(0)
) ,
c(1) d ((0) ) d (
vn +
6

(5.97)

which involves the rst-order transport eld solution c(1) .


With c(1) = 0 the transport eld equation is satised identically to
leading-order. The rst-order transport eld equation reads

(0)
vn (0) (0) ,
c
(1)
vn d ((0) ) = D
(5.98)

2
which, when simplied by using (5.93), can be integrated to result in
(0)

vn (0)

(0)
(0)

c
(1) vn (0) 1 vn ((0) )2 = D

C+ vn .

C = D
D

2
n
n

(5.99)

Here 1 = 1 . The constant of integration has been evaluated in two


ways by taking and applying the matching conditions (5.63). The
solution can then be expressed in the equivalent forms

(1)

(1)

(1)

() = C +

() =

(1)
C (x)

(0)

C
(x) + +
n

(0)

C
+
n


(0)

vn

1 vn (0) 2
( )
+
2


d ,
(5.100)


(0)

vn (1

1 vn (0) 2
( )
)+
2


d .

(5.101)
The discontinuity of the adatom concentration eld across the boundary
ij now reads:
1 vn
.
(5.102)
12
Evaluation at the front and at the back of each boundary ij and inserting
into the full expansion for C up to the rst order results in:
(1)

(1)

C+ C =

78

5. Asymptotic Analysis

12X
C+
=
,

n
1 l

(5.103)

12X
C
,
=
n
1 l

(5.104)

(C+ Ceq )
and
(C Ceq )

respectively. Thus the equivalents of (5.69) and (5.70) are recovered. This
validates that the model equations (5.80) and (5.81) are an equivalent formulation of the full set of equations constituting the sharp interface formulation
(5.68)(5.71).

5.3 A Generalized Approach Towards the Asymptotic


Analysis of Diuse Interface Models
Whereas in the previous section the thin interface analysis arose as an extension of the sharp interface analysis, in this section I discuss a generalized approach towards matching, which directly incorporates a notion of the
diuse interface. It was proposed by Elder, Grant, Provatas and Kosterlitz
in [102] and works for a specic class of moving boundary problems only.
These include order/disorder transitions, spinodal decomposition, Ostwald
ripening, dendritic growth and eutectic alloy solidication. For this class of
phenomena the projection operator method can be employed to extract the
sharp interface limit from diuse interface models with interfaces that are
diuse on a length scale . In particular, phase-eld equations are mapped
onto sharp interface equations in the limits
1 and v/D
1, where
and v are the interface curvature and velocity, respectively. D is the diusion
constant in the bulk. The approach follows the projection operator method
of Kawasaki and Ohta [173, 174, 175, 176, 225]. It is generally applicable in
the sense that the free energy functional F does not need to be specied.
This implies that the free energy density can be chosen freely, such that it
guarantees strict relaxational behavior of the thermodynamic potential and
thus ensures thermodynamic consistency inherently.
5.3.1 Role of the Governing Thermodynamic Potential
Following [102] I will consider two elds, a non-conserved phase-eld and
a conserved eld c. The phase-eld distinguishes between the two phases,
e.g. liquid phase and solid phase. The c eld can be understood as a concentration. The free energy functional describing the system can be written
as



1
1
2
2
F{, c} = dr K || + Kc |c| + f (, c) ,
(5.105)
2
2

5.3 A Generalized Approach Towards the Asymptotic Analysis

79

where f (, c) is the local bulk free energy density. The gradient terms are
related to the phase-separating interface itself. The dynamics of these elds
are described by an equation of motion for the non-conserved variable ,

f 
F

,
=
= K 2 +

(5.106)

as well as an equation of motion for the conserved variable c, i.e. the concentration:


c
f
2 F
2
2
.
(5.107)
= c
= c Kc c +
c
c
t
Here F{, c}/c = refers to the chemical potential. c and are the
respective mobilities (compare to (4.42) of Sect. 4.1).
A generic bulk free energy density f (c, ) can give rise to a variety of
equilibrium as well as a variety of metastable states. To illustrate this, I will
consider the following bulk free energy
T
[c ln c + (1 c) ln (1 c)]
f (, c) = u
T

M
1
1
1 2
+ T (c ) () 2 + 4 .
2
2
4

(5.108)

Here () 2 43 /3 + 25 /5, T (T TM )/TM , where TM denotes


the melting temperature. The parameters , and u are phenomenological
parameters and have to be determined by calibration to experimental phase
diagrams. If these parameters are chosen as = = 1.0 and u = 0.6, a phase
diagram as displayed in Fig. 5.3 emerges. A dierent type of phase diagram
arising from parameters = = 1.0 and u = 0.45 is pictured in Fig. 5.4.
Obviously the choice of parameters determines dierently shaped domains
of pure phases (i.e. liquid phase L and solid phase S) as well as coexistence
regimes S + L (solid/liquid) and S + S (solid/solid) (compare to the phase
diagram explained in detail in Sect. 3.1, depicted by Fig. 3.1). Within the
current section the focus is on the dynamics of phase-separating interfaces.
The dynamics of a system can result in transitions from one point of the
phase diagram to another. Dierent such transitions are shown in Fig. 5.3 and
Fig. 5.4. Each of them denotes a dierent kind of growth phenomenon. They
are termed quenches. More precisely a quench is dened as a rapid change
in temperature taking a system from one domain within the phase diagram
to another. It is often considered instantaneous in theoretical modeling. The
generalized calculations of the following subsection are carried out including
all the possible quenches illustrated in Fig. 5.3 and Fig. 5.4. As stated above,
this does not require to specify the explicit form of the bulk free energy
density f . It is simply assumed that f is chosen in a way such that all the
phases of interest are well dened.

80

5. Asymptotic Analysis

Dendritic
Growth

L
(T-TM)/TM

0.2

L+S

L+S
Spinodal Decomposition

Ostwald
Ripening

S+S

-0.2

0.2

0.4

0.8

0.6

Fig. 5.3. Mean-eld phase diagram obtained from a bulk free energy given by
(5.108) with parameters = = 1.0 and u = 0.6. In this gure the regions
containing vertical and horizontal lines are liquid/solid (L + S) and solid/solid
(S +S) coexistence regions, respectively. The gure follows, just as further gures in
this section, [102]. Here liquid/solid and solid/solid coexistence regimes are disjunct.
Decreasing the parameter u within the energy density f lowers the critical point Tc
of possible solid/solid coexistence, so that for smaller values of u the two coexistence
regimes will merge. Such a situation is depicted in Fig. 5.4. A phase diagram with
merging coexistence regimes gives rise to a qualitatively dierent quench (see text),
which refers to the growth phenomenon of eutectic solidication.

Dendritic
Growth

L
Eutectic Solidification

(T-TM)/TM

0.2

L+S

L+S
S
S+S

-0.2

0.2

0.4

0.6

0.8

Fig. 5.4. Same as the previous gure apart from dierent parameters for the bulk
free energy given by = = 1.0 and u = 0.45.

5.3 A Generalized Approach Towards the Asymptotic Analysis

81

5.3.2 The Expansion Procedure in Detail


The goal of this section is to derive the sharp interface formulation for the
dynamics of a system described by a free energy functional F as given in
(5.105) and governed by (5.106) and (5.107) via the generalized asymptotic
analysis. It proceeds by expanding around a planar equilibrium interface.
The interface is out of equilibrium as soon as it is either gently curved or
one of the two bulk phases is metastable. In the former case, a curvature
is assumed gentle, if the radius of curvature 1/ is large compared to the
interface width or correlation length. This yields one small dimensionless
expansion parameter . In the latter case, the dierence between the free
energy of the stable and metastable phases causes the interface to propagate
into the metastable one. If the free energy dierence is small the propagation
velocity v is small, as well. In this context a velocity can be interpreted to be
small, if the interface moves so slowly that a steady state diusion eld is
allowed to form in front of the interface. This implies that the time for the
diusion eld to relax during interface motion of a distance should be much
smaller than the time /v the interface itself requires to move that distance.
Since the diusion time is given by = 2 /D, this leads to a second small
dimensionless parameter v/D. The latter is known as the interfacial Peclet
number. In the following analysis the interface equations will be obtained to
lowest order in both small parameters. Technically the expansion to lowest
order in these two parameters can be achieved by regarding them to be of the
same order in the expansion. In the calculations to follow both parameters
will be considered O() with
1.
The succeeding expansion exploits the fact that the elds behave qualitatively dierent close and far from the interface. In the region close to
the interface, they vary rapidly over distances O(), while far from the interface they vary on a scale O(/). If a length scale exists such that
1
/
1/, distinct inner and outer regions can be dened. Those
are depicted in Fig. 5.5. For the analysis to follow it is appropriate to expand
in both inner and outer regions and match the solutions order by order at
the length scale . In this sense the expansion follows the ideas described
previously in Sect. 5.1.2. It diers from the example provided to illustrate
the usual way to carry out matching in the thin interface limit with fully
specied potentials and respective thin interface analysis model equations in
Sect. 5.2 in three points:
1. The generalized analysis is carried out to lowest order in two dimensionless parameters, i.e. and the interfacial Peclet number P eInt = v/D.
Technically this is possible by considering them both of O() (see above).
2. The expansion involves usage of the Greens function formalism and further the projection operator method as discussed in the introductory part
of this section. This allows us to remain on a general level.

82

5. Asymptotic Analysis

3. An additional solvability condition is obtained by employing the denition of the Gibbs surface. It is required to carry out the matching again
remaining on a general level.
Following [102] all of the above is featured in the succeeding paragraphs.
Interface
S+

S-

Inner
Region

Inner
Region
Outer

Outer
V+

V-

c(u)

Fig. 5.5. Illustration of inner and outer regions dened in the context of the
expansion procedure described within this section.

Inner Expansion I start by discussing the inner expansion, which applies


to a region dened by < u < . Here u is a coordinate normal to the
interface and 1
/
1/. The aim is to obtain asymptotic expansions
for the solutions of the evolution equations (5.106) and (5.107) valid in this
inner region. The latter can be written in a compact form as
1 c
= 2
c t
where

(5.109)

f
(5.110)
eq .
c
The rst step is to partition the system into two regions V+ and V
bounded by surfaces S+ and S , respectively. The region V+ is dened by
0 < u < , V by < u < 0. The position of the interface between two
= (r) eq = Kc 2 c +

5.3 A Generalized Approach Towards the Asymptotic Analysis

83

bulk regions is dened as u(r, t) = 0. This denition is a formal one. To get


an impression one may picture the surface u(r, t) = 0 as the surface near
which the elds c and vary rapidly over a distance O(). However the exact
position remains unspecied at this point. Moreover it is useful to dene
Greens functions G (r, r  ) in the regions < u < 0 (G ) and 0 < u <
(G+ ), respectively, obeying
2 G (r, r  ) = (r r  ) .

(5.111)

These Greens functions satisfy the following boundary conditions:

G(r, r ) = 0 at u = 0 and u = 0,
G(r, r )/u = 0 at u = ,
G(r, r )/u = 0 at u = ,
periodic at u = and u = .

Note that both r and r  lie in the same region V+ or V , respectively.


Multiplying (5.109) by G and integrating over r  V yields

G (r, r  ) c
(r) =
dr 
c
t
V



 
+
(5.112)
dS G G   ,
S


where  = (r  ) is dened as in (5.110). V+ dr  denotes integration over

V+ (i.e. 0 < u(r  ) < ). Moreover, V dr  refers to integration over V .

In analogy, S+ dS denotes integration over the boundary S+ enclosing V+ .

Similarly S dS denotes integration over the boundary S enclosing V .
Multiplying (5.112) by cin
0 /u and integrating over < u < yields


cin
0




cin f
Kc 2 c eq ,
=
du 0
u
c

B+S

du

where B = B + + B and S = S + + S with




1
cin
c
0

B =
,
du
dr  G (r, r  )
c 0
u V
t
S =


0

du

cin
0
u


S


dS   G G   .

(5.113)

(5.114)
(5.115)

An analogous formula for can be derived by multiplying (5.106) by in


0 /u
and integrating over < u < + to obtain

84

5. Asymptotic Analysis

du

in
0
=+
u t

du

in
0
u

K 2


.

(5.116)

Each term in the above equations can systematically be expanded in powers


of . In this section, attention is restricted to the terms O(), which will prove
to be sucient to obtain the relevant sharp interface conditions.
To facilitate the expansion, c(r, t), (r, t) and the chemical potential
(r, t) are expressed by the help of a power series in as follows:
in
2 in
c(r, t) = cin
0 (u(r)) + c1 + c2 +
in
in
(r, t) = 0 (u(r)) + 1 + 2 in
2 +
in
2
in
+

(r, t) = in
(u(r))
+

0
1
2 + ,

(5.117)

where the superscript in refers to the inner solution. To expand the Laplacian in powers of , it is useful to introduce a curvilinear coordinate system
with one coordinate u along the local normal to the interface and (d 1) coordinates s perpendicular to u and tangential to the interface. For simplicity
a two-dimensional system is considered, where s is the scalar arc length. Note
in
that within (5.117) the O(0 ) terms cin
0 and 0 are the equilibrium planar
interface solutions.
At this stage I come back to the point, that the exact position of the
interface has not yet been specied. Its choice is somewhat exible to within a
distance . It is one of the key elements of this generalized analysis to choose
it in accordance to the denition of the Gibbs surface, so that the excess
surface concentration is equal on both sides of the interface. This results in
an additional solvability condition, which ensures that the chemical potential
is continuous across the interface.
The transformation from Cartesian to curvilinear coordinates (see Appendix A) results in the formal expansion
2 2 = L0 + BL1 + B2 L2 + ,

(5.118)

where the specic form of Ln depends on the expansion. In the inner region,
derivatives of the elds with respect to u are much larger than derivatives
with respect to s, which are zero, if the curvature and the interfacial Peclet
number vanish. This is taken into account by introducing the dimensionless
variables u
and s. These are of O(0 ) and given by u = u
and s =
s/. As
proven in Appendix A, this scaling results in:
2
L0 = 2 / u
L1 =
/ u

L2 = 2 /
s2
2 u
/ u
,
where the dimensionless curvature
/ is of order unity.
A nal step is to expand the time derivatives within (5.114) and (5.116)
in B. For these calculations, it is convenient to choose a frame of reference, in
which the interface position is stationary, such that

5.3 A Generalized Approach Towards the Asymptotic Analysis



v.
 = 
t (u,s)
t r

85

(5.119)

Here v is the interface velocity which has components normal and tangential
to the interface. The time derivative on the right hand side corresponds to a
relaxational dynamics independent of interfacial motion. If this operator acts
on the elds c and , the tangential component and time derivative are of
order 3 . Thus they can be dropped4 . To O(), /t|r becomes:
v1

(5.120)
+ O()2 + ,
 =
t r
u

where the normal velocity has been expanded in a power series in , i.e.:

 m
u

vn
vm .
m=1
t

(5.121)

in
Using these expansions and further expanding f around cin
0 and 0 , the
right hand sides of (5.113) and (5.116) become

and


 

cin
cin  
0
c 2 2 c f =
c L0 cin f (1,0)
0 + K
0 + K
d
u 0
0
i
u
c
u




c (L1 cin + L0 cin ) cin f (2,0) in f (1,1) + O(B2 )
+B K
(5.122)
1 i
1
1 i
0

du



in
f
0
2 2

K
du




in 

L0 in f (0,1) + B K
(L1 in
K
=
d
u 0
0
0
i



(1,1)
in (0,2)
+ O(B2 ) .
cin
+L0 in
1 ) 1 fi
1 fi

(5.123)

(n,m)
c Kc / 2 , K
K / 2 and =
in , K
n+m f /cn m |in
Here fi
0 ,c0
0
/. Terms of O(B ) vanish by construction. For later use, it is convenient to
perform partial integrations on combinations of terms:
4

In two dimensions the tangential component of the time derivative can be expressed as vT /s, where vT accounts for ow of concentration (or ) along the
interface. vT must be O() since it is zero for = 0. Furthermore in the inner
region s is scaled with such that vT /s O(2 ). If this term operates on c or
, it becomes O(3 ), since both c and are independent of s. The time derivative
in the moving frame /t|(u,s) is at least of order O(3 ). This term accounts for
the relaxation of uctuations around the steady state prole. For the concentration eld uctuations relax as t (length)2 , which implies /t|(u,s) O(2 ).
Just as the tangential component this term becomes O(3 ) when operating on
c. Moreover, uctuations relax exponentially fast, thus rendering the term
negligible.

86

5. Asymptotic Analysis



2
cin
0
c f (2,0) cin =
K
d u

1
i
u
2
u


in
(1,1) 0
d
u cin
=
1 fi
u

d
u cin
1

2
c f (2,0)
K
i
u
2

cin
0
u

(5.124)

and



 in



0
2
2
in
(0,2)
(0,2)
0
in
in

K 2 fi
1 =
d
u 1 K 2 fi
d
u


in
(1,1) c0
d
u in
=
,
(5.125)
1 fi
u

in
= in the limit = / >> 1.
since derivatives of cin
0 and 0 vanish at u
Here the nal equations (5.124) and (5.125) were obtained using the relationships


f (c, )
2 f (c, ) c
c

2 f (c, )
=
+ f (1,1)
= f (2,0)
+
u

c u

c
c2 u

and

f (c, )


=

2 f (c, ) 2 f (c, ) c

c
+
= f (0,2)
+ f (1,1)
.
2

c u

To complete the calculation, the left-hand sides of (5.113) and (5.116) are
expanded to lowest order in B. The expansion for in (5.116) is straightforward


1
B
in
in in
=
du 0
d
u 0 v1 0

u t

u

v1
=
+ O(2 ) ,
(5.126)
K
where


K


du

in
0
u

2
.

(5.127)

The equivalent expansion for c is more complicated. The algebra is given


in Appendix B. Formally the results of these calculations can be expressed
as
B + S = B(B1 / + S1 ) + B2 (B2 / + S2 ) + ,

(5.128)

where Bn and Sn are given in Appendix B.


Combining all of the above yields the following two equations to O(B):

5.3 A Generalized Approach Towards the Asymptotic Analysis

v1


A1
=
K

87

(5.129)

and


c
v1 (s) 2
u) cin
u)]2
d
u[cout
+ A1
0 (
0 (
c

  0
in
1 

d
u[cout
u) cin
u)]

0 (
0 (
u

 
in
1 

d
u[cout
u) cin
u)] ,
(5.130)

0 (
0 (
u
0

c in
1 (0, s) =

in

where c cin
0 () c0 () is the miscibility gap, and
 in

u
<0
c0 ()
out
c0 (
,
u)

(+
)
u
>0
cin
0


c Kc


du

cin
0
u

2
,



in
cin
(1,1)
0
in 0

c
fi
d
u in
.
1
1


A1 =

(5.131)

(5.132)
(5.133)

Equation (5.130) yields the chemical potential of the inner solution at


the interface (i.e., at u
= 0). This must be matched to the outer solution at
An expression for in ()
can be obtained by expanding (5.109)
u
= .
1
to O(B):
cin
c
v1 0 = 2 L0 in
(5.134)
1 .
u

then from 0 to - yields


Integrating (5.134) twice - rst from u
to ,
in 
2 
s) = in (0, s) + 1  + v1
in
(
,
d
u[cout
u) cin
u)] . (5.135)
1
1
0 (
0 (
u
c 0

A similar procedure has to be applied to in


1 (). From (5.130) and (5.135)
one obtains

cin
1 (, s) =



in
v1 2
1 

d
u[cout
u) cin
u)]2
c /B + A1 c

0 (
0 (
u
c
  0
 
in
in
1 
1 
out
in
d
u[c0 (
u) c0 (
u)]
d
u[cout
u) cin
u)]



0 (
0 (
u

u
0

v1 2
+c
d
u[cout
u) cin
u)] .
(5.136)
0 (
0 (
c 0

88

5. Asymptotic Analysis

The integrals in (5.136) can be rewritten in a more suitable form by noting


Thus
u| .
u) = cin
that  1 in the inner region, so that cin
0 () for |
0 (
(5.136) results in
) =
cin
1 (, s


c

s) 2
in
v1 (
1 

d
u[cin
u) cout
u)]2
+ A1 c

0 (
0 (

u

c

  0
 
in
in
1 
1 
in
d
u[cout
(
u
)

c
(
u
)]

d
u[cout
u) cin
u)]



0
0
0 (
0 (
u

u
0

s) 2
v1 (
d
u[cout
u) cin
u)] .
(5.137)
+c
0 (
0 (
c
0
At this point the expression for v1 is still lacking. It is obtained by integrating
(5.110) over < u
<
v1 =

 

c  in
in
1 
1 
.



u

c 2
u

(5.138)

This solution for in


1 () has to be matched to the solution of the expansion
within the outer region, to which I will turn next.
Outer Expansion Far from the interface, the elds and c vary slowly in
space and are close to the bulk equilibrium values eq and ceq . Variations
of the elds in the bulk regions far from the interface take place on length
scales O(/B) in all directions. Therefore suitable dimensionless space and
time coordinates are (
u, s, t) (Bu/, Bs/, B2 t/ ).
Expanding (r) around the bulk equilibrium solution (r) = out (r) +
eq yields

out
f 
2 f 
= K 2 out

 out
t
eq 2 eq 
1 3 f 
1 4 f 
out 2

(
)

(5.139)

 (out )3 .
0
2! 3 eq
3! 4 eq
By denition, (f /)eq = 0. Since out = 0 in the limit B 0, (5.139)
is linear at O(B). Furthermore, ( 2 f /2 )eq > 0, so that out
vanishes ex0
ponentially with time for all wavelengths. Thus, out
is
trivial
in the outer
0
region and can be ignored. It is convenient to expand cout and out in the
outer region as
out
cout (r) = cout
0 + Bc1 +
out
out
(r) = 0 + Bout
1 + ,

(5.140)

is given by (5.131). At O(B3 ) in dimensionless variables:


where cout
0
cout
c 2 out
1
= 2
1 ,

(5.141)

5.3 A Generalized Approach Towards the Asymptotic Analysis

89

(/B) is the scaled dimensionless derivative suitable for the outer


where
region. Equation (5.140) is evaluated in a laboratory frame. It simplies to
a linear diusion equation for the chemical potential inside the bulk phases.
In dimensional units it reads
out
= Dc 2 out ,
t

(5.142)

where Dc c (out /cout )eq is a diusion constant. The value of Dc depends on the bulk equilibrium phase considered.
The Gibbs Surface Solvability Condition To solve (5.142), initial
values out
u = 0, s) are required. These result from matching to the inner
1 (
To obtain out (
s)
solution in
u, s) at u
= .
u = 0 , s) from out
u = ,
1 (
1
1 (

with B, it is useful to Taylor expand around u


=
) + (

u )
u, s) = out
out
1 (
1 (, s

out
1
| + .
u

In the outer region,


1. This expansion is valid at u
=0
out 
) = out (0, s) 1  + .
out
1
1 (, s
u

(5.143)

(5.144)

in

Since out
1 () = 1 (), one can use (5.137) and (5.144) to obtain


s) 2
c
v1 (
+
A
(0,
s

)
=

d
u[cin
u) cout
u)]2
cout
1
0 (
0 (
1

  0
in

1

d
u[cout
u) cin
u)]

0 (
0 (
u

 
in
1 
d
u[cout
u) cin
u)]


0 (
0 (
u
0

s) 2
v1 (
d
u[cout
u) cin
u)] .
(5.145)
+c
0 (
0 (
c
0

This yields the appropriate boundary value for out


). The inner solution
1 (0, s
out
(0),
since
the
matching is done
in
(0)
diers
from
the
outer
solution

1
1
at u
= and extrapolated linearly to u
= 0 by (5.144). This extrapolation
out
of in
is illustrated in Fig. 5.6. It is essential that out
and in
1 to 1
1
1

coincide at u
= , which is in analogy to taking the thin interface limit
rather than the sharp interface limit discussed in the context of the example
of the previous section. The right-hand side of (5.145) appears to depend on
whether the inner and outer solutions are matched at u = or u = . This
ambiguity is eliminated by dening the interface to be a Gibbs surface at
u = 0, given by the condition

in
du[cout
(5.146)
0 (u) c0 (u)] = 0 .

90

5. Asymptotic Analysis

out
1 (0)
in
1 (0)

out
(u)
1
in
(u)
1

Fig. 5.6. Matching in the context of the generalized asymtotic analysis applied to
out
in
1 (u) (dashed line) with 1 (u) (solid line) at u = .

It is always possible to satisfy (5.146) by choosing the position of the interface


(i.e. u = 0) accordingly. In essence, the interface position is determined by
the condition that the excess concentration on one side of the interface is
exactly compensated by the decit on the other, as depicted in Fig. 5.7. This
can be interpreted as a solvability condition. It is one of the relations, which
is required to obtain sharp interface model equations via this generalized
asymptotic expansion.
A further necessary relation is obtained by matching the rst derivative
of the chemical potential. In dimensional units it reads:
vc = c

 in 
in  

.


u
u

(5.147)

Matching derivatives of the inner and outer solutions for and extrapolating
to u = 0 by (5.143), yields the standard result
vc = c

 out 
out  

,
+

u 0
u 0

(5.148)

since out (u) is linear for 0 < |u| . Finally, combining (5.145), (5.146) and
(5.148), results in the chemical potential of a moving, curved interface
c[out (0, s) eq ] = c + E 2 v + A1 + O(B2 ) ,

(5.149)

5.3 A Generalized Approach Towards the Asymptotic Analysis

out

111111
000000
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111
000000
111111

91

in

du[c0 (u)-c 0 (u)]

in

out

du[c0 (u)-c 0 (u)]

Fig. 5.7. Gibbs surface as dened by (5.146) pictured for u = 0. It matches the
concentration decit on one side with the concentration excess on the other.

where A1 is given in (5.133) and





1
2
in
2
du [cout
(u)]

[c
(u)]
.
E2
0
0
c

(5.150)

All of the above can be combined into a single set of boundary layer equations
written in terms of the concentration. In the outer region these boundary
layer equations are related at order O() to the chemical potential by the
relationship


c .
(5.151)
=
c eq
Combining this result with (5.149) and (5.129) yields the GibbsThomson
relation. In dimensionless units it reads
c(0, s)
= do (s) kin v ,
c

(5.152)

where do is the capillary length given by


do =

,
(c)2 (/c)eq

c + the total surface tension given by

(5.153)

92

5. Asymptotic Analysis


=

du K

in
0
u

2


+ Kc

cin
0
u

2 
,

and kin is the coecient of kinetic undercooling, which reads





1
2
.
kin =

E
(c)2 (/c)eq ) K

(5.154)

(5.155)

Equation (5.152) provides a boundary condition at the interface for the diffusion equation (5.142), which can be written

where

c
= Dc 2 c ,
t

(5.156)



Dc c
.
c eq

(5.157)

This must be solved in conjunction with (5.148), which can be transformed


into




c
c
c v(s) = Dc
.
(5.158)
Dc
u 0+
u 0
The latter represents the well known Stefan condition, ensuring conservation
of energy at the interface. Thus (5.152)(5.158) recover the sharp interface
formulation of this general class of moving interface problems represented
by an energy functional F as given in (5.105). Equations (5.153)(5.155)
relate the physical parameters of the sharp interface description to the phenomenological parameters of the diuse interface model equations. They are
determined by K , Kc and by the exact position of the interface u = 0, as
expressed by E 2 . However, there remains a freedom to choose f in such a way
that thermodynamic consistency can be guaranteed.

5.4 Discussion
Here I will briey summarize the impact of asymptotic matching approaches
for the overall context of diuse interface modeling. Such methods can serve
to establish the correspondence of diuse interface models and sharp interface
models. In this sense an asymptotic analysis does not necessarily involve a discussion of thermodynamic consistency. Instead the latter model is validated
completely by its equivalence to the sharp interface model in the respective
limits (either the sharp or the thin interface limit). However, such a
procedure misses the chance to apply diuse interface modeling to growth
phenomena for which sharp interface formulations are not yet established.
In this case the thermodynamic theory underlying diuse interface models
opens the possibility to proceed in the opposite direction, i.e. formulate the

5.5 Appendix

93

diuse interface model on the basis of a variation of the thermodynamic


potential(s) rst and then derive the sharp interface model from the latter
through the asymptotic analysis by at the same time taking into account
additional conditions to ensure strict relaxation of the potentials.
In this sense the great advance of the generalized asymptotic analysis
as described in Sect. 5.3 is to ensure thermodynamic consistency inherently
through general algebraic operations, which are independent of the specic
form of the free energy functional5 . Thus the calculations are universal in
the sense that a freedom to choose F remains. Exploiting the concept of the
Gibbs surface explained in detail in Sect. 5.3.2 yields a solvability condition
to carry out the matching within this generalized framework.
Consequently this generalized asymptotic analysis opens the way to construct physically consistent sharp interface descriptions of more complicated
multiple phase systems such as a solid in contact with a uid which can
support ows. This will involve mode coupling terms in the dynamical equations. Once such governing equations are constructed, there is no longer any
conceptual diculty in deriving the corresponding sharp interface equations.
In Chap. 6 I will present a detailed discussion of growth inuenced by diffusive as well as hydrodynamic transport in the bulk phase following these
ideas of exploiting the diuse interface approach to establish the precise sharp
interface boundary conditions of the system.

5.5 Appendix
Appendix A: Transformation to Curvilinear Coordinates
The curvilinear coordinates (u, s) used in the text are related to the Cartesian
coordinates as [103]
r = R(s) + u
n(s) ,
(5.159)
(s) is the normal vector. The
where R is the position of the interface and n
metric tensor g of the transformation from Cartesian to curvilinear coordinates is given by


1
0
,
g=
0 (1 + u)2
where = /s is the curvature with being the angle between the r1 axis and the tangent to the curve. The Laplacian in (u, s) is obtained in the
following manner
 1



2 =
|g|g

r
|g| r
,

=
5

us
1
+

+
,
2
2
u
1 + u u (1 + u) s2
(1 + u)3 s

This holds provided that F describes well dened phases.

(5.160)

94

5. Asymptotic Analysis

where r1 = u and r2 = s. (r1 and r2 denote the components of r.) g are


the components of the inverse of the matrix g and s /s.
Within the inner region, the elds vary more rapidly in u direction than
in s direction. This allows us to rescale coordinates (u, s) to dimensionless
variables as (
u, s) (u/, s/). The dimensionless curvature is expressed by

= / and
s = 2 s /2 , respectively. In terms of these dimensionless
variables the Laplacian reads
2

2
B

B3 u

B2
+

+
u
2
1 + B
u
u
(1 + B
u
)2
s2
(1 + B
u
)3
s




2

2
=
(n + 1)(B
u
)n
+
B
(B
u
)n
+ B2
2
u

s2
u

n=0
n=0



s 3 
n
B

(n + 1)(n + 2)(B
u
)

s
2
n=0

2 2 =

= L0 + BL1 + B2 L2 + B3 L3 + ,

(5.161)

where
L0 = 2 / u
2
L1 =
/ u

(5.162)
(5.163)
2

L2 = /
s
u
/ u

L3 = 2
u
2 /
s2 +
3 u
2 / u
u

s/
s.

(5.164)
(5.165)

Appendix B: Greens Functions


In addition an expansion for the inverse of the Laplacian, i.e. the Greens
function, is required [103]. The Greens function of interest is dened by
2r G(r, r  ) = (r r  ) .

(5.166)

An expansion of this function can be derived in the following manner. Let


G(r, r  ) = G0 (
u, s; u
 , s ) + G1 (
u, s; u
 , s ) + , where
L0 G0 = 0
L0 G1 + L1 G0 = (
uu
 )(
s s )
L0 G2 + L1 G1 + L2 G0 = 0

(5.167)
(5.168)
(5.169)

... .
The solution for G0 is 0. Thus to lowest order the relevant solution for G is
G1 . G1 satises the equation
2
G1 (
u, s; u
 , s ) = (
uu
 )(
s s ) ,
u
2

(5.170)

5.5 Appendix

with solutions
G
u, s; u
 , s ) =
1 (

G+
u, s; u
 , s )
1 (


=

95

 < u
<0
u
(
s s ) < u


u
(
s s ) < u
<u
 < 0 ,

s s )

u (

u(
s s )

0<u
 < u
<
0<u
<u
 < .

The surface terms S of (5.115) can be expanded as





u , s ) 
G
(
u , s )

G
d
s (

u

u

B
0
B



in 


u, s; u
 , s )
G (
c
u , s ) m
d
u 0
d
s
Bn+m1 n (
=
u

u
B
0
n=1 m=0
u , s ) 
n (
G
u, s; u
 , s )
(5.171)
 = BS1 + B2 S2 + ,
m (
u

B

S =

d
u

cin
0
u

where
S1


=

S2 =

cin
d
u 0
u


B

d
s

G
1
1
1
G
1
u

u


G
1
d
s 2
u


cin
0
+
u
B
0

1
2 
G

G
 .
2
1
u

u
 B
d
u





G
2
1
u

(5.172)

Here the subscript B indicates that the integrands are evaluated on the

 = .
boundary at u
 = 0 and at u
Moreover the O() surface contribution becomes:
S1 = S1 + S1+


 0
u , s) 
cin
in
1 (
=
d
u 0 in
(0,
s

)
+
u


1
u

u




 +

in
u , s) 
cin
1 (
in

+
d
u 0 u
+

(o,
s

)

1
u

u

+
0


0
in
1 
in

= in
) +
d
u[cin
u)]

1 (0, s
0 () c0 (

u

 
in
1 
in

+
d
u[cin
u)] .

0 (+) c0 (
u
 + 0

(5.173)

To evaluate the bulk contribution B, (5.114) is expanded in powers of .


Then (5.114) reads

96

5. Asymptotic Analysis

B =

1
c

du

cin
0
u

dr  G (r, r  )

cin (r  , t)
.
t

(5.174)

Further a variable transformation r  u


 , s yields
B =

2
c


d
u


d
u d
s
Bn+m1


1+u
B
(
s ) n=1 m=1

cin
0
(cin + Bcin
1 + )
u
u
 0
= BB1 + B2 B2 + ,
vn G
u, s; u
 , s )
m (

(5.175)

where
B1 =
B2 =

2
c
2


d
u

d
u d
s v1 G
1


d
u

in
cin
0 c0

u
u

d
u d
s (v2 G
1 + v1 G2

in
in
in 
cin
0 c0
c0 c1
.
u
 G
)
+
v
G
v1
1
1
1
u
u

u
u


(5.176)

For the Greens function introduced above, B1 results in


B1 =

B1

B1+

s)
2 v1 (
=
c

d
u[cin
u) cout
u)]2 .
0 (
0 (

(5.177)

6. Application of Diuse Interface Modeling


to Hydrodynamically Driven Growth

Taking up the line of argumentation of the previous chapter, here I will start
with a diuse interface model of hydrodynamically driven growth at solid
liquid interfaces. In particular I will be interested in the dendritic morphology
which is known to occur for quenches taking the system under investigation
from the liquid region of the underlying phase diagram to the liquid/solid
coexistence region (see Fig. 5.3 and Fig. 5.4). Basically dendritic growth
can be pictured as the advance of a solid phase shaped like a dendrite (see
Fig. 6.1) into the liquid phase. One question tied to this growth phenomenon
is that of the precise shape and the precise velocity selected during the growth
process. This is referred to as the selection problem of dendritic growth. Here
I will reconsider the selection problem of the dendritic morphology starting
from a thermodynamically consistent phase-eld model of hydrodynamically
inuenced dendritic growth. I will take its thermodynamic consistency as
validation of the model. I will then continue to carry out the asymptotic
analysis of the diuse interface model. This analysis reveals an additional
term beyond the ones classically known within the sharp interface equations
describing dendritic growth. Due to the thermodynamic consistency of the
diuse interface model from which it is derived I claim this additional term
to reveal further insight into the physics at the dendritic interface which
has not been considered classically. This is relevant for the comparison of
classical dendritic growth theory and related experiments, which still reveals
unsettled issues. Within this chapter I will recall these issues and meet them
with an analysis of the new model equations including the non-classical term
obtained only by the analysis of the related thermodynamically consistent
diuse interface model in the beginning.
Classically the problem is treated based on a set of equations containing
the NavierStokes equations and the relevant diusion equations in the bulk
phases. These have to be supplemented by the correct boundary conditions
given by the GibbsThomson relation, the Stefan condition, the no-slip condition and the condition for the conservation of mass at the interface. The
basic structure of this set of equations will remain unchanged within this
chapter, however an additional curvature term within the Stefan condition
implies an energetic contribution in the interfacial region relevant for further
consideration.
Heike Emmerich (Ed.): LNP m73, pp. 97130, 2003.
c Springer-Verlag Berlin Heidelberg 2003


98

6. Application of Diuse Interface Modeling

Fig. 6.1. Picture of a single dendrite. Its precise shape and velocity are the main
issues of this chapter. This dendrite was grown at the Laboratory of Solid State
Physics of the ETH Z
urich in the group of Prof. Bilgram. Picture kindly provided
by Stalder.

A couple of diuse interface models for solidliquid interfaces including hydrodynamic transport in the bulk can be found in the literature so
far [12, 13, 15, 17, 85, 250]. The basic assumption underlying their formulation is that the solid phase is well approximated by a highly viscous isotropic
liquid. This requires the inclusion of Korteweg stresses, which can drive convection [190]. Korteweg stresses are known from the theory of uids near a
critical point, where they are given by a capillary tensor acting as the reversible part of the stress tensor1 . Having in mind the structure of diuse
interface models for diusion limited growth, the basic formalism of models
extended to hydrodynamic transport arises from the transformation
D

:=
+v,
t
Dt
t

(6.1)

where v is the vector of the uid velocity eld. Moreover, consistency with
a formulation per unit mass rather than per unit volume has to be ensured,
since obviously mass is a conserved quantity whereas local density is not.
One can recognize the quantity on the right hand side of (6.1) as the material
derivative with respect to time, i.e. the phase-eld equations become invariant
under Galilean transformation.
1

The irreversible part is provided by the standard viscous stress term of a Newtonian liquid.

6.1 Diuse Interface Modeling for Hydrodynamically Inuenced Growth

99

The diuse interface model, which underlies the analysis of this chapter, was presented before at the end of Chap. 4. An asymptotic analysis of
the model was given by Anderson et al. in [18]. Here I will recall their argumentation and summarize some of their results as starting point for the
reconsideration of the dendritic selection problem with density change ow
included.

6.1 Diuse Interface Modeling for Hydrodynamically


Inuenced Growth
The formulation of the headline of this section reects some of the diculties one encounters when talking about growth of a phase-separating interface taking into account hydrodynamic transport in the liquid bulk phase.
Whereas diusion limited growth is a xed expression, convection limited growth does not seem to describe the possible physical set-up belonging to that expression correctly: usually hydrodynamic as well as diusive
transport of the temperature eld2 will be active in the bulk. In most cases
diusive transport will even be the dominant one with hydrodynamic transport only imposing corrections. Nevertheless it remains challenging to fully
grasp these corrections quantitatively for the big range of applications appearing in e.g. solidication, thin lm growth, epitaxial growth or electrodeposition, to mention only a few. Modeling certainly is the rst step to deal
with these problems. Within the framework of diuse interface modeling it
was demonstrated at the end of Chap. 4 following [14] how a suitable
model can be derived by expressing entropy production in terms of transport
variables and identifying the forms which ensure entropy production to be
non-negative. One important feature of this model is that it couples uid
motion to a non-conserved order parameter description. This oers the possibility to treat quasi-incompressible systems [208], in which the densities of
the two bulk phases are each spatially uniform. However, the overall density
can be formulated as function of the order parameter varying smoothly
within the interfacial region. Thus
() = S r() + L [1 r()] ,

(6.2)

where S and L are the densities in liquid and solid, respectively. The function r() is monotonically increasing with r(0) = 0 and r(1) = 1. Suitable
choices include r() = and r() = 2 (3 2). Assumption (6.2) is called
quasi-incompressible, since density may vary in the interfacial region but
thereby does not depend on pressure. It places a constraint on the form of
the underlying thermodynamic potential as stated in [208]. In particular it
requires that the underlying Gibbs free energy density is given by
2

In case of multicomponent materials diusion of solute will be involved in the


growth process as well.

100

6. Application of Diuse Interface Modeling

g(T, p, ) = g0 (T, ) +

p pR
,
()

(6.3)

where pR is a reference pressure associated with an isothermal, stationary


planar interface at melting temperature TM .
In their analysis in [18] Anderson et al. take g0 (T, ) as


T
1
Hm () (1
)
g0 (T, ) = e0 cTM r()L
4aS
TM
1
T
) + Hm () ,
(6.4)
c T ln(
TM
4a
in which case the corresponding expressions for the internal energy and entropy densities are
1
pR
Hm ()
4aE
()
1
T
1
s=
Hm ()] + c ln(
).
[e0 r()L +
TM
TM
4aS

e = e0 + c(T TM ) r()L +

(6.5)
(6.6)

Here 1/aE = 1/a 1/aS . 1/a is simply the height of the double well of the
Gibbs free energy density at T = TM . 1/aE and 1/aS are heights of the
double wells within the internal energy and entropy densities, respectively.
Moreover, the quantity e0 is a constant reference energy and both the heat
capacity per unit mass c and the latent heat per unit mass L are assumed to
be constant.
Note that the doublewell function Hm appears in the per unit mass
quantity g0 (T, ). The analog per unit volume quantity associated with
()g0 (T, ) is ()Hm () Hv (). A common form of a doublewell potential is the function 2 (1 )2 . In standard phase-eld models, which do not
include convection, it is usually associated with a per unit volume quantity (see e.g. [289]). In applications, in which the density of the two phases
is constant and equal, the per unit mass and the per unit volume specication of the double well are equivalent. However, in the present situation
the bulk densities are not necessarily equal. Thus the per unit mass formula
Hm () = 2 (1 )2 and the per unit volume formula Hv () = L 2 (1 )2
with subsequent Hm () = L 2 (1 )2 /() result in dierent approaches.
For a detailed description of the algebra involved in the asymptotic analysis of (4.98)(4.101) the interested reader is referred to [18]. Here I will merely
state the basic steps and results: Assuming an expansion of outer variables
as
u = u0 + u1 + 2 u2 + ...
2

p = p0 + p1 + p2 + ...
= 0 + 1 + 2 2 + ...
= 0 + 1 + 2 2 + ...

(6.7)
(6.8)
(6.9)
(6.10)

6.1 Diuse Interface Modeling for Hydrodynamically Inuenced Growth

101

and an analogous expansion of inner variables as


U = U0 + U1 + 2 U2 + ...
1
P = [P0 + P1 + 2 P2 + ...]

= 0 + 1 + 2 2 + ...
= 0 + 1 + 2 2 + ...
1
M = [M0 + M1 + 2 M2 + ...]

(6.11)
(6.12)
(6.13)
(6.14)
(6.15)

leading-order equations in the bulk phases read:


Du0
= p0 + P r 0
Dt
u0 = 0
D0
SP r
0
= k2 0 +

Dt

0 in the liquid phase
0
=
1 in the solid phase .

(6.16)
(6.17)
(6.18)
(6.19)

Here is related to the internal energy e via


Sp 1

.
= e r() + Hm ()

2

(6.20)

Apart from the last term on the right hand side of (6.18) the above set
of equations (6.16)(6.18) are nothing but the incompressible NavierStokes
equations and the diusion equation for the temperature eld respectively.
0 denotes the leading-order terms of the viscous stress tensor given at the
end of Chap. 4. P r refers to the well known Prandtl number representing
the ratio of kinematic viscosity to diusivity in the material. The additional
r
term SP

0 can be understood as a source term resulting from the rate of


work of forces acting at the boundary of each volume element (compare to
(4.80)). For small coecients Pr this term can be neglected.
To obtain information about the physics in the interfacial region resulting
asymptotically in boundary conditions, I turn to the analysis of the behavior
of inner variables invoking matching and solvability conditions as described
in the previous chapter:
Continuity Equation
The leading-order problem for the continuity equation (4.98) can be
solved to obtain the form of U03 through the interface as
U03 =

J0
+ Vn ,
(0 )

(6.21)

Vn ) = S (u0 |S n
Vn ). Thus matching the
where J0 = L (u0 |L n
velocities results in

102

6. Application of Diuse Interface Modeling

Vn )|L
(u0 n
S =0,

(6.22)

which is nothing but a boundary condition stating conservation of mass


through the interface.
Momentum Equation
As a second condition for the velocity components one would expect
the so called no-slip condition for the tangential velocity components to
arise at the interface. Indeed it can be recovered by an analysis of the
momentum equation (4.99). At leading-order [O(1 )] one nds that the
normal3 derivatives of the velocity components satisfy
(u0 n
)]|L
[u0 n
S =0.

(6.23)

Together with (6.22) this implies that the tangential components of the
velocity eld vanish.
Phase-Field Equation
Analyzing the phase-eld equation (4.100) at leading- and at rst-order
yields a solvability condition, which can be expressed as
L(1

TI
0
L 1
Vn )
) = K (u0 n
.
TM
L
TM mob

(6.24)

Here TI is the interface temperature and mob the interface mobility. K


represents the curvature of the interface in a two-dimensional system,
0 the surface energy. Thus (6.24) is nothing but the classical Gibbs
Thomson condition with the two terms on the right-hand side of the
equation summarizing the eects of curvature and attachment kinetics,
respectively.
Energy Equation
Last the energy equation (4.101) has to be taken into account. Leading
order contributions originate from dissipation, from thermal diusion and
from the advective term in the energy equation. Examining the O(1)
problem one can identify a heat ux boundary condition. The respective
analysis involves the inner velocity component correction U13 , which is
determined from the correction to the continuity equation. It yields a
heat ux boundary condition as follows
T |L
Vn )[L L + 0 K] .
kn
S = (u0 |L n

(6.25)

This is the classical Stefan condition modied to account for motion in


the solid and liquid bulk phases as well as the curvature of the interface.
The eect of ow only enters in the rst factor on the right-hand side,
3

normal with respect to the interface

6.1 Diuse Interface Modeling for Hydrodynamically Inuenced Growth

103

where it arises as the normal velocity of the material relative to the


2
( ) D
interface. The curvature eect follows from the term E
Dt in
(4.101). It represents the internal energy gradient and doublewell terms
in the model.
In [18] Anderson et al. discuss some earlier work considering this term:
In the context of phase-eld modeling it had been identied before by Fife
and Penrose [116] and Fried and Gurtin [122]. Within sharp interface formulations Wollkind and Maurer [299], Umantsev and Davis [284], Zhang and
Garimella [301] and Schlitz and Garimella [247] had studied its eect on
the stability of an interface in the context of diusion limited growth and
Lemieux and Kotliar [205] its inuence on velocity selection during diusion
limited dendritic growth. They showed that in this case the curvature term
increases the values of selected velocities. Viewed in the light of experiments
by Glicksman et al. summarized in 1994 in [129], the ndings of Lemieux et
al. dating back to 1987 appear quite interesting. Glicksman et al. researched
the dierence between the theory of diusion and hydrodynamically inuenced dendritic growth in detail based on microgravity experiments. Prior to
these experiments Huang and Glicksman had performed terrestrial growth experiments using the transparent organic material succinonitrile (SCN) [149].
The data of the latter was in good agreement with the analytical theory
for velocity selection in diusion limited dendritic growth4 except for low
undercoolings. A plausible reason appeared to be the fact that at low undercoolings the driving force due to the temperature gradient gets smaller
and smaller. In turn hydrodynamic eects start to become more and more
important and could become visible exactly in such a manner that it would
explain the deviations from the analytical curve. Microgravity experiments
were thought to be the way out, since buoyancy as main force driving hydrodynamic ow is suppressed. Nevertheless the experiments by Glicksman
et al. revealed deviations from the analytical curve at low undercoolings, as
well, even though somewhat smaller in their absolute value (Fig. 6.2). Indeed
there exists some ow, which occurs under microgravity as well, namely the
so called density change ow. This expression density change ow refers to
hydrodynamics with the only mechanism driving hydrodynamic ow being
the density change the material undergoes when solidifying. Thus at this
point one could wonder whether density change ow does lead to noticeable
deviations from diusion theory in particular at low undercoolings. However,
steady state solutions of the classical sharp interface equations governing density change ow had been determined by McFadden and Coriell in 1986 [212].
They showed that the eect of density change ow is of the order of the density dierence between solid and liquid phase normalized with respect to the
density in the liquid, which I will call $ in the following:
4

A thorough review discussing velocity selection in the case of diusion limited


dendritic growth is given in [42].

104

6. Application of Diuse Interface Modeling

$=

s l
.
l

This implies that it is minor for most experimental set-ups, in particular for
SCN with $ = 0.028. (Other values are, e.g., $ = 0.065 for aluminium and
$ = 0.09 for silicon. Note that for silicon density increases upon solidication.) Moreover, in comparison to diusion limited dendritic growth their
analysis reveals a decrease of growth rates with $. The experimental curve
by Glicksman et al., on the other hand, displays an increase (see Fig. 6.2).
Thus even qualitatively density change ow proved to have rather the opposite eect than the one desired to explain the experiments. McFadden and
Coriell did not proceed their studies with a perturbative analysis as necessary to investigate the inuence of surface tension on the steady state solutions. A comprehensive solution including the dependence of growth rates on
anisotropic surface tension obtained via an irregular perturbation expansion was rst given for diusion limited dendritic growth by Brener [43]. It
reveals corrections as 7/4 . From this one can assume that for SCN with
anisotropy factor = 0.55 [224] considering surface tension eects within the
McFadden/Coriell-solution cannot help to explain the experiments by Glicksman et al.. Thus at this point it seems useless to carry out the full irregular
perturbation analysis for the classical problem of density change inuenced
dendritic growth hoping to gain new insight compatible with experimental
evidence. However, turning from the classical sharp interface formulation of
the problem towards the non-classical one with an additional curvature term
in the energy equation as in (6.25), one can hope that the ndings of Lemieux
and Kotliar for diusion limited dendritic growth will carry over, i.e. that also
in the case of density change inuenced dendritic growth an irregular perturbation expansion taking into account the additional curvature term will result
in an increase of selected velocities and thereby explain the experiments.
The following section is devoted to an investigation of this idea. A rst
step is an analysis of (6.16)(6.18) and boundary conditions (6.22)(6.25) in
the framework of dendritic selection theory. This analysis is carried out in
a way originally proposed by Kruskal and Segur in [191]. In the context of
dendritic growth this method was employed by various authors, see e.g. [236]
and references therein. It starts from the derivation of a zero surface tension
solution. This solution is subsequently analyzed further within the framework
of regular as well as irregular perturbation theory.

6.2 The Selection Problem


of Dendritic Growth Revisited
To begin a detailed study of dendritic selection in the case of density change
ow at this point I recast the governing equations and boundary conditions
D
is given
suitably for the parameter regime Pr
1 [117]. Here the operator Dt

6.2 The Selection Problem of Dendritic Growth Revisited

105

Fig. 6.2. Figure taken from [129]. The dotted line refers to the selection theory of
diusion limited dendritic growth, the solid data points to microgravity experiments
and the white squares and circles to two dierent terrestrial experiments. The
theory curve for density change inuenced growth would run below the dotted
curve depicting the diusion theory (compare to Fig. 6.4).

by the full form t


+u to elucidate the necessary coordinate transformations
in the following. Moreover, the thermal diusivity D, the volumetric specic
heat c = s cs = l cl and Ts cs = Tl cl are assumed equal in both phases.

Basic Governing Equations


In the liquid phase the heat conduction equation for the temperature eld
T corresponding to (6.18) can be written
T
+ (u) T = D 2 T ,
t

(6.26)

where u denotes the velocity of the hydrodynamic eld satisfying mass conservation
u=0
(6.27)
and the momentum equation
u
1
+ (u) u = p + 2 u .
l
t

(6.28)

106

6. Application of Diuse Interface Modeling

is the kinematic viscosity of the material in the molten (liquid) state, l the
density of the melt and p the pressure eld. (6.26)(6.28) hold for a frame of
reference at rest as well as moving at constant velocity.
Within the solid phase, the heat conduction equation is of the same type
as (6.26), however the velocity is constant and determined by the velocity of
the reference frame. Choosing it as Vfr = V ex , one obtains
T
+ (V ex ) T = D 2 T .
t

(6.29)

At the crystalmelt interface the uid ow satises conservation of mass as


well as a no-slip boundary condition:
l (u V ex ) n = (l s )(v V ex ) n ,
(u V ex ) t = 0 ,

(6.30)
(6.31)

where v denotes the velocity of the interface. These equations correspond


to (6.22) and (6.23), respectively. n and t denote the unit normal and unit
tangential vector to the moving interface. In (6.30) and (6.31) the velocity
uV ex refers to a frame of reference at rest. The temperature directly at the
interface is specied claiming local thermodynamic equilibrium, which yields
continuity of temperature
T l = Ts ,
(6.32)
as well as the GibbsThomson relation



.
T = TM 1
s Q

(6.33)

Equation (6.33) corresponds to (6.24) neglecting kinetic terms. Here, TM is


the crystallization temperature of the planar interface, s Q the latent heat
per unit volume of solid. It is related to the mean latent heat L appearing
in (6.24) via L = sl Q. denotes the anisotropic surface tension (see (2.8))
and the local curvature of the interface.
Moreover, energy is conserved for the phase transition occurring at the
interface. Thus:



(u V ex ) n = [(kT )s (kT )l ] n .
s Q 1
(6.34)
s Q
This is the nal boundary condition at the interface corresponding to (6.25).
The left-hand side of formula (6.34) expresses the latent heat release per unit
volume of the solid phase. It is equal to the total energy ux away from the
interface into both phases as given on the right-hand side of (6.34). Introduc which takes into account unequal specic heats
ing a modied latent heat L,
cl and cs as follows
= Q + cs TM cl TM ,
L
(6.35)

6.2 The Selection Problem of Dendritic Growth Revisited

107

it is possible to replace (6.34) by

s L

(1
)(u V ex ) n = D [(T )s (T )l ] n .
l cl
s Q

(6.36)

Boundary conditions at innity (x ) are assumed to be given by:


u V ex

(6.37)
(6.38)

T T .
Dimensionless Governing Equations

The analysis of the above set of equations (6.26)(6.38) with respect to the
selected growth velocities is simplied choosing dimensionless variables by
2
rescaling spatial coordinates by and the time coordinate by D .5 Here
is a reference length scale specied as tip radius of curvature of the zero
surface tension solution. The velocity eld u is expressed in units of D
, the

temperature T in units of cLl and pressure by D2l . Temperature is measured


relative to the melting temperature of the at interface TM , i.e. T = (T 


TM
)/ cLl , where the prime indicates the dimensional temperature.
Expressed through dimensionless variables heat conduction in the liquid
phase reads
T
+ (u) T = 2 T .
(6.39)
t
Mass conservation (6.27) remains formally unchanged. The momentum equation is now given by
u
+ (u) u = p + P r2 u ,
t

(6.40)

is the Prandtl number. To eliminate pressure one determines


where P r = D
the curl of equation (6.40). This results in the vorticity equation

+ (u) = P r2 .
t

(6.41)

where = ( u) ez is the scalar vorticity. Heat conduction in the solid


phase takes the dimensionless form:

T
+ Pc (ex ) T = 2 T ,
t

(6.42)

Actually the analysis of the phase-eld model as summarized in the previous section is simplied by choosing dimensionless variables, as well. Moreover curvilinear coordinates were employed within the original work. To keep the summary of
the previous section short such details were omitted. Again the interested reader
is referred to [18].

108

6. Application of Diuse Interface Modeling

s
6
where Pc = V
D is the Peclet number . With l = 1 + $, the boundary conditions for the hydrodynamic eld at the interface expressed in dimensionless
variables read:

u n = Pc (1 + $)ex n $v n ,
u t = Pc ex t .

(6.43)
(6.44)

Condition (6.32) stating continuity of temperature remains unchanged. The


GibbsThomson equation now reads
T = a() ,

(6.45)

T c

o
where = do = M
Q2 is the capillary length do in units of .
Finally, energy conservation can be expressed as

(1 + $ ) (Pc ex + v) n = [(T )s (T )l ] n .

(6.46)

The boundary conditions at innity become (6.38) and


u Pc e x .

(6.47)

Equations in Parabolic Coordinates


A further step to simplify the following analysis can be taken by transforming
the above equations to parabolic coordinates (, ) (see Fig. 6.3) dened by
x = 12 ( 2 2 ) and y = [219]. This transformation allows for a separation
of variables. Moreover a stream function (, ) dened as
w = 
w = 

(6.48)

2 + 2

(6.49)

1
2

+
1

allows me to rewrite the governing bulk equations as follows:


Liquid Phase
( 2 + 2 )

T
T
2T
T
2T
+

=
+
2
t

(6.50)

Note that Pc is not exactly the same as the interfacial Peclet number P eInt =
v/D (see Sect. 5.3) in the sense that the length is related to the diuse
interface thickness of a phase-eld model, whereas the length scale denotes
a characteristic length scale of the sharp interface model (6.26)(6.38). Here it
can be identied with the tip radius of the zero surface tension solution of the
problem.

6.2 The Selection Problem of Dendritic Growth Revisited


=4

y=10

=3

109
=2

=2

=1
=1

y=0

=0

=0

=-1

=-2
y=-10

x=-10

x=0

x=10

Fig. 6.3. Denition of parabolic coordinates (, ) as underlying the analysis in


the remainder of this section [219].


( + )
+

= Pr
t


2

2T
2T
+
2
2

2 2
+
+ ( 2 + 2 ) = 0
2
2

(6.51)
(6.52)

Solid Phase
( 2 + 2 )



2T
T
T
2T
T
=
+
.
+ Pc

2
2

(6.53)

Interface With the interface shape function s dened by s (, t) = 0


one can rewrite the boundary conditions expressed in parabolic variables
as follows:
Mass Conservation



+ s = Pc (1 + $) (s ) + $ 2 + s2 t s .
(6.54)
Continuity of tangential components of the hydrodynamic eld

110

6. Application of Diuse Interface Modeling

s + = Pc ( s s ) ,

(6.55)

where the prime denotes dierentiation with respect to .


Energy conservation




(1 + $ ) Pc (s ) + s2 + 2 t s = (s + ) (Ts Tl ) .
(6.56)
Continuity of temperature (6.32)
GibbsThomson Relation
T = a[s ()][s ()] ,

(6.57)

where the curvature is determined by




s
s
s

(6.58)
[s ()] = 

3 +
1

2 2
2 + s2

2
2
1 + s
( + s ) 1 + s 2 2
and anisotropy by
a[s ()] = 1 + 8

(s + s ) ( s s )
2 .
2
( 2 + s2 ) 1 + s 2

(6.59)

Innity Boundary conditions at innity are now given by (6.38) and


Pc ,

(6.60)

Pc .

(6.61)

(The underlying derivation of operators in parabolic coordinates can be


found in the appendix of this chapter.)
Steady State Solutions
A rst step when searching for the solutions to the problem given by (6.50)
to (6.61) is to derive steady state solutions for the case of zero surface tension. These solutions can then be used as basis to nd the solutions of the
full problem in a perturbative manner. In the case of zero surface tension
the temperature at the interface is constant (T = 0) and a family of exact
solutions can be found. In order to determine these solutions one assumes
that T depends on only, so that the shape of the interface is a parabola
s () = 1. In this case equation (6.50) reduces to:

T
2T
.
=

2

Thus for the stream function the following ansatz

(6.62)

6.2 The Selection Problem of Dendritic Growth Revisited

= f ()

111

(6.63)

must hold. With (6.52) this yields a vorticity

= 
f  .
2
+ 2

(6.64)

The simplest solution of (6.51) is the one for which the vorticity vanishes everywhere, i.e. f  = 0. Integrating twice one nds that all boundary conditions
can be met:
f () = Pc ( + $) .
(6.65)
Inserting relation (6.65) into (6.62) and integrating once results in




Pc
T
Pc
(1 + $)2 exp ( + $)2 .
= Pc (1 + $) exp

2
2

(6.66)

One further integration yields




Pc
T () = Pc (1 + $) exp
(1 + $)2
2

 
1



Pc
2
dx exp (x + $) .
2

(6.67)

This can be expressed in terms of the complementary error function as







Pc
Pc
P
c
(1 + $) exp
(1 + $)2 erfc
( + $) . (6.68)
T () T =
2
2
2
Moreover, the dimensionless undercooling can be obtained with = 1 in
(6.68) as:





P
Pc
Pc
c
(1 + $)2 erfc
= T =
(1 + $) . (6.69)
(1 + $) exp
2
2
2
As stated at the end of Sect. 6.1 steady state solutions of dendritic growth
considering density change ow have been determined before by McFadden
and Coriell in [212]. Their analysis was carried out in cylindric coordinates
V
rather than V
with same scaling, however dening Pc as 2D
D as assumed
here. The conclusion they draw from their analysis coincidences with (6.68),
i.e. the eect of density change ow is of order $ and hence minor for most
experimental set-ups. The remainder of this chapter is devoted to the question
whether surface tension eects alter this preliminary view on dendritic growth
inuenced by density change ow noticeably. Within the context of classical
sharp interface equations this analysis would have been useless as pointed out
at the end of Sect. 6.1. However in connection with the results by Lemieux
and Kotliar one can hope that the additional curvature term within the
energy condition (6.34) will increase values of selected velocities in a way,
that they provide an explanation for the experimental data in [129] at low
undercoolings.

112

6. Application of Diuse Interface Modeling

Regular Perturbation Expansion


A regular perturbation expansion serves to determine the stability of the
above solution. To carry it out, the following expansions are employed:
T = T 0 + Pc (1 + $)T 1 + . . .
s0 +
0

s1

s =
+ ...
= + Pc 1 + . . .
= 0 + Pc 1 + . . . ,

(6.70)
(6.71)
(6.72)
(6.73)

where = Pc (1+
) . Inserting these expansions the following set of rst-order
equations is obtained:

Liquid Phase
2T 1
0 T 1
1 T 0
2T 1
0 T 1
+
=
+

2
2

(6.74)

T 1
T 1
2T 1
2T 1
1 T 0
+
P

(
+
$)
+

P
=

c
c

(6.75)

2 1
0 1
2 1
0 1
=0
+ Pr

+
2
2

(6.76)

2 1
1
2 1
1
+
Re(
+
$)
=0
+

Re

2
2

(6.77)

2 1
2 1
+
+ ( 2 + 2 ) 1 = 0
2
2

(6.78)

2T 1
T 1
T 1
2T 1
+ Pc
=0
+
Pc
2
2

(6.79)

Pr

Solid Phase

Interface
 
1 = $ s1
 
1 = $ s1


s1



+ Pc (1 + $)s1 + Pc (1 + ) 2 +

(6.80)
(6.81)

 1
T Ts1 = 0
l

Ts1 = Tl1 s1

(6.82)
(6.83)

23

Ts1 = (1 + )

(6.84)

6.2 The Selection Problem of Dendritic Growth Revisited

113

Innity
T1 0
1 0
1 0

(6.85)
(6.86)
(6.87)

at as well as at .
Smooth Tip Condition
 1 
s (0) = 0

(6.88)

s1 (0)

(6.89)

=1

at = 0.
Solving the First Order Equations
A solution of the rst-order equations can be obtained via separation of
variables [117]. The starting point of the following analysis is the homogeneous heat conduction equation associated with the inhomogeneous
equation (6.75) valid in the liquid phase7 :
Th1
2 Th1
Th1
2 Th1
+
P
=0.
(
+
$)
+

c
c
2

2
Based on the ansatz

(6.90)

Th1 (, ) = X()Y (),

this can be rewritten as


X  Pc X  + 2Pc X

=0,
Y  + Pc ( + $) Y  2Pc Y = 0 .
Choosing

x=

Pc 2
2 ,

X = X(x) ,

(6.91) is transformed into Kummers equation




1

xX +
x X  + X = 0 ,
2
with a fundamental set of solutions given by

M (, 12 , x) (regular at = x = 0)
X(x) =
U (, 12 , x) (singular at = x = 0) .
7

(6.91)
(6.92)
(6.93)

(6.94)

(6.95)

The denitions of special functions appearing in the succeeding analysis follow


mainly [3], partly [219].

114

6. Application of Diuse Interface Modeling

Here M and U are the conuent hypergeometric functions. Suppressing the


divergence of the solution at = x = 0, an appropriate choice of X(x) is
restricted to
()1
()2 2
()n n
1
X(x) = M (, , x) = 1+ 1 x+ 1
x +. . .+ 1
x +. . . , (6.96)
2
( 2 )1
( 2 )2 2!
( 2 )n n!
where (a)n is the Pochhammer symbol, i.e. (a)n = a(a+1)(a+2) . . . (a+n1)
and (a)o = 1. On the other hand, M (, 12 , x) grows exponentially, i.e.:
 
12 1 x
1
2e
M (, , x)
as x ,
(6.97)
x
2
()
unless () is innite. This implies = 0, 1, 2, . . ., in which case the series
in (6.96) truncates. Thus the acceptable solutions of (6.91) are the Hermite
polynomials


n



1 Pc 2
1
n!
X() = M n, ,

=
H2n
Pc ,
(6.98)
2 2
(2n)!
2
where = n = 0, 1, 2, . . .
Moreover, turning towards (6.92) it is convenient to introduce
y=

Pc ( + $), Y = e

y2
4

Z(y) .

This results in Webers equation:



 2
1
y

+ 2n +
Z=0,
Z
4
2

(6.99)

(6.100)

which fundamental set of solutions read



 
U 2n + 12 , y  = D2n1 (y)
Z(y) =
,
V 2n + 12 , y

(6.101)

where U and V are parabolic cylinder functions. For large y they display the
asymptotic behavior:
e

y2
4

x2n1 e
D2n1 (y) 


y2
e 4 V 2n + 12 , y 2 y 2n .

y2
2

(6.102)

For exponential decay of the solution as , one has to choose D2n1 (y)
accordingly:


Pc
2
Pc ( + $) .
(6.103)
Y () = e 4 (+ ) D2n1
Thus the solution of the homogeneous equation (6.90) can be expanded as

6.2 The Selection Problem of Dendritic Growth Revisited

Th1 (, )

n H2n



Pc

 e P4c (+
e

n=0

Pc
4

(1+

115


Pc ( + $)

 . (6.104)
)2 D
Pc (1 + $)
2n1
)2

D2n1

As a consequence the overall solution of (6.75) for inhomogeneous heat


conduction can be expressed based on a Greens function formalism as:
Tl1 (, )

Th1 (, )

d
0

d  G(, ;  ,  )

1 T 0  
( , ) .

(6.105)

Here one can dene the general formula for the Greens function G(, ;  ,  )
as a solution of
2G 2G
G
G
= (  )(  ) ,
+
Pc
+ Pc ( + $)
2
2

(6.106)

with suitable boundary conditions at = and = 0. The Greens function


can be obtained following the general outline given in [219]. As a function of
it can be expanded into a set of Hermite polynomials as given in (6.98). A
general approach to G taking into account - as well as -dependence reads
G(, ;  ,  ) =


An (;  ,  )H2n ( Pc ) .

(6.107)

n=0

Inserting (6.107) into (6.106) yields:



 2



A
G(, ;  ,  ) =
A
+
P
(
+
$)

2P
nA
n
c
n
c
n H2n ( Pc )
2

n=0
= (  )(  ) .

(6.108)

At this point it is appropriate to take advantage of the orthogonality relation


of the Hermite polynomials:


t2

dt e 2 Hn (t)Hm (t) =

for n = m
.
2n! for n = m

(6.109)

Since H2n (t) is an even function of t an alternative way to express this orthogonality is given by:

0

de

Pc 2
2





H2n ( Pc )H2m ( Pc ) = nm
(2n)! .
2Pc

(6.110)

Applying the projection operator of (6.110) to (6.108) one obtains


2

Bn + Pc ( + $) Bn 2Pc nBn = (  ) ,
2

(6.111)

116

6. Application of Diuse Interface Modeling

where
Bn =

An
e

Pc
2

2Pc (2n)!

2 H

2n (

Pc  )

Thus the problem is reduced to nding a one-dimensional Greens function


dened by (6.111) with general form [3]:

1
y1 ()y2 (  ) for 
Bn =
.
(6.112)
(y1 , y2 ) y2 ()y1 (  ) for > 
Here the Wronskian (y1 , y2 ) is evaluated at  . y1 and y2 are two independent solutions of the homogeneous equation
2

y + Pc ( + $) y 2Pc ny = 0 ,
2

(6.113)

as given by (6.99) and (6.101), or a suitable linear combination of both.


The choice of y1 and y2 depends on the boundary conditions one imposes.
Suppressing divergence at innity results in:


Pc
2
y2 () = e 4 (+ ) D2n1
Pc ( + $) .
(6.114)
For y1 one is more exible and can introduce an additional parameter n ,
which allows for a variety of boundary conditions one does not need to specify
at this point. Thus:



y1 () = n D2n1
Pc ( + $)


Pc
2
1 
V 2n + , Pc ( + $) e 4 (+ ) .
(6.115)
2

For the subsequent calculation of the Wronskian (y1 , y2 ) = y1


y2 y2
y1
it is appropriate to dene the abbreviations:



x = P ( + $) , c = n , u(x) = D2n1
Pc ( + $)
,


1 
v(x) = V 2n + , Pc ( + $) .
2

With these notations solutions y1 and y2 read:


y1 = [cu(x) v(x)] e

x2
4

, y2 = u(x)e

x2
4

Denoting dierentiation with respect to x by a prime, the Wronskian can be


rewritten as


2
2

(y1 , y2 ) = y1 y2 y2 y1 = Pc (uv  vu )ex /2 = Pc (u, v)ex /2 ,

6.2 The Selection Problem of Dendritic Growth Revisited

117

where (u, v) = uv  vu is given in [3] as:



2
(u, v) =
.

Thus:

2Pc P ( + )2
e 2
.

This is independent of c = n . With these expressions the Greens function


reads:
Pc

2
2
(6.116)
G(, ;  ,  ) = e 2 [( + ) ] g(, ;  ,  ) ,
(y1 , y2 ) =

where
g(, ;  ,  )
(6.117)







P
2
1
c
H2n ( Pc )H2n ( Pc  )D2n1
=
Pc (> + $) e 4 (> + )
(2n)!
n=0





Pc
2
1 
n D2n1
Pc (< + $) V 2n + , Pc (< + $) e 4 (< + ) .
2
Here > refers to >  , whereas < refers to <  .
To solve for heat conduction in the solid phase, separation of the
two variables and is employed as well:
Ts1 (, ) = X()Y ().
This yields (6.91) and (6.92) with $ = 0. Following the steps to derive (6.98),
X() is determined in the same manner as for the homogeneous heat conduction problem in the liquid, yielding (6.98). To solve (6.92) for $ = 0 it is
convenient to choose
(6.118)
y = P2 2 , Y = Y (y) .
Thus (6.92) for $ = 0 is transformed into Kummers equation (see (6.94)).
The solution regular at y = = 0 is M (n, 12 , y). As a consequence

Y () = M

1 Pc
n, , 2
2
2


=

n!
(2n)!

1
2

n



H2n i Pc .

(6.119)

Therefore the solution of heat conduction in the solid can be expanded as:


 H iP 

2n
c
1
  .
(6.120)
Ts (, ) =
n H2n
Pc
i
H
P
2n
c
n=0

118

6. Application of Diuse Interface Modeling

Linear Solvability Theory


Accordingly y1 of (6.115) is selected as
 



1 
y1 () = V 2n + , Pc (1 + $) D2n1
Pc ( + $)
2


 
Pc
1 
D2n1
Pc (1 + $) V 2n + , Pc ( + $) e 4 (+
2

)2

As a consequence n is determined via (6.115) and thereby species the


Greens function G based on (6.117). In turn G enters (6.105) and thus selects
Tl , which has to satisfy the rst-order boundary conditions (6.82)(6.84). At
this point one has to realize that it is impossible to satisfy all three of these
boundary conditions together. Thus the steady state solutions derived in the
beginning of this section are destroyed by isotropic surface tension.
Irregular Perturbation Expansion
However, from diusion limited dendritic growth it is known as well that
anisotropy of surface tension can cure the problem we encountered at the
end of the last section, i.e. destruction of steady state solutions through
isotropic surface tension. In fact, anisotropy of surface tension selects a discrete spectrum of growth velocities for the case of diusion limited dendritic
growth, of which only the fastest growth mode is linearly stable8 . In this
case it is necessary to study the additional eects arising from anisotropy in
the complex plane, e.g. via a WKB formalism [236]. The key point of such
an irregular perturbation expansion is the evaluation of a related eigenvalue
problem in the complex plane. This eigenvalue problem itself arises from the
dierence between the governing equation with zero surface tension solution
inserted respectively the perturbed solution inserted. Its analysis results in a
matching condition between the regular and the irregular parts of the solution. One can analyze this solvability condition further to derive the selected
tip velocity as:
D(1 + $)() 2
Pc (, $) ,
(6.121)
VTip =
d0
where () refers to the dimensionless growth rate known from diusion
limited dendritic growth [42] and Pc2 (, $) to the respective Peclet number,
which here is a function of $ as well. The precise relation determining Pc2 (, $)
reads





Pc
Pc
Pc
2
=
(1 + $) exp
(1 + $) erfc
(1 + $) .
(6.122)
2
2
2
8

A concise review article discussing the selection problem in the case of diusion
limited dendritic growth is given in [42].

6.3 Comparison to Experimental Data

119

It is obtained just as in the case of diusion limited dendritic growth from


the relation expressing the temperature for the zero surface tension steady
state solution right at the interface. The as well as the $dependence of
Pc is visualized in Fig. 6.4.

6.3 Comparison to Experimental Data


l
Obviously it is $ = s
which contains the density dierence between liquid
l
and solid phase and thus the origin of the additional driving force, which
characterizes this growth problem compared to diusion limited dendritic
growth. Five dierent solutions for Pc corresponding to ve dierent values
of $ are plotted in Fig. 6.4.

Fig. 6.4. The Peclet number Pc as obtained in (6.121) plotted for ve dierent
values of
. The relation of these curves to experimental data is discussed in more
detail in the text.

The case realized by most materials, namely $ > 0, obviously results in


reduced growth. For a real material such as e.g. succinonitrile with $ = 0.028,
at = 0.01, velocities are only 95.38% of the ones in the diusion limited
case and only 94.63% at = 0.7. The reason is that for s > l density
change ow drives a hydrodynamic ow towards the dendritic interface. This
hinders the transport of latent heat away from the dendrite and thus slows
down the solidication process.

120

6. Application of Diuse Interface Modeling


00000
11111

(a)

00000
11111
00000
11111
1111111111111111111111111111111111111111111111
0000000000000000000000000000000000000000000000
00000
11111
00000
11111
0000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111
00000
11111
00000
11111
0000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
0000000000000000 11111
1111111111111111
00000
0000000000000000
1111111111111111
00000
11111
0000000000000000
1111111111111111
00000
0000000000000000 11111
1111111111111111
00000
11111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111
0000000000000000
1111111111111111

(b)
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111
00000000000000000000000000000000000000000000000000

11111111111111111111111111111111111111111111111111
0000000000000000000000000000000000000000000000000
0
0000000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111111
0
1
0000000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111111
0
1
00000000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111111

1111111111111111111111111111111111111111111111
0000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111
0000000000000000000000000000000000000000000000
1111111111111111111111111111111111111111111111

Fig. 6.5. Comparison of a no-slip condition for (a) sharp versus a no-slip condition
for (b) diuse interface models. The term forcing a hydrodynamic ow as depicted
at the right enters as source term within the hydrodynamic equations of a diuse
interface model. It can be interpreted as a dissipative boundary force.

Moreover, the above results indicate that the eect of density change ow
on a dendrites tip evolution increases at low undercoolings. Again such a
dependence on the undercooling can be explained by the fact that at low undercoolings the temperature gradient as main driving force becomes smaller
and smaller. Thus it overrides hydrodynamic eects less and less. As a consequence one would expect to observe the eects of density change ow within
these parameter regimes at least in microgravity experiments, where buoyancy does not override density change ow. Indeed as indicated at the end
of Sect. 6.1 a dierence between the theoretical solution for purely diusion limited dendritic growth and data of microgravity experiments has been
reported previously. One well known experiment along this line is the one
by Glicksman et al. [129]. However, comparing the above theoretical results
including the eect of density change ow to the experimental data by Glicksman et al. one realizes that the dierence between theory and experiment is
even larger than for diusion limited dendritic growth theory (Fig. 6.2). This
result is found in numerical simulations of respective phase-eld model equations as well.
The precise model equations for the simulation results presented below
are
Heat Conduction Equation
T
+ (w) T
t

= (DT T )

L ()
t

Phase Field Equation

+ (w) = 2 2
t

+ V0 ( 3 ) + M0

1 2
(T 1 + L ())
(1 + 2 )2

Navier Stokes Equation


w
+ (w) w
t

1
p + 2 w + N S
l

6.3 Comparison to Experimental Data

121

Fig. 6.6. Simulation depicting four dierent stages of density change inuenced
growth for
= 0.3. The evolution runs from upper left to upper right, lower left and
lower right. The corresponding time steps of the numerical simulation are displayed
above each gure. Further details are given in the text.

Incompressibility Condition
w

1

2

.
and L () =
1 + 2

= 0 with N S =

Diusion advection type equations are implemented on the basis of an


ADI (alternating direction implicit) scheme [133]. The NavierStokes type

122

6. Application of Diuse Interface Modeling

equations are solved employing a MAC (Marker and Cell) solver with discretization of space coordinates as proposed in [137]. All simulations are
done on a quadratic grid. Numerical parameters for the simulation runs are
= 0.9, V0 = 0.4, = DT = 1, Re = 1500, t = 0.5, x = 0.3, N x = 1500
and N y = 1200, where N x and N y denote the grid size in x and ydirection
respectively. The initial condition is a parabolic interface of height 75 grid
units with the hydrodynamic velocity eld identical to zero in the solid phase.
Nevertheless after a few time steps hydrodynamic ow is encountered within
the solid phase and clearly visible in the graphs. In the literature it is known
as parasite current. It arises due to the fact that in the phase-eld simulation
the no-slip condition gets smeared over the diuse interface region (Fig. 6.5).
These parasite currents have been investigated in more detail by Lafaurie
et al. [195]. Applying their results allows me to choose appropriate parameter regimes, where the eect of these parasite currents remain bounded and
quantitative changes to the simulation results are negligible. The side branching activity in dendritic growth as visible in the simulation pictures is known
to result from thermal uctuations [22, 44, 180, 201, 234]. These have to be
included in the simulations by transforming
Ti,j Ti,j + R0 (ri,j

ri+1,j + ri1,j + ri,j+1 + ri,j1


)
4

(6.123)

with noise amplitude R0 and random numbers ri,j evenly distributed in the
interval [ 12 ; 12 ]. Taking into account the ri1,j1 as in (6.123) ensures local
energy conservation. The noise amplitude R0 corresponds to the magnitude
of thermal noise encountered in dendritic growth experiments, which reads
(in dimensional units):
2
kB TM
cp
.
(6.124)
2
2
L d0
This is equal to the mean square uctuation of T inside a two-dimensional
volume of d20 . For a comparison Fig. 6.6, Fig. 6.7 and Fig. 6.9 depict dendritic
evolution at dierent driving forces and noise levels, i.e. M0 = 4, M0 = 2.5
and M0 = 1, respectively. For M0 = 2.5 the evolution of the temperature eld
is displayed, as well, through colored contour plots in Fig 6.8. Simulations
were carried out keeping the position of the dendritic tip at approximatly
same height by repeated transformations of the x-coordinate (coordinate in
growth direction) of all eld variables. In any case the growth rate of the
dendritic tip is smaller than in the case of corresponding simulations without
coupling to the NavierStokes equations resulting in an increased deviation
from the experimental values by Glicksman et al.. Thus at this point one
clearly has to draw the conclusion that even taking into account surface
tension corrections within the energy condition, density change ow cannot
explain the dierence between experiment and theory.
Other physical mechanisms active in microgravity experiments must have
an even stronger inuence of reversed sign. Studying the literature one will

6.3 Comparison to Experimental Data


t=250100

t=250000

1000

1000

600

600

200

200

200

600
t=250200

200

1000

1000

1000

600

600

200

200

200

600

ydirection of space

123

1000

200

600
t=250300

1000

600

1000

y direction of space

Fig. 6.7. Four dierent stages of growth demonstrating the eect of a reduced
driving force with M0 = 2.5 and otherwise same parameters as in Fig. 6.6.

nd a couple of articles discussing possible mechanisms, as e.g. nite-size


eects through container walls [83] or nutation of the dendrite under microgravity [251]. Meanwhile, the experimentalists themselves called the experimental conditions within the space shuttle a reduced gravity environment (see
e.g. [194]) thus hinting at the possibility of some left-over buoyancy convection inuencing growth at low undercoolings. To my knowledge quantitative
studies of the possible eects of left-over buoyancy convection have not
been reported. Rather the recent focus of work by Glicksman et al. is concerned with uctuations of the dendritic tip velocity [84, 193, 194]. However,

124

6. Application of Diuse Interface Modeling


t=250000

t=250100

1000

1000

600

600

200

200

t=250200

t=250300

1000

1000

600

600

200

200

200

600

y direction of space

1000

200

600

y direction of space

1000

Fig. 6.8. The evolution of the temperature eld corresponding to the four dierent
stages of growth depicted in Fig. 6.7.

an explanation of the systematic deviation between diusion theory and the


experimental curve for 0.3 K receiving community wide acceptance has
not yet been given.
In this context a phenomenon, which to my knowledge has never been
studied in detail even though it has been assumed to inuence dendritic
growth at low undercoolings quantitatively [32], is the inuence of thermocapillary convection. It arises from a weak temperature dependence of surface
tension[248]. In dendritic growth the temperature along the interface is not
constant. Rather - taking into account a fourfold crystalline anisotropy - there

6.3 Comparison to Experimental Data


t=250100

t=250000

1000

1000

600

600

200

200

200

600
t=250200

200

1000

1000

1000

600

600

200

200

200

600

y direction of space

125

1000

200

600
t=250300

1000

600

1000

y direction of space

Fig. 6.9. Four dierent stages of growth demonstrating the eect of a further
reduction of driving force at M0 = 1.0 (otherwise same parameters as in Fig. 6.6).

are two coldest points to the left and to the right of the tip position as indicated in Fig. 6.10. Thus there is a temperature gradient along the dendritic
interface from the tip towards these points. This drives a thermocapillary ow
as depicted in Fig. 6.10. Comparing its direction to the one of ow driven by
density change (Fig. 6.6), one nds it to be reversed. Thus also its quantitative eect on the dendritic growth velocity should be reversed, i.e. it should
enhance growth compared to diusion limited dendritic growth. In that sense
it could be an explanation for the experimental data in [129].

126

6. Application of Diuse Interface Modeling

Fig. 6.10. The two coldest points of the parabolic interface are to the left and to
the right of the tip as indicated by the dark dots. If surface tension is temperature
dependent, the respective temperature gradient comes along with a surface tension
gradient, which is the origin of thermocapillary convection.

The problem is to fully grasp the quantitative eects of thermocapillary


ow on the evolution of a dendritic tip. From the point of view of sharp
interface modeling one would include it as a balance condition for pressure,
viscous ow stresses and surface tension:
(ps pl ) n + (u + (u)T ) n = 0 ,

(6.125)

where u + (u)T represents the viscous stress tensor for a liquid obeying
the incompressible NavierStokes equations. The surface tension term
can be expanded in a Taylor series truncated after its rst term, i.e.:
=

T .
T

(6.126)

Here T
is a material parameter, which is usually less than 0. Equation (6.125) yields a condition for the gradients of the ow velocity eld
normal to the interface. It occurs as an additional boundary condition at the
interface, which has to be solved together with (6.30) and (6.31).
Trying to apply a perturbative analysis as sketched in section 6.2 to model
equations (6.26)(6.38) with additional boundary condition (6.125) one will
nd that a solution, for which vorticity vanishes everywhere, is no longer
compatible with all boundary conditions. It is not obvious how to construct an
ansatz solving the NavierStokes equations with boundary conditions (6.30),
(6.31) and (6.125), which could serve as starting point for the subsequent
perturbative analysis.
A numerical investigation would be an alternative. However, in this case
it is not clear, whether the phase-eld method with its smearing of boundary
conditions over the diuse interface could handle the problem with necessary
precision. In particular the ndings of Lafaurie et al. [195] with respect to
parameter regimes, in which eects of parasite currents remain bounded,
cannot be carried over directly.
Preliminary results without thorough investigation of stability and convergence of the algorithm are depicted in Fig. 6.11 just to demonstrate qualitatively the eect of thermocapillary convection at a dendritic interface in

6.4 Summary
t=10000

127

t=12000

300

300

300

400

300

t=14000

400
t=16000

300

300

300

400

y direction of space

300

400

y direction of space

Fig. 6.11. Simulation depicting the inuence of thermocapillary convection at


a dendritic interface. Calculation were carried out at = 0.03 assuming quasiincompressibility and setting (u + (u)T ) n = within the interfacial region
(no noise included). Judged based on the velocity components depicted here the
dynamics of uid ow seems to be restricted to a small layer around the dendritic
interface. However the pressure eld displays a distribution which reaches far ahead
of the dendritic tip (see Fig. 6.11).

simulations. They are obtained by assuming quasi-incompressibility and setting (u + (u)T ) n = within the interfacial region. For the material

N
parameter T
= 0.028 mK
these preliminary numerical investigations indicate an increase of vn of as much as 14.7 % at = 0.03, clearly demonstrating
the eect required to explain the experiments9 . A more thorough investigation is in preparation [111].

6.4 Summary
Within this chapter a phase-eld model for hydrodynamically inuenced dendritic growth was employed to derive the sharp interface equations governing
9

This value T
is an estimate, since to my knowledge experimental measurements

for SCN are not yet available. However, it has been


of the precise value of T

N
measured for silicon as T
= 0.28 mK
[144]. On the other hand, the surface
tension itself is usually an order of magnitude smaller for organic materials than

for silicon [24]. Thus to get an estimate for T


applying to SCN I scaled the
value obtained for silicon with the ratio of surface tensions of the two materials.
The idea to proceed like this is due to Glicksman [131].

128

6. Application of Diuse Interface Modeling


t=16000

600

300

300

400

y direction of space

Fig. 6.12. The pressure eld around a dendrite growing inuenced by thermocapillary convection. This gure corresponds to the nal graph in the evolution of
Fig. 6.11.

density change ow. Compared to the classical approach towards dendritic


growth the resulting model equations contain an additional curvature term
in the energy boundary condition. The eects of such a term had been investigated by Lemieux et al. for diusion limited dendritic growth [205]. In that
case it results in an increase of selected velocities. In that sense including it
into an analysis of density change inuenced dendritic growth was accompanied by the motivation to nd an explanation for the experimental data
in [129]. The analysis itself was carried out as proposed originally by Kruskal
and Segur [191]. In the context of dendritic growth this method was employed
by various authors [236]. It starts from the derivation of a zero surface tension
solution. An analysis of this solution within the framework of regular perturbation expansion reveals that isotropic surface tension destroys the steady
state solutions found without surface tension. An irregular perturbation expansion allows me to analyze the problem taking into account crystalline
anisotropy. The selected velocity thereby obtained is (6.122). Its dierence
to the velocity selected in diusion limited dendritic growth is contained in
the terms depending on $, the density change between liquid and solid normalized with respect to l . Eects of the additional curvature term obtained
via the asymptotic analysis of Sect. 6.1 are of higher order in the expansion.
Thus other than for pure diusion limited dendritic growth, where exactly
this curvature term gives rise to increased values of selected velocities [205],
within the mathematical treatment underlying (6.90)(6.121) it cannot be
employed to explain larger growth velocities than for diusion limited den-

6.5 Appendix

129

dritic growth. Thus the contradiction to the experimental data obtained in


the experiments of [129] cannot be solved at this point. To explain the latter
experiments other possible inuences as indicated in Sect. 6.3 remain to be
investigated and understood quantitatively.

6.5 Appendix
Appendix A: Transformation to Parabolic Coordinates
Interface

s(, , t) = s (, t) = 0

(6.127)

Unit Normal and Tangent Vector



 
2

s e + e
+

2
2
(s ) +
|s| =
+
 

s
1

n=
s e + e
=
|s|
2
(s ) +



1
e + s e
t= 
2
(s ) +
s =

Normal Velocity
vn=
Curvature

|s|


s ()
1
ys () = (s () )
2

xs () =

[s ()] = 

s
t

(6.128)
(6.129)
(6.130)
(6.131)

(6.132)

(6.133)
(6.134)

ys xs xs ys


(s s )2 (1 s ) 2( + s )s s
=
3
3
3

( + s ) 2 (s + (s )2 ) 2
2
2 2
(ys ) + (xs )
(6.135)

130

6. Application of Diuse Interface Modeling

Anisotropy

s + s

s + (s )2 + s

(ns 1)

sin = t ex =
s + (s )2 + s

cos = n ex =

cos(4) = 8 sin2 cos2 + 1 = 8


A[s ()] = 1 +

(6.136)
(6.137)

s (s 1)(s + s )

( + s )2 ((s )2 + s )

8s (s 1)(s + s )
2

( + s )2 ((s )2 + s )

+1

(6.138)
(6.139)

7. Application to Epitaxial Growth


Involving Elasticity

In Sect. 5.2.3 the phenomenon of epitaxial surface growth has been introduced as an application of an asymptotic analysis in the thin interface limit.
It was demonstrated, how a set of phase-eld model equations can be developed from two thermodynamic potentials, namely the inner energy functional
E and the entropy functional S. Thereby an additional degree of freedom is
obtained. This additional degree of freedom allows me to recover the precise
attachment kinetics of the two-sided epitaxial growth problem involving unequal attachment kinetics and discontinuities at the interface asymptotically
at rst order (thin-interface limit). The sharp interface equations constituting
this two-sided epitaxial growth problem are given by (5.68)(5.71).1 Moreover, the corresponding phase-eld model equations are (5.80) and (5.81).
For further details the reader is referred to Sect. 5.2.3.
The scope of the succeeding two sections is to exploit the variational priciples of irreversible thermodynamics to extend the above phase-eld model
equations to elastic eects at strained surfaces [110]. In this manner the
phenomenon of epitaxial growth at strained surfaces becomes numerically
tractable. On the other hand, trying to extend the above sharp interface
equations to elastic driving forces results in boundary conditions, which in
their general formulation cannot be solved, as will be demonstrated in more
detail in Sect. 7.1. This possibility to obtain tractable models where the sharp
interface approach fails can be viewed as one of the advantages of phase-eld
modeling justifying the development of this method. I will come back to this
point in Chap. 8.

Note that in the case of epitaxial growth anisotropy due to the underlying crystal
structure of the material is contained in eq = eq
0 (1 cos4). Detailed studies
of the inuence of crystalline anisotropy in epitaxial growth can be found in
[106, 109].

Heike Emmerich (Ed.): LNP m73, pp. 131140, 2003.


c Springer-Verlag Berlin Heidelberg 2003


132

7. Application to Epitaxial Growth Involving Elasticity

7.1 Elastic Driving Forces Within the Diuse Interface


Approach: Extended Model Equations for Strained
Surfaces
If an epitaxial surface consists of strained layers, the correct free energy to
describe the system has to include the respective elastic strain energy. Thus
(5.79) changes into:
f (, , uij ) =

VE + VS

( eq )
k() ln( eq ) + + f ()el . (7.1)
d() +
2
eq

Here fel is a free energy density, which takes into account the global elastic
energy in both phases. A natural starting point to construct a suitable form
of fel for the phase-eld approach is the relation between the stress tensor
ij and the strain tensor uij as given by Hookes law:
f = uij uij +

2
u ,
2 ii

(7.2)

where summation over double subscripts is implied. and are the Lame
constants. For plane strain, these elastic constants are related to Youngs
modulus E and the Poisson ratio via = E/[2(1 + )] and = E/[(1 +
)(1 2)]. f the free energy per unit volume of a general system, of which
the only energy contribution is elasticity. The stress tensor ij is then given
by:
ij =

ij
f
= 2uij + ukk ij = 2(uij ukk ) + Kukk ij ,
d
uij

(7.3)

where K = + 2/d is the bulk modulus and d referring to the dimension for
which the problem is solved (i.e. d = 2 or d = 3). If there is a mist on the
surface of the substrate, additional mist strain arises. Following [261] this
m
additional strain can be modeled through an additional contribution ij
:


1+
m
(7.4)
= 2m
ij
ij ,
1 2
where m quanties the mist. It is then straightforward to formulate the
following ansatz for the -dependence of f :



1+

fel (, {uij }) = k() uij uij + u2ii 2m


2
1 2

2
+ [1 k()] uii .
(7.5)
2
Here
is the bulk modulus of the non-solid phase, i.e. liquid or vapor depending on the employed epitaxial technique. As a result the governing equations

7.1 Elastic Driving Forces Within the Diuse Interface Approach

133

of the phase-eld model given in an adiabatic approximation, i.e. Fel /ui = 0


(for implications of this approximation see the paragraph following (7.7)),
now read:

eq 
VE + VS 

d ()
k ()
= ( + )2
2
eq
t


1+

2

m
],
(uii ) 2
+k ()[uij uij +
1 2
2

= D2 + G()

+F + .
t

(7.6)

(7.7)

These have to be solved together with the equations for the evolution of
the elastic variables, where the strains uij (i, j = 1, ..., d) are dependent
quantities. Therefore, the variational derivatives F/uij are dependent, as
well. To proceed one may exploit the fact that the components ui (i, j =
1, ..., d) of the displacements are independent variables. Assuming that the
relevant time scales of the problem are large compared to sound propagation
times, the variational derivatives Fel /ui of the elastic free energy can be
assumed to be zero, which is exactly the adiabaticity assumption. Hence one
obtains
0=

Fel
Fel
=
{k()ij [1 k()]ukk ij } .
=
xj
ui
xj uij

(7.8)

The second term can be related to the pressure of the system via
p = p0
ukk ,

(7.9)

where p0 has been chosen as equilibrium pressure in the liquid phase. This
term contributes only in the liquid phase. Assuming that during the experiment on the liquid side deviations from this equilibrium pressure remain
small, (7.8) can be simplied as follows:



0=
k()ij .
(7.10)
xj
Equations (7.6), (7.7) and (7.10) become closed equations by replacing ij
and uij by the eld variables ui using the denition of the strain tensor,


1 ui
uj
uij =
,
(7.11)
+
2 xj
xi
and Hookes law. Its applicability is demonstrated in the following section.
Again an ADI scheme is employed for the diusion-advection type equations
(7.6) and (7.7). The elastic problem, on the other hand, is solved via successive
overrelaxation, where the time integration is performed by a formally secondorder accurate midpoint scheme [220].

134

7. Application to Epitaxial Growth Involving Elasticity

Turning towards the sharp interface model given by (5.68)(5.71) accounting for elastic eects requires a coupling to the following evolution equations
for the displacements of the solid phase:
(1 2)k2 ui + i k uk = 0 .

(7.12)

To evaluate these equations, boundary conditions of the displacement eld at


the moving, phase-separating interface have to be determined at every time
step of the calculation. The correct physical boundary condition is a balance
of the force acting on the lm surface and the pressure in the liquid (vapor)
phase, i.e.:
ij nj = pl/v |ij .
(7.13)
Obviously in this case the construction of the correct boundary values for the
displacements uij based on (7.13) is less straightforward than the formulation
of the phase-eld equations (7.6)(7.11). A more detailed discussion of this
point can be found for the special case of an assumed zero-pressure condition
above the substrate in [261].

7.2 Comparing Simulations and Experiments


For strained surfaces grown by epitaxy two qualitatively dierent morphologies are known to result from the epitaxial process, namely
1. rippling of continuous surface layers,
2. a morphology made up of an array of islands.
Based on the experimental observations reported in [6, 86, 87] ripples were
associated with layers of a low mist surface. On the other hand, islands were
understood to be the dominating morphology in high mist regimes [94, 97,
204, 218]. In contrast to this, Dorsch et al. [95] found a morphology transition
towards islands even for low mist surfaces. Their experiments were done for
strained Gex Si1x grown from In solution at low driving forces. Dorsch et
al. claimed that their ndings coincide with theoretical work by Gao [123],
indicating, that in all mist regimes island morphologies are the ones to result
in maximum strain energy relaxation. Following that point of view one would
expect islands to dominate the morphological surface evolution independent
of the mist, once growth follows essentially energetics. Accordingly epitaxial
growth at strained surfaces could be viewed as a two stage process:
1. During a rst stage of growth a rippled morphology with well dened
initial wavelengths arises.
2. At a subsequent stage of growth evolution is dominated more and more
by the principle of maximum strain energy relaxation, which drives the
system into a transition towards an island morphology. This transition is

7.2 Comparing Simulations and Experiments

135

independent of the surface miscut. It is an expression of the rippled morphology itself, the geometry of which results in a redistribution of strain.
In turn this changes the surface potential such that it favors the strain
relaxation mechanism. During this stage of growth the island morphology
essentially inherits the wavelength of the initial ripple instability.
Thus a transition to islands should occur independent of the mist at advanced stages of growth. A necessary presupposition is diusion in the nourishing phase (liquid or vapor, depending on the epitaxial technique) as material transport mechanism. Previous to the experiments by Dorsch et al. a ripple morphology was usually observed resulting from MBE (molecular beam
epitaxy) experiments. Compared to LPE (liquid phase epitaxy) the diusion coecients encountered in MBE are four orders of magnitude smaller.
This leads to the conjecture that during MBE diusion processes are simply
not fast enough to trigger the morphology transition towards islands before
the end of the experiment. In that sense the ripple morphologies observed
in MBE experiments would have to be considered transiental morphologies.
Moreover, pseudomorphic islands were the true stationary growth mode to
be expected from any kind of epitaxial setting at a strained surface.
From point of view of a nonlinear analysis carried out for strained epitaxial
surfaces by Spencer et al. [261] this is not necessarily the case. Their analysis
revealed two dierent steady state solutions within the near critical parameter
regime:
solution 1: a spatially periodic small amplitude cusp like solution
solution 2: a spatially periodic large amplitude solution of sinusoidal
shape.
Both solutions were found to be unstable, leaving open the question, whether
the evolution of the strained surface would be steady outside the realm of
long-wave theory or characterized by a transient state displaying coarsening.
To reconcile these ndings with the experimental observations described
above, one would have to identify the island morphology as the steady state
solution outside the realm of long-wave theory, which develops eventually as
secondary instability from solution 1 as well as from solution 2. However, the
question remains: Which of the two qualitatively dierent types of primary
solutions 1 or 2 are the ripples observed initially in the experiments by Dorsch
et al.? Are they spatially periodic nite amplitude rounded cusp solutions
or rather near critical spatially periodic small amplitude solutions? If either of them is true, might the morphology observed in MBE experiments
actually be the other kind of solution? And if so, might the secondary instability of MBE growth be the second kind of scenario depicted above, namely
a transient state displaying coarsening? If this were true, it could still be reasonable not to observe the coarsening process due to slow diusion as already
pointed out by Albrecht et al. [95]. Nevertheless the precise understanding
of the dynamics of strained surface epitaxial growth would be dierent - a

136

7. Application to Epitaxial Growth Involving Elasticity

dierence which might become important to predict the correct morphology


for dierent parameter settings. Thinking in this direction one could for example wonder, whether high driving forces and/or high temperatures might
result in a transition towards islands even for MBE growth.
Table 7.1. Amplitudes and cell widths for varying diusivities D, m = 0.05
D
Amplitude
Cell width

1
73
6

102
59
5

104
45
4

106
13
1

To answer the questions above a full simulation of model equations (7.6)


(7.11) is a valuable tool. Note that for the simulations reported in this chapter
the terms F + (see (7.7)) are discarded. This refers to the situation of no
desorption, which is true for MBE experiments as well as for LPE. Moreover,
to model LPE growth as employed in the experiments by Dorsch et al., one
has to keep in mind that the experimental conditions are chosen such that the
solutal eld is approximately at equilibrium. The true driving force is rather
the temperature eld T , which takes the place of the variable in (7.6) and
(7.7). This exchange of and T is possible, since under the assumption of
local thermodynamic equilibrium the phase diagram allows us to relate the
two to each other. Approximating solidus line and liquidus line of the diagram
as straight lines (see Fig. 2.4) the relevant relations read:
 S
 dC 
S
 (T Tm )
C = 
(7.14)
dT 
 L
 dC 
 (T Tm ) = 1 (T Tm ) ,
C L = 
(7.15)
dT 
ml
where C S is the concentration of solute in the solid phase, C L its concentration in the liquid phase, Tm the melting temperature and ml the slope of
the liquidus line, which is usually negative. In this way the model allows us
to investigate the evolution of strained MBE as well as LPE surface growth
taking into account all of the nonlinear modes and thus to identify steady
state solutions outside the realm of long-wave theory.
Table 7.2. Amplitudes and cell widths for varying mists m , D = 1
m
Amplitude
Cell width

0.05
73
6

0.01
34
3

0.12
21
2

0.15
9
1

7.2 Comparing Simulations and Experiments

137

In particular it enables an investigation of the amplitude and the shape of the


evolving morphology (sinusoidal versus cusp like) depending on the diusion
constant as well as the mist angle. The results are summarized in the tables.
In Table 7.1 the mist m is kept xed, whereas D varies. In Table 7.2, on
the other hand, D is kept xed and m varies. The cell widths reported
in the two tables are measured at one half of the amplitude height. The
remaining parameters of our investigations read x = y = 0.005, t =
0.0018, 2 = 1, = 1, = 0.7417, = 0.0021, = 1.4371, = 0.3212,
= 0.6666,
= 1 and = 0.3341. Again all parameters are scaled with
respect to x and t. Note that within the context of this work we identify
a morphology to be cusp like if its rst derivative does not display a turning
point between two extrema2 (Fig. 7.1). Obviously, the higher the mist m
the lower the amplitude of the primary instability. This coincides with the
experimental ndings of Dorsch et al. just as well as with theoretical work
by Srolovitz [262]. The diusion constant D displays the same qualitative
behavior. Moreover, the unproportionally large drop in amplitude for D below
104 indicates that the transition to a dierent type of solution occurs. Indeed
the small amplitude solutions found by simulation of (7.6)(7.10) for D =
106 display a cusp like shape with considerably smaller widths of the cell
shaped instabilities. Thereby they dier from the solutions for D = 1 to
D = 104 . Examples for both solutions are given in Fig. 7.1.
Taking into account that low diusion constants belong to MBE growth,
whereas large diusion constants indicate use of the LPE technique, one important result of this numerical investigation is that for equal mists the
primary ripple like instabilities of MBE and LPE dier: rst, amplitudes
of LPE instabilities are larger; second, their tips are rounder. On the other
hand, tips of MBE instabilities display a cusp like shape. The question which
remains is whether these dierent solutions display dierences with respect
to the long time evolution as well. Figure 7.1 presents a view on the temporal
evolution of the respective solutions. Four subsequent stages of growth are
given with intervals of 104 time steps in between. In Fig. 7.1b a transition
of ripples to islands indicated by negative growth values in the valleys of the
morphology becomes visible. In contrast valleys of the cusp like instabilities
keep positive growth rate. After t = 6 104 time steps their shape is approximately stationary. Signs of coarsening cannot be detected. These numerical
ndings are indeed characteristic for the long time behavior of cusp like solutions versus large amplitude solutions, i.e. they were obtained for other sets
2

A fully cusped tip could not be resolved via phase-eld simulations, for which the
interface width l puts an inherent length scale cut o. l has to be resolved by 8-10
grid points to avoid numerical metastability [38]. The results reported here were
obtained for four dierent values of l with dierent mesh sizes to fulll the above
criterion. They are independent of this change of numerical parameters. Thus in
the investigated parameter regimes no fully cusped solutions can be expected.

138

7. Application to Epitaxial Growth Involving Elasticity


(a)

(b)

0
0

60

60

50

50

40

40

30

30

20

20

10

10

10

20

30

40

50

60

70

y direction of space

80

90

100

10

20

30

40

50

60

70

80

90

y direction of space

Fig. 7.1. Contour plots of surface morphologies at t = 3 104 (turquois line),


t = 4104 (red line), t = 5104 (green line) and t = 6104 (blue line) for  = 0.15: (a)
evolution of the cusp like morphology, D = 106 , (b) large amplitude morphology,
D = 1. The pink lines depict the local slopes of the contours at t = 6 104 (blue
line). The derivatives are generated from the data points of the contour plots. Note
that they are aligned with respect to a second zero point on the vertical axis.

of diusion constants and mists, as well. To elucidate these ndings further,


Fig. 7.2 displays the evolution of the principal component of the strain tensor in growth direction. The dierent strain elds arise as an answer to the
dierent evolutions of the two respective surfaces and in turn inuence their
further growth as well. I.e., surface regions subject to low strain have high
growth rates, whereas the regions exposed to large strain exhibit slow growth
or even recede. In this sense in Fig. 7.1b enhanced amplitude growth of the
morphology can be attributed to the large dierence of strain values around
its crest versus the one around its valleys. For the cusp like solutions, on the
other hand, strain is distributed more evenly over the surface.
Together these numerical studies indicate that even though MBE and LPE
experiments at strained surfaces can both result in similar rippled surface
morphologies, these ripples are not exactly the same type of solution of the
growth problem. While the ripples found for low diusion constants, i.e. MBE
growth conditions, are cusp like instabilities, the ripples found in LPE are
large amplitude solutions with rounded tips. In the long time evolution the
latter will undergo a transition towards islands. Cusp like solutions do not
display such a transition even after simulation of time intervals much longer
than any experiment. However, there is no sign of coarsening either. The latter
nding can be compared to recent numerical studies of a strained surface
in contact with its melt, but not subject of epitaxy [172]. Results of these
studies indicate that a primary instability of periodic grooves will undergo

7.2 Comparing Simulations and Experiments


(a)

139

(b)

Fig. 7.2. Evolution of the principal component of the strain tensor in growth
direction: (a) cusp like morphology, (b) large amplitude morphology. The time
steps, for which the elds are displayed, coincide with the ones for the contour
plots of Fig. 7.1, i.e. from top to bottom: t = 4.5 104 , t = 5 104 , t = 5.5 104 and
t = 6 104 .

140

7. Application to Epitaxial Growth Involving Elasticity

coarsening only in the absence of gravity. Once gravity is taken into account,
several stationary grooves coexist within the nal surface morphology. Within
(7.6) the last term on the right-hand side modeling the surface miscut could
be interpreted as having the same inuence as gravity. In that sense it would
stabilize the cusp like rippled morphology in a way, such that coarsening does
not occur.

7.3 Discussion
Here epitaxial growth at strained surfaces was chosen as a second example to
demonstrate the convenient applicability of the phase-eld approach when it
comes to the point of including additional driving forces into an interfacial
growth problem. Since such additional driving forces can very easily be related
to an additional term in the systems free energy density functional, they
directly enter the variational principles on which the derivation of phase-eld
model equations is based upon.
In this manner one can exploit the framework of irreversible thermodynamics to extend the model equations of diusion limited epitaxial growth
as derived in Sect. 5.2 to stress induced surface eects. The respective model
equations have been developed in detail in Sect. 7.1.
In Sect. 7.2 they have been applied to experiments by Dorsch et al. [95].
In this context their simulation sheds new light on the morphology transition from ripples to islands. In particular the numerical investigations reveal
that the ripple instability occurring in MBE has to be distinguished from the
one in LPE growth. It seems to fall into the class of cusp like instabilities.
The non-appearance of a morphology transition towards islands in principle
hints at the possibility that the true secondary instability belonging to this
kind of cusp like solution is a transiental coarsening regime. The reason that
the coarsening is depressed, can be explained by comparison to the numerical studies in [172], where the authors report the absence of coarsening in a
similar situation, if gravity eects are taken into account. It seems plausible
that within the model equations derived here, the mist term has the same
stabilizing eect as gravity in [172]. These ndings reconcile with the interpretation given by Dorsch et al. [95] in the sense that the formal reason for
the morphological dierences of MBE compared to LPE instabilities is given
by the large ratio of the respective diusion constants. As a consequence a
morphology of cusp like ripples should be the one to be expected in MBE
experiments even at high driving forces and high temperatures after extended
deposition time a point, which certainly could be claried by experiments.
If it were true, MBE would generally fail as a technique to tailor regular
island arrays with specied island distance as functional devices.

8. Conclusions and Perspectives

Concluding the text at this point can only result in a partial conclusion:
Research in the eld of diuse interface modeling is still young and fully
active. It is likely, that while I am writing, further progress of considerable
impact is achieved. As discussed progress could basically be twofold, either
concerning modeling related issues or numerical ones.
With respect to the latter the numerical appendix of this book summarizes the most recent computational approach, which achieved outstanding
computational eciency for the case of 3D dendritic growth involving uid
ow [155] on the basis of a semi implicit approximated projection method,
implemented fully adaptive and parallelized. It is an example of a growth phenomenon, which certainly cannot be treated numerically without employing
enhanced techniques. However, for this kind of moving boundary problems an
implementation making use of enhanced numerical techniques is possible only
based on diuse interface modeling. The reason is as pointed out in Chap. 2
that only the diuse interface approach allows for an Euler scheme1 . Thus
one conclusion to draw at this point is, that a diuse interface approach for
moving boundary problems becomes necessary, if one intends to apply the full
power of high performance computing, which today is essentially connected
to parallelization, to their simulation.
With respect to modeling-related progress, the concept of thermodynamic
consistency was described in detail in Chaps. 35 as a most important underlying viewpoint. It was explained, how a thermodynamically consistent
diuse interface model can be obtained by the variation of underlying thermodynamic potentials. Thermodynamic consistency itself can be interpreted
as a validation of the model. Consequently one can go on to examine the
derived model equations with the intention to understand more about the
behavior of physical elds in the region of the moving, phase separating interface. The mathematical method allowing for the respective analysis, is the
1

As discussed in Chap. 2 a level set approach would basically yield an alternative


for the implementation of a moving boundary employing an Euler scheme, as
well. However, from point of view of numerics such a level set approach imposes
diculties a diuse interface approach does not (see Chap. 2). Moreover the
level set approach is lacking the thermodynamic background of diuse interface
models and thus qualies only as a numerical tool for classically known moving
boundary problems.

Heike Emmerich (Ed.): LNP m73, pp. 141144, 2003.


c Springer-Verlag Berlin Heidelberg 2003


142

8. Conclusions and Perspectives

method of matched asymptotic expansion described in Chap. 5. Applications


of this concept of diuse interface modeling given in the text are twofold:
The rst is devoted to hydrodynamically driven dendritic growth (Chap. 6).
This example was chosen to demonstrate, how the asymptotic analysis of a
diuse interface model can be employed to revise a traditional sharp interface formulation of a moving boundary problem. Subsequent analysis of the
specied model equations via a perturbation expansion around a zero surface
tension steady state solution yields new insight in the dynamics of dendrites
growing at low undercoolings. In particular it sheds new light on the well
known and still not fully explained experiments [129], in which Glicksman et
al. carefully determined the dierence between dendritic tip velocities in microgravity experiments and diusion limited growth theory. In this context
it is the achievement of the diuse interface approach to allow for a rigorous
derivation of a specied set of model equations revising the relation between
experiment and theory.
A second example was presented in Chap. 7 to illustrate the usefulness of
diuse interface modeling with respect to a dierent class of moving boundary problems: As demonstrated for the case of elastically inuenced epitaxial
surface growth, sharp interface formulations of moving boundary problems
involving more complex transport than just diusive one might yield boundary conditions, which numerically are no longer tractable. Here the diuse
interface approach oers the great advantage of overcoming the need to solve
boundary conditions at the phase separating interface. Thus for this type
of problems one can conclude, that to nd a numerical solution at all, it is
unavoidable to turn towards diuse interface modeling.
These examples combined with the theoretical background given in the
rst ve chapters are intended to provide the reader with an idea of what
diuse interface modeling is all about, for what classes of problems it is useful
and what kind of knowledge it allows us to derive based on a combination
of thermodynamic modeling, mathematical analysis and simulation. If interested in the eld for the purpose of an own application, a rst step would
be to formulate the underlying free energy or entropy functional as basis
for the variational approach, by which the relevant model equations are to
be obtained (Chap. 4). Todays applications range from dendritic growth to
electrodeposition, from image processing to superconducting materials, from
alloy solidication to polymer systems. Certainly depending on the focus of
ones interest the complications when applying diuse interface modeling to
any of these phenomena are more but just one step beyond diusion limited
dendritic growth, from which I started my discussion taking it as paradigmatic problem in Chap. 2. Therefore it seems appropriate to reect for a
moment on what really has been achieved in the eld of diuse interface
modeling and what problems one would judge to be in reach based on the
knowledge already gained.

8. Conclusions and Perspectives

143

In the following I will restrict the discussion of what has been achieved to
the phenomenon of crystallization from the melt, which is directly linked to
the paradigmatic dendritic growth problem. Thus it provides a representative
overview of what one has to meet when turning from pure diusion limited
growth to further complications originating from additional transport elds2 ,
or from anisotropic kinetics and energetics in the interfacial region.
In the context of crystallization diuse interface modeling refers back to
Cahn [56]. His approach was taken up again several years later by Halperin
et al. [143], Langer [200], Fix [120] and Collins and Levine [74], which
then started a rapid development of diuse interface modeling for crystal
growth phenomena. Caginalp [49, 50] and others demonstrated the relation
of these approaches to classical sharp interface models by means of asymptotics and distinguished limits. For suciently thin diuse interfaces the
GibbsThomson boundary condition was proven to be incorporated in the
diuse interface approaches [52]. First derivations from an entropy functional
were due to Penrose and Fife [231]. This came along with an awareness of
the question of thermodynamic consistency of such models. In this context
Wang et al. [289] introduced the concept of local positive entropy production and non-classical uxes. At the same time Kobayashi [186, 187] demonstrated the relevance of diuse interface modeling for real computations by
simulating rather complicated solidication patterns. Since then many others [182, 239, 240, 290, 297] have used the diuse interface approach successfully to study solidication morphologies. Great advance in the eld was
achieved through the introduction of the thin interface limit by Karma and
Rappel [168]. It lessened the constraints with respect to the thickness of
the interfacial region and thus rendered models computationally more efcient. The authors themselves used the increased capability of diuse interface modeling gained from the thin interface limit amongst other things
to study side branching eects in dendritic growth [168, 169]. Moreover the
diuse interface approach could be extended to study alloy solidication,
i.e. a growth problem with coupled transport elds. Respective models were
formulated by Wheeler et al. [295, 296], Bi et al. [252] and Charach and
Fife [67, 68]. These could be employed to study the phenomenon of solute
trapping [69, 76, 77, 296], dependence of surface tension on composition [69],
isothermal dendritic growth and microsegregation [75, 291], Ostwald ripening and coalescence [292], recoalescence during dendritic solidication [35],
cell to front phase transitions during directional solidication [36] and eutectic alloy solidication [96, 100, 101, 138, 166, 298]. A further issue was
to blend hydrodynamics with diuse interface modeling. Such models have
been developed [26, 51, 272] and used to some extent in computation [85, 108].
Moreover it was shown, that with an additional potential modeling van der
Waals and polarization forces, the models can be employed to simulate wetting phenomena, as well [221]. Only recently thermodynamically consistent
2

which might not be continuous at the interface

144

8. Conclusions and Perspectives

diuse interface models including hydrodynamics could be formulated for


pure materials as well as for binary alloys by Anderson et al. [12, 13]. The
same authors also developed a model including convection and anisotropy for
a pure material [15, 17]. This model was subsequently extended by Sekerka
and Bi [250] to multicomponent alloy solidication. In the area of solidsolid
transformation Braun et al. [39, 40] have used a diuse interface eld model
with multiple order parameters to explore orderdisorder phase boundaries
and transformation in alloys. Moreover, Kobayashi et al. [188] have developed a phase-eld model involving two order parameters, expressed in terms
of polar coordinates, that facilitate modeling of polycrystalline materials with
grain boundaries. Finally Karma [171] was able to solve the problem of twosided dendritic growth. Thus a wide range of algorithmic means to tackle
complexities arising from coupled transport elds as well as kinetic and energetic anisotropies in the interfacial region are available from this research
in the eld of crystallization by now.
Turning from the discussion of What has been achieved? towards the
question of What can be achieved? it seems natural to point out, that there
are still questions to investigate for the phenomena of electrodeposition, of
image processing, of polymer systems and so on, where one would simply have
to carry over steps taken in the eld of crystallization from the melt. However, there is a further complexity tied to some of these phenomena, namely
an inherent multi-scale nature of the problem due to an internal structure
of the material. Examples for such materials with inner structure are liquid
crystals and polymers. For these kind of systems the diuse interface approach opens an additional interesting perspective, since it can be used as
basis of a rigorous mathematical multi-scale approach. Thus the sentence of
the introductory chapter One might wonder if in the end this approach can
provide a framework to tackle the behavior of still more complicated systems,
e.g. systems with an inherent multi-scale nature due to an internal structure,
such as liquid crystals or polymer solutions, as well. is actually an anticipated perspective at this point. There is a clear methodological framework to
tackle the multi-scale problem, namely homogenization. Examples of how homogenization can be applied to diuse interface modeling are [209] and [98].
Further steps to treat phase transitions in materials with inner microstructure in a thermodynamically consistent manner are provided by Kalospiros et
al. [159, 160] in the eld of polymer growth. Now entering the Greek culture
allows me to close these remarks unscientically by turning towards Plato:
And with respect to the other points I would not dare to adhere to my view
with great insistence, but ... I ... believe that we must try to nd what is not
known, for we should be ... less idle than if we believed that what we do not
know it is impossible to nd out and that we need not even try.
Plato words from the mouth of Socrates, Menon 86C [243]

A. Numerical Issues
of Diuse Interface Modeling

This appendix is concerned with the questions: What happens after a diuse
interface model has been established? How straightforward is its implementation? Is there anything one has to take special care of with respect to that
implementation? The answer to the latter is a clear yes. Answering it in
more detail will leave the question of the previous sections whether or not a
diuse interface model under consideration is thermodynamically consistent
behind. Rather one is concerned with a choice of model parameters which
will yield the best numerical performance. For dendritic growth it could be
proven that from point of view of run time performance models which do not
display thermodynamic consistency and can only be validated by the method
of matched asymptotics expansion can be superior to thermodynamic consistent ones [182]. In that sense referring to the implementation point of view
one nds once more1 the expression universality of diuse interface models
in the previous citation.
Since the early attempts of phase-eld computation in the context of
solidication by Smith [259] great advances in the accuracy of those computations have been encountered. The reasons are twofold: First, computer
technology has advanced signicantly and thus allows us to integrate the unsteady phase-eld equations for increasingly complex congurations. Second,
there has also been progress in the numerical techniques applied to the implementation of phase-eld models. This involves choosing an optimal set of
numerical parameters and model potentials just as well as considering enhanced numerical techniques as parallel, adaptive or multi-grid algorithms.
The rst of the above two issues arises since
1. dierent potentials can be employed, resulting in dierent phase-eld
models, which nevertheless describe the same growth phenomenon in
their sharp interface limit and
2. to establish the correspondence between the diuse interface model and
the sharp interface model there is some freedom in how to arrange the
numerical parameters of the diuse interface, i.e. a physical variable ap1

once more refers to the discussion at the end of Chap. 5, where the notion
of universality in the context of phase-eld modeling has been used to denote a
freedom in the precise choice of F .

Heike Emmerich (Ed.): LNP m73, pp. 145163, 2003.


c Springer-Verlag Berlin Heidelberg 2003


146

A. Numerical Issues of Diuse Interface Modeling

pearing in the sharp interface model is usually expressed by more than


one numerical parameter of the diuse interface model.
I will demonstrate the above in detail in Sect. A.1 and Sect. A.2 for the case
of dendritic growth, for which it has received much attention by the scientic
community over the last decade, e.g. [113, 182, 232]. Whereas Sect. A.1 is
devoted to the relationship between physical variables and numerical parameters, thereby pointing out the freedom in the precise choice of model parameters, Sect. A.2 compares the computational eciency of dierent phase-eld
models depending on the choice of the underlying themodynamic potential.
In Sect. A.3 I will proceed to discuss some state of the art numerical
approaches to interfacial growth phenomena and thereby give a look at how
advanced numerical techniques can be employed to speed up diuse interface
computation a step which is certainly required to deal with more complex,
challenging growth conditions.

A.1 Relationship Between Physical Variables


and Numerical Parameters
With respect to phase-eld computations in dendritic solidication Kobayashi [187] was the rst to become famous for quite beautiful animations of
3D phase-eld simulations of evolving dendrites. However, this work was still
very qualitative. To use phase-eld simulation for quantitative predictions of
dendritic growth morphologies one has to take a closer look at the relationship
between the physical variables of the sharp interface limit and the numerical
parameters of the phase-eld model equations.
For diusion limited dendritic growth the GibbsThomson relation, the
Stefan condition and the diusion equation for the transport of heat in the
two bulk phases constitute the free boundary problem, which describes the
evolution of the dendritic front. The respective model equations (2.5)(2.7)
are rewritten here by transforming to a dimensionless temperature T via
T =


cp (T  T
)
.
L

(A.1)

This transformation employs the specic heat of the solid given by cp . More
over, T
refers to the dimensional temperature of the undercooled melt far
from the interface. In addition one can introduce a dimensionless supercooling
=



cp (TM
T
)
L

(A.2)

of the liquid phase which is normalized, so that for = 1 the solidication of


one volume element of the liquid phase results in precisely the heat required to


heat up the same volume element from T
to TM
. With these transformations
the set of dimensionless equations corresponding to (2.5)(2.7) reads:

A.1 Relationship Between Physical Variables and Numerical Parameters

T
= D2 T
t
T = d0 ()
v n = D [(T )s (T )l ] n .

147

(A.3)
(A.4)
(A.5)

For large driving forces the consideration of local thermodynamic equilibrium is not valid any longer. This is taken into account by introducing a
kinetic term kin vn into the GibbsThomson relation and thus replacing
the dimensionless equation (A.4) by:
T = d0 () kin v n .

(A.6)

Here kin denotes a kinetic coecient, which is a material parameter (compare to (2.10)). d0 () is the capillary length given by (2.12).
For quantitative application of a derived phase-eld model to a concrete
dendritic growth problem, the capillary length d0 and the kinetic coecient
kin of (A.6) have to be related to the numerical parameters of the phaseeld model. To demonstrate how this can be achieved, I introduce a concrete phase-eld model for dendritic growth of a binary alloy by specifying
G(, X, r) in the remainder of this section. Following [2] I choose
2
4
2
()2 + V0 (

)
(A.7)
2 
4
2

Q2
C2
C
+ Q + 1 + L ()(1 C Q) ,
+ M0
2
2

G(, Q, C) =

with

.
(A.8)
+
2 1 + 2
Here C denotes the solutal eld. To overcome the problems related to phaseeld modeling based on the dimensionless temperature T (see Sect. 4.2) the
dimensionless energy density Q is used instead. It is related to T by

T
: solid phase
Q=
(A.9)
T + 1 : liquid phase .
L

With this choice of G and after retransformation from Q to T the set of


phase-eld model equations I discuss here reads



1 2
2 2
3
, (A.10)

T+
= + V0 ( ) + M0
t
(1 + 2 )2
1+
C
(A.11)
= (DC (C L ())) ,
t
L ()
T
= (DT T )
.
(A.12)
t
t
In (A.10)(A.12), it is the set of parameters , V0 , and M0 for which
relations to the capillary length d0 and the kinetic coecient kin have to be

148

A. Numerical Issues of Diuse Interface Modeling

found. To do so it is convenient to assume that d0 and kin are constants,


which are independent of the temperature eld at the interface [61]. Moreover the TolmanBu eect [226], which describes a correction of the surface
tension originating from curvature, is neglected. Then it is justied to derive
the respective dependencies for a planar front as follows:
Since the capillary length d0 is proportional to the surface tension , d0
is obtained by comparing the Gibbs free energy GSIM of an innite, onedimensional system with sharp interface to the Gibbs free energy GP F M of
the corresponding phase-eld model with nite interface width. Assuming
>
the interface at x = 0, it is the limit 0, i.e. (x) = 1 for x < 0, in
which GSIM follows from GP F M :

2

 2
(x)
(x)4

(x)2
(A.13)
GP F M =
+ V0

4
2
x
2


C(x)2
Q(x)2
+M0
C(x) +
+ Q(x) + 1 + L ()(1 C Q) dx
2
2

  2
CS
Q2S
GSIM = M0
(A.14)
+
CS + QS + 1 dx .
2
2

CL,S and QL,S are thermodynamic equilibrium values of C and Q in liquid


and solid phase, respectively. Since the surface energy is dened as the
dierence of GP F M and GSIM , one obtains an expression for by inserting
the stationary solution of the phase-eld model, which is given by
x


2

(x)

1 + y2

V0 (y 4 1)2 + M0 (y 2 1)2
C(x) = C0 + L ((x)) and
Q(x) = Q0 + L ((x)) .
0

dy ,

(A.15)
(A.16)
(A.17)

Thus
2




(x)
(x)4
1
(x)2
2
+ M0
=
+ V0

L ((x)) dx
x
4
2
2


V0
 1
4
2
2
2

M0 (y 1) + (y 1)
dy .
(A.18)
= 2 M0
1 + y2
0


2
2

Equation (A.18) relates the surface energy to the phase-eld parameters , V0


and M0 . From this the , V0 and M0 dependence of d0 can be obtained as

 1
V0
4
2
2
2
M0 (y 1) + (y 1)
2

dy .
(A.19)
d0 =
1 + y2
M0 0

A.1 Relationship Between Physical Variables and Numerical Parameters

149

A more detailed derivation based on a generalized GibbsThomson relation,


which starts from the specication of liquid and solid chemical potentials
with respect to an idealized phase diagram, is given in [2].
To derive an analogous expression for kin , I will turn to the quasistationary solutions of the phase-eld model equations. For a velocity v 1 they
solve the dierential equations
v

= 2 2
x
x

1 2
(1 + 2 )2
L ((x))
+v
x
L ((x))
+v
,
x

+V0 ( 3 ) + M0
T
x
C
v
x
v

2T
x2
2C
= DC 2
x
= DT


C +T +2

1 + 2


(A.20)
(A.21)
(A.22)

with stationary solutions (v = 0) (A.15) and


T0 (x) = C0 (x) = const .

(A.23)

If one considers the physically relevant sharp interface limit, for which the
diusion lengths lD = 2Dv T and lC = 2Dv C are large compared to the diuse
interface width, in (A.21) and (A.22) the terms with factor v can be neglected
and C as well as T can be assumed to be constant in the framework of
an adiabatic approximation: T (x) = T0 ; C(x) = C0 . From the phase-eld
equation (A.20) one obtains the following expression after integration of in
the interval [1; 1]:


V0

4
2
2
2
2 v 1 M0 (y 1) + (y 1)
T0 + C0 =
dy .
(A.24)
1 + y2
M0 0
Since TInt + CInt = T0 + C0 = kin v, this can be rewritten to yield


V0
 1
4
2
2
2
M0 (y 1) + (y 1)
2
dy .
(A.25)
kin =
1 + y2
M0 0
Moreover because of (A.19) one obtains a relation between the capillary
length and the kinetic coecient, which reads
kin =

d0
.
2

(A.26)

Thus in contrast to the sharp interface model the phase-eld model displays model inherent kinetic eects [192]. Therefore, carrying out quantitative
phase-eld simulations for materials with vanishing kinetic eects requires to

150

A. Numerical Issues of Diuse Interface Modeling

work in parameter regimes, for which the eect of the kinetic term kin vn
is small compared to the curvature term d0 . However, for those parameter
regimes phase-eld modeling is usually inecient due to the large diusion
lengths originating from the low driving forces.
The application of phase-eld models to dendritic growth of vanishing kinetic eects at moderate to large velocities became possible only after the introduction of the isothermal approach proposed by Karma and Rappel [167],
as mentioned already in Chap. 1 and Chap. 5. The basic idea underlying
that approach is to take explicitly into account the dependence of the temperature and solutal elds on the space variables within the interfacial region
|x| <
. Essentially this implies carrying out the asymptotic analysis described in Sect. 5.1 up to the rst order. This yields a solvability condition,
by which the expression for the kinetic coecient kin can be obtained. For
the solidication of a binary alloy it reads:





2
1
1
2 I3
1
kin =
1
.
(A.27)
+
I1 + 2I2
I3 DT
DC
2
M0
In (A.27) the abbreviations I1 , I2 and I3 denote integrals, the specic form
if which depends on the precise version of the phase-eld model equations.
For (A.10)(A.12) they are of the form
 1
(y 1)2

I1 =
dy
(A.28)
V0
4 1)2 + (y 2 1)2
0
(y
M0
 y
 1
1 y2
y

I2 =
d
y dy
(A.29)
2 )2
(1
+
y
V
0
0
0
y 4 1)2 + (
y 2 1)2
M0 (

V0
 1
4
2
2
2
M0 (y 1) + (y 1)
I3 =
)dy .
(A.30)
1 + y2
0

I3
The prefactor 2
in (A.27) corresponds precisely to (A.25), i.e. the sharp
M0
interface limit of the kinetic coecient derived neglecting the spatial changes
of the temperature and solutal eld within the interfacial region. Thus in
(A.27) the expression in parentheses denotes a correction term originating
from the consideration of non-constant behavior of the two diusive elds in
the region |x| <
. For





3
1
1
M0 2
1
1
(A.19)
0
(A.31)
+
=
+
I3
DT
DC
d0 DT
DC

the limit of the correction term is unity, which is the condition for the applicability of the sharp interface limit as given by [168]. Thus in this limit
(A.27) is equal to (A.25).
The great importance of (A.27) results from the fact that it allows us to
minimize the correction term by enlarging or reducing V0 /M0 . Finally for

A.2 Computational Universality of Phase Field Models

2 (DT + DC )
=
DT DC

1
2 I1

I3
+ 2I2

or

d20 M0 (DT + DC )
=
2 DT DC

1
2 I1

I33
+ 2I2

151

(A.32)

it vanishes. Equation (A.32) provides a condition, which allows us to carry


out quantitative phase-eld computations for the solidication of binary alloys with vanishing kinetic coecient. Together with (A.19) and (A.25) it
constitutes the set of expressions relating numerical phase-eld parameters
to physical parameters.
As mentioned above (A.28)(A.30) and as a result (A.32), as well, depend
on the specic choice of phase-eld model equations. Therefore within the
context of the phase-eld
 model given by (A.10)(A.12) a necessary condition

DT DC
for kin 0 is > 2
DT +DC . For other models this is not necessarily the
case. Some of them allow for a vanishing kinetic coecient simply by variation
of M0 [167].

A.2 Computational Universality of Phase Field Models


Here I will turn from discussing a freedom in the precise choice of phase-eld
parameters towards a freedom with respect to the precise choice of the potential underlying the model equations by referring to the notion of universality
of phase-eld models. The rst time the term universality of phase-eld models appeared within this text, it was used exactly to denote such a freedom
with respect to the choice of the free energy F underlying the derivation of a
diuse interface model. In this sense it refers to the fact that a large class of
free energies give rise to the same sharp interface equations. Speaking of computational universality exploits this fact and goes a step further: If there is a
freedom in the choice of the precise potential and if there is also some freedom
in the choice of numerical parameters of a phase-eld model as described in
the previous section, then how does this manifest itself in the computational
eciency of a phase-eld model? Are CPU times and convergence rates in the
end factors to distinguish between dierent models and classify them? The
discussion in this section following [182] will reveal that the latter is not the
case. In this sense it will establish the notion of computational universality.
In the remainder of this section time-dependent solutions of dierent
phase-eld models for dendritic solidication in two dimensions are compared.
To this end extensive computations using a specially developed adaptive mesh
renement algorithm [239, 240] are reported. These indicate that when properly used, all phase-eld models give precise results, i.e. not only does each
phase-eld model converge to the steady state predicted by theory, but also
the transient dynamics approach the steady state uniquely. Indeed, once one
has established that there is genuine universal dynamic behavior, the only
remaining question is that of computational eciency. The results discussed
here clearly indicate that the CPU times required for the dierent models are

152

A. Numerical Issues of Diuse Interface Modeling

identical. In particular, there is no advantage for thermodynamically consistent models. Moreover nite discrepancies of the interfacial Peclet number
P eInt encountered between dierent models are shown to be eliminable by
adjusting the phase-eld parameters.
The phase-eld equations considered here are of the form
T
t
(n)

= D2 T +

1 h()
2 t

F (, T )

= (W 2 (n))
+
t

x



W
(n)
+
||2 W (n)
,
y
y

||2 W (n)

W (n)
x

(A.33)


(A.34)

as in [168, 169]. Here the order parameter is denoted by , with = +1 in


the solid and = 1 in the liquid phase. The interface is dened by = 0.
Moreover x = /x and y = /y represent partial derivatives with
respect to x and y.
The function F (, T ) = f ()+T g() is a phenomenological free energy
where f () has the form of a doublewell potential, controls the coupling
between T and , and the relative height of the free energy minima is determined by T and g(). The function h() accounts for the release of latent
heat. Anisotropy has been introduced in (A.34) by dening W (n) = a(n)
and (n) = o a2 (n). o is a time characterizing atomic movement in the
solidliquid interface region, is a length characterizing the width of the
interface, and


4 (x )4 + (y )4
(A.35)
a(n) = (1 3) 1 +
1 3
||4
with a(n) [0, 1]. a is identical with the function a in (2.8), except that here
it is given depending on n rather than on , where n is the normal vector
at the contours of . For phase-eld computations this is convenient, since n
can easily be determined via
+ y y
)/(2x + 2y )1/2 .
n = (x x

(A.36)

The constant parameterizes the deviation of W (n) from and is a measure


of the anisotropy strength as in (2.8).
The asymptotic relationships of Karma and Rappel [168, 169] are used to
map the phase-eld model onto the sharp interface free-boundary problem,
where (A.33) and (A.34)
In terms of a(n), (n) =

reduce2 to (A.3)(A.5).

a(n)
o a(n) and d(n) = do a(n) +
,
where

is
the angle between n and
2
the x-axis, these expressions become (2.8) and d(n) = do (115 cos 4) in the
free-boundary problem (note that tan() = y /x ). The parameters of the
phase-eld model are related to the free-boundary parameters by = a1 /do

A.2 Computational Universality of Phase Field Models

153

and o = 3 a1 a2 /(do D)+ 2 o /do . The positive constants a1 and a2 depend on


the exact form of the phase-eld equations. In choosing to simulate particular
material characteristics, the experimentally measurable quantities do , , and
D are xed, leaving as a free parameter, which determines and o .
Within the simulations reported here fourfold symmetric dendrites are
computed in a quarter-innite space using a nite-element adaptive grid
method as in [239, 240]. Solidication is initiated by a small quarter disk
of radius Ro centered at the origin. The order
parameter is initially set to
its equilibrium value o (x) = tanh((|x| Ro )/ 2) along the interface. The
initial temperature is T = 0 in the solid. Without loss of generality it is
assumed to decay exponentially from T = 0 at the interface to T =
as x , where the far-eld undercooling is given by (A.2). This implies
vanishing in (A.6).

Table A.1. Overview over the phase-eld models under investigation


Model
1
2
3
4
5

g()

1
(1 2 )2
(1 2 )3
(1 2 )4
(1 2 )2

h()

15
( 23 3 + 15 5 )
8

a1
1

2
5

4 2

1.0312
1.1601
4

a2
5
6
47
75

0.52082
0.45448
0.39809

The dierent phase-eld models studied are summarized in Table A.1. To


satisfy the asymptotics, f () is chosen to be an even function, and g() and
h() are odd. All of the models employ f () = 4 /42 /2. For computational
purposes, g() is chosen such that the two minima of F (, T ) are xed at
= 1. Model 1 is a form used by Almgren [11], model 2 by Karma and
Rappel [168, 169], and model 5 is the thermodynamically consistent form
used by Wang et al. [289]. Models 3 and 4 are forms created in particular
for this numerical study. Note that model 1 requires to be less than 1/.
Otherwise the = 1 state becomes linearly unstable.

Table A.2. Overview of simulation parameters

0.45
0.55
0.65
0.65

L
1000
800
800
800

Ro
17
15
15
15

t
0.010
0.016
0.016
0.004

D
3
2
1
2

do
0.5
0.5
0.5
1.5

v
0.00545
0.0170
0.0469
0.0469

P eInt
0.011
0.034
0.094
0.031

154

A. Numerical Issues of Diuse Interface Modeling

In the simulations, the computational domain is an L L square box. Computations are performed at = 0.65, 0.55 and 0.45. A summary of the parameters used for each simulation run is given in Table A.2, where v = vdo /D
is the dimensionless tip velocity predicted by linear solvability theory, x is
the minimum grid spacing of the mesh [239, 240], and t is the simulation
time step. The phase-eld parameters are chosen for each model so that they
all simulated the same free-boundary problem. For all simulations = 0.05,
= 1 and x = 0.39. As long as = 0.45 or = 0.55 these simulation
results reveal that all of the phase-eld models studied produce identical results for the entire temporal evolution of the dendrite and also converge to
steady state solutions that are within a few percent of those predicted by
linear solvability theory.
At = 0.65 (with do = 0.5), on the other hand, signicant quantitative
dierences between the various phase-eld models are encountered. These
discrepancies can be attributed to nite corrections of the interfacial Peclet
number P eInt at higher orders of the asymptotic expansion. I.e., one can
interpret the deviations as an indication that the solutions do not converge
as function of the expansion parameter. Moreover it implies that the phaseeld equations do not operate within the sharp interface limit. The universal
behavior of the dierent models can be recovered via a decrease of P eInt .
The price to pay is a reduction of computational eciency, since the only
parameter, which can be controlled within P eInt is the interfacial width.
This interfacial width, however, always has to be resolved by an appropriate
grid spacing. At this point a detailed investigation of how to push those nite
P eInt -corrections to still higher orders by careful choices of f (), g() and
h(), seems to be a chance to render phase-eld models computationally more
ecient. However, to my knowledge such a study is yet unpublished.

A.3 Selected State-of-the-Art Numerical Approaches


Within this section I will present two examples, which display how enhanced
computational techniques can be applied to phase-eld simulations. The rst
makes use of an adaptive mesh renement algorithm and employs dynamical
data structures [240] to compute dendritic growth in two dimensions. It is
the basis of the second [155], in which the idea of adaptive mesh renement
is extended to 3D simulations of dendritic growth including hydrodynamic
ow in the liquid phase. It employs a semi-implicit approximation projection
method (SIAPM) along with parallelization. Certainly within the 3D code
the precise algorithm for the grid renement runs somewhat dierent from
that in the 2D case. However, for the sake of simplicity I will discuss grid
renement only for the 2D example. Afterwards I will continue to give an
operational description of the additional features concerned with simulating
uid ow in case of the 3D example. These outlines follow [240] and [155],
respectively. Together the discussion of these two examples provides concrete

A.3 Selected State-of-the-Art Numerical Approaches

155

numbers of which performance can be achieved with respect to diuse interface simulations today.
A.3.1 2D Adaptive Mesh Renement Computation
The model employed with respect to the numerics presented here is of the general form of model equations (A.33)(A.34). These equations are solved using
the Galerkin nite element method on dynamically adapting grids of linear,
isoparametric quadrilateral and triangular elements. The grid is adapted dynamically based on an error estimator that utilizes information from both the
and T elds. In the broadest sense, the algorithm performs functions that
can be divided into two classes. The rst deals with the establishment, maintenance and updating of the nite element grids, the second with evolving
and T on these grids according to (A.33) and (A.34).
Maintaining a grid of nite elements on a data structure known as a
quadtree [92, 229, 255] is the main issue of the rst class of functions. The
quadtree is a tree-like structure with branches up to a pre-specied level
creating children elements. These are themselves data structures that contain information analogous to the parent, from which they branch, but one
level down. This quadtree element data structure is depicted in Fig. A.1.
Every entry on the quadtree contains information pertaining to a four-noded
isoparametric quadrilateral nite element. This information includes the following:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.

values of and T at the four nodes,


the nodal coordinates of the element,
the level of renement of the element on the quadtree,
the value of the current error estimate,
the element number, which contains information about the coordinates
of the element and its level of renement,
an array mapping the four nodes of the element onto the entries of a
global solution array,
pointers to the nearest neighbors of the element sharing a common edge
at the same level of grid renement,
a variable that determines whether or not an element contains further
sub-elements termed child elements,
pointers to the child elements of an element,
a pointer to the parent element, from which an element originates.

A parent element and its four child elements are referred to as a family.
Renement produces a ner mesh within the connes of the original parent
grid by bisecting each side. Unrenement, which consists of fusing the four
child elements back into the parent, has the opposite eect, locally creating a coarser mesh. Both renement and unrenement proceed via dynamic
memory allocation, making the code scalable. Unrenement can occur only
if the child elements do not possess further children of their own. Also, in

156

A. Numerical Issues of Diuse Interface Modeling

order to avoid regions of dierent renement bordering each other, the restriction that any two neighboring quadrilateral elements may be separated
by no more than one level of renement is imposed. Moreover one denes a
level of renement of an element le , so that a uniform grid at renement level
le would contain 2le 2le grid points in a physical domain LB LB .

(b)

(a)

(d)

(e)

(c)

(f)

Fig. A.1. An illustration of the quadtree element data structure. (a) An element
and four child elements. (b) Splitting of the children. (c) Splitting of the childrens
children. (d)(e) Evolution of branching in the quadtree. Branches with triangles
indicate square elements which are bridged with triangular or rectangular elements.
Figure following [240].

Cases, in which an element has no children, a missing neighbor, or no


parent are handled by null pointers. The latter case occurs only for the root
of the quadtree. All elements at a given level of renement on the quadtree
are strung together by a linked list of pointers, referred to as the G-list.
There are as many G-lists as there are levels of renement in the quadtree.
Each pointer in the G-list accesses the location in memory assigned to one
element of the quadtree. The purpose of the G-list is to allow traversing of
the quadrilateral elements sequentially rather than recursively. The latter
procedure would be memory intensive and relatively slow.
In addition the code maintains two independent grids representing special
linear isoparametric triangular and rectangular elements. These elements are
used to connect the extra nodes, which arise when two or more quadrilateral
elements of dierent renement levels border each other. These element types
are referred to as bridging elements. They are maintained as two linked lists of
derived data types, one containing information about triangular elements, the
other rectangular. Elements of both grids include the following information:
1. the values of and T at the three nodes (four for rectangles) of the
element,

A.3 Selected State-of-the-Art Numerical Approaches

157

2. the nodal coordinates,


3. node numbers that map the nodes of the element onto the global solution
array.
The main set of operations performed on the grids described above concern renement of the nite element mesh as a whole. The renement process
is performed only on the quadrilateral mesh. The triangular and rectangular
grids are established after this process is completed. To rene the grid the
code traverses the elements of the quadtree, rening (unrening) any element,
for which the error estimate is above (below) a critical value h (l ). Fusion
of four quadrilateral elements can only occur, if all four error estimates are
below the critical value l , where l < h . Simulations show [240] that for
l = h the grid starts to oscillate, i.e. a large numbers of elements become
alternatively rened at one time step, then unrened at the next.
The processes described so far are grouped into modules that encapsulate
various related tasks. They can cross-reference the data and instructions of
each other. The module highest in hierarchy contains the denition of the
quadtree data structure as well as routines that construct the initial uniform
grid, rene and unrene individual quadrilateral elements and impose the initial conditions. Another module constructs the G-lists. It contains routines
that construct the initial G-list from an initially uniform quadtree data structure. Moreover it adds (deletes) element pointers from the G-list as elements
are created (deleted) from the quadtree. Another module accessing the data
structure of the previous ones has the role of creating the triangular and
rectangular element grids. It contains denitions for creating triangular and
rectangular elements data structures and routines that search the quadtree,
building the linked lists of triangles and rectangles that make up these grids.
The second class of functions is concerned with updating of the nite
element grids. With respect to this point the nal modul runs the integration
of (A.33) and (A.34). This module performs four main processes:
1. It maps the internal element node numbers to the indices of a global
solution vector.
2. It advances the T and eld vectors by Nr time steps on the nite
element grids dened above.
3. It calculates an error estimate for each element of the quadtree, based on
an error estimate of the quadrilateral elements.
4. It invokes the routines to rene the grid according to the error estimator.
Steps (1)-(4) are repeated until a sucient time evolution of the microstructure is established. The variable Nr is set, so that the interface remains within the regions of ne meshing between regriddings. The latter is
typically done every 100 time steps. Step (1) involves searching for all elements and the respective neighbors, as well as assigning each node a unique
number, if it has a counterpart on a global solution vector.

158

A. Numerical Issues of Diuse Interface Modeling

The nite element discretization of (A.33) and (A.34) is done using


Galerkins weighted residual method [78]. The method begins by assuming
that and T are interpolated within an element as
N

e =

ei Ni (x, y)

i=1

Te =

Tie Ni (x, y)

(A.37)

i=1

where ei and Tie are the eld values at the N nodes of the element e and
their interpolated values in its interior. The functions Ni (x, y) are standard
linear interpolation functions appropriate for the used element [303]. They
satisfy
Ni (xj , yj ) = i,j ,
(A.38)
where i,j is the Kronecker delta. Rewriting the dierential equations for
in (A.33) and (A.34) as L = 0, as well as of the T -equation as LT T = 0,
the Galerkin method requires that

(A.39)
Ni (x, y)L e (x, y)dxdy = 0
e

Ni (x, y)LT T e (x, y)dxdy = 0 ,
e

for i = 1, 2, 3, . . . , N , where e represents the area of an element e. Substituting (A.37) into (A.39), one obtains two linear algebraic equations for i
and Ti , i = 1, 2, 3, . . . , N in the element e.
The next step is to dene e = (1 , 2 , 3 , , N )T and Ye = (T1 , T2 , T3 ,
, TN )T , where the superscript T denotes the transpose, making e and Te
column vectors. Then the linear algebraic statement of the nite element
form of (A.33) and (A.34) reads


e
+ E en + Fe ()
e ) d = A
C(
dt
dTe
1 de
= DATe + C
,
C
dt
2 dt

(A.40)

A, A
and E and the vector Fe () are given by
where the matrices C, C,

C=
NT Ndxdy ,
(A.41)


=
C


A=
=
A


e

NT NA2 ((e ))dxdy ,



NT Nx + NT Ny dxdy ,

(A.43)


NT Nx + NT Ny A2 ((e ))dxdy ,

(A.44)

(A.42)


E=

A.3 Selected State-of-the-Art Numerical Approaches


NT Nx NT Ny A((e ))((e ))dxdy ,
e

F () =


e

NT f (e , U e ; )dxdy .

159

(A.45)
(A.46)

Here Nx , Ny denote the partial derivatives of the vector of shape functions


with respect to x and y, respectively. The function A((e )) is given by
(A.35). () is proportional to the derivative of A(). It reads:
((e )) = 16

tan (1 tan2 )
.
(1 + tan2 )2

(A.47)

[78] is employed. In
A lumped formulation for the matrices C and C
this procedure, the row vector of shape functions N in (A.41) is replaced by
the identity row vector I = (1, 1, 1, ). The resulting matrix C consists of
identical columns, each of which contains the element Ni (x, y) at the position
of the ith row. A lumped term is dened as a diagonal matrix of entries which
take the values
nodes

1
Lc =
Ni (x, y)dxdy .
(A.48)
nodes i=1 e
The use of a lumped matrix for C enables the assembly of a diagonal matrix
for the left hand side of (A.40), stored as a one-dimensional vector. (Twodimensional arrays would be required if the consistent formulation for the
assembly of the C matrices were used. However, microstructures evolving at
low undercooling can produce interfaces with over 2 105 elements, making
the storing of 2 105 2 105 matrices impossible.)
The global (obtained after assembly of the element equations in (A.40))
is time-stepped using a forward dierence (explicit) time scheme. For each
time step of the eld, the global T eld is solved iteratively using a Crank
Nicholson scheme. Convergence of T is obtained within a few iterations.
Regridding is based on an error estimator function discussed next. It
is obtained following Zienkiewicz and Zhu [303], based on the dierences
between calculated and smoothed gradients of the and T elds. In particular
a composite eld
= + T
(A.49)
is dened, for which is a constant. This denition enables a regridding in
accordance to the requirements of both the and T eld, as opposed to using
merely the gradients of the -eld in establishing the grid [39]. Since and
T have to be determined rather than their gradients, one does not expect the
gradient of to be continuous across element boundaries, due to the order
of the interpolation used. Thus the dierence between the calculated and the
smoothed gradients, which are both continuous across element boundaries,

160

A. Numerical Issues of Diuse Interface Modeling

can be expected to provide a reasonable estimate of error. This method appropriately meshes regions of both steep gradients and regions where the
and T elds change rapidly.
The error estimator function e is dened as
e = qs qc ,

(A.50)

where qc and qs are the calculated and smoothed gradients of , respectively.


To determine qs , one assumes it to be interpolated in the same way as the
and T elds, namely
qs = NQs ,
(A.51)
where N is the row vector of element shape functions and Qs a 4 2 matrix,
which columns represent the nodal values of uxes of in the x and y direction, respectively. To nd Qs , Galerkins method is employed, minimizing
the weighted residual


(A.52)
NT (NQs qc )d = 0 .
NT ede =
e

This calculation is simplied by lumping the left hand side of (A.52), leading
to



e

NT Id Qs =

NT qc d .

(A.53)

Assembling (A.53) for all quadrilateral elements yields an equation for the
smoothed gradients Qg of the global eld , at all element nodes, of the form
DQg = b ,

(A.54)

where D is a diagonal matrix, due to masslumping, and Qg is a N 2


matrix for the global, smoothed ux.
For the actual error updating on the elements of the quadtree one employs
the normalized error

|(qs qc )|2 de
2
e

.
(A.55)
Ee = 
2
e |qs | d
The domain of integration in the denominator denotes the entire domain
of the problem. Thus Ee2 gives the contribution of the local element error
relative to the total smoothed error calculated over the entire grid.
The CPU time of this approach scales with the arc length of the problem
under investigation. For most computations this implies an increase in computational eciency compared to computations on a uniform grid, which scale
with the size of the required grid itself. E.g., for = 0.70, D = 2, dt = 0.016,
chosen to simulate = 0 and system size 800 800 with xmin = 0.4,
simulations of 105 time steps take approximately 15 CPU-hours on a Sun
UltraSPARC 2200 workstation with 512 MHz.

A.3 Selected State-of-the-Art Numerical Approaches

161

A.3.2 3D Simulations Including Fluid Flow


The phase-eld model underlying the computational approach presented
within this section employs the same phase-eld model equations as the previous section. To extend it to hydrodynamic ow in the liquid phase a method
proposed by Beckermann et al. [26] is employed. Within this approach a phase
average k of a variable for phase k over volume V is dened as

2 V Xk dV
k =
,
(A.56)
1
where Xk {0, 1} is an existence function. The formulation ensures that the
uid velocity is extinguished in the solid, and further that the shear stress at
the liquidsolid interface is handled correctly.
The governing equations for simulating dendritic growth with uid ow
are the mixture continuity equation:


1

u =0,
(A.57)
2
where u is the velocity vector, the averaged momentum equation:



 

1
1
1

u +u
u +
p
t
2
2
2


h2 (1 2 )(1 + )
1
= 2
u
u,
(A.58)
2
8 2

where t is time, p the pressure, the kinematic viscosity, = W0 / 2 the


characteristic interface width, and h a constant (=2.757), which ensures that
the interface shear stress is correct for a simple shear ow (see Beckermann
et al. [26]), and the averaged energy conservation equation in terms of
the dimensionless temperature T :


T
1
1
+
u T = D2 T +
.
(A.59)
t
2
2 t
The phase averages of velocity and pressure are used for deriving the mixture continuity equation, averaged liquid momentum equation and averaged
energy conservation equation. These are solved together with the 3D phaseeld evolution equation, which is a straightforward extension of (A.34).
Here the 3D ow equations are solved using the SIAPM by Gresho [136].
SIAPM is a predictor-corrector method, which can solve (A.57) and (A.58)
eectively, especially for large 3D problems, because it uses relatively small
amounts of memory. The velocity degrees of freedom are solved in a segregated form. Pressure is updated using a projection method. For a detailed
discussion of the algorithm, the reader is referred to the original paper [136].
The algorithm consists of three steps, which are summarized briey in the
following:

162

A. Numerical Issues of Diuse Interface Modeling

n+1 from
1. Compute an intermediate velocity u




1
1
1
1
n+1
M K+F u
M
+
K
uni
=
i
t
2
t
2
A(un )uni Gi pn ,

(A.60)

n+1
is the vector of nodal values of the intermediate velocity
where u
i
component i at time step n + 1, uni is the corresponding vector at time
step n, and pn is the vector of nodal pressures at time step n. The
coecient matrices are dened in terms of the velocity shape functions
N as follows:

(1 ) T
(A.61)
M =
N Nd
2




( 1) NT N
NT N
d
(A.62)
K =

+I
x x
xk xk
2


h(1 2 )(1 + ) T
F =
N Nd
(A.63)
8 2

(1 ) T n N
N uk
A(un ) =
d
(A.64)
2
xk


(1 ) T N
Gi =
d .
(A.65)
N
2
xi

2. The velocity eld found in the rst step is generally not divergencefree. The next step corrects the pressure to obtain an approximately
divergence-free velocity eld by solving a Poisson equation for
pn+1 = pn+1 pn :
Lpn+1 =
where

 n+1

1

D u
un ,
t

(A.66)

(1 ) NT N
I
d
2
xk xk

(1 ) T N
N
d .
D=
2
x

L =

(A.67)
(A.68)

pn+1 is then updated according to


pn+1 = pn + pn+1 .

(A.69)

3. Finally, the projected velocity un+1 is computed in a corrector step by


solving
n+1 tML 1 Gpn+1 ,
un+1 = u
(A.70)
where


ML =

1
Nd .
2

(A.71)

A.3 Selected State-of-the-Art Numerical Approaches

163

The computations are started using an initial velocity eld u0 determined


from the boundary conditions. In order to obtain a eld u0 , whose discrete
divergence is close to zero, the pressure is adapted by iteration. The initial
pressure is zero everywhere.
Equations (A.60) and (A.66) are solved by the conjugate gradient (CG)
method with diagonal preconditioning [217]. The SIAPM can calculate the
velocity eld for large 3D problems much faster than fully implicit timestepping methods, because convergence for (A.60) is reached within a few
iterations. Moreover, the number of degrees of freedom of (A.66) is one. The
CG iteration for (A.66) converges more slowly, typically within 50-200 iterations.
The averaged energy equation (A.59) is solved using the CG method with
diagonal preconditioning as well. Streamline upwind schemes [45] are employed for the convection terms in (A.60) and (A.59). The 3D phase-eld
equation is a nonlinear system. In order to solve the system implicitly, an
iterative method such as the NewtonRaphson method is required. Alternatively an explicit time-stepping scheme can be employed. Stable solutions are
obtained from the explicit scheme, since the variation of the eld exists
only in the interface region and a suciently small time increment t can
be used.
In analogy to the previous section the algorithm is implemented along
with adaptive grid renement. Moreover it is parallelized on the basis of the
Charm++FEM framework [158].
Based on this implementation in [155] for a set of problem parameters
corresponding to SCN, i.e. t = 0.016, W = 1, 0 = 1, = 6.383, d0 (=
a1 W
) = 0.139, = 92.4 and P r = 23.1 computations of 100 time steps
of a growing 3D dendrite are reported to require 190 hours of CPU for an
underlying minimum grid spacing of xmin = 0.4. 80% to 90% of the CPU
time were spent for the update of the velocity and the pressure eld [155].
To my knowledge these 3D simulations of SCN including uid ow are the
most extensive computations of dendritic growth so far. Everything beyond
still remains an open challenge from point of view of numerics as well as
from point of view of modeling, which might help to simplify numerics.

A. Numerical Issues
of Diuse Interface Modeling

This appendix is concerned with the questions: What happens after a diuse
interface model has been established? How straightforward is its implementation? Is there anything one has to take special care of with respect to that
implementation? The answer to the latter is a clear yes. Answering it in
more detail will leave the question of the previous sections whether or not a
diuse interface model under consideration is thermodynamically consistent
behind. Rather one is concerned with a choice of model parameters which
will yield the best numerical performance. For dendritic growth it could be
proven that from point of view of run time performance models which do not
display thermodynamic consistency and can only be validated by the method
of matched asymptotics expansion can be superior to thermodynamic consistent ones [182]. In that sense referring to the implementation point of view
one nds once more1 the expression universality of diuse interface models
in the previous citation.
Since the early attempts of phase-eld computation in the context of
solidication by Smith [259] great advances in the accuracy of those computations have been encountered. The reasons are twofold: First, computer
technology has advanced signicantly and thus allows us to integrate the unsteady phase-eld equations for increasingly complex congurations. Second,
there has also been progress in the numerical techniques applied to the implementation of phase-eld models. This involves choosing an optimal set of
numerical parameters and model potentials just as well as considering enhanced numerical techniques as parallel, adaptive or multi-grid algorithms.
The rst of the above two issues arises since
1. dierent potentials can be employed, resulting in dierent phase-eld
models, which nevertheless describe the same growth phenomenon in
their sharp interface limit and
2. to establish the correspondence between the diuse interface model and
the sharp interface model there is some freedom in how to arrange the
numerical parameters of the diuse interface, i.e. a physical variable ap1

once more refers to the discussion at the end of Chap. 5, where the notion
of universality in the context of phase-eld modeling has been used to denote a
freedom in the precise choice of F .

Heike Emmerich (Ed.): LNP m73, pp. 145163, 2003.


c Springer-Verlag Berlin Heidelberg 2003


146

A. Numerical Issues of Diuse Interface Modeling

pearing in the sharp interface model is usually expressed by more than


one numerical parameter of the diuse interface model.
I will demonstrate the above in detail in Sect. A.1 and Sect. A.2 for the case
of dendritic growth, for which it has received much attention by the scientic
community over the last decade, e.g. [113, 182, 232]. Whereas Sect. A.1 is
devoted to the relationship between physical variables and numerical parameters, thereby pointing out the freedom in the precise choice of model parameters, Sect. A.2 compares the computational eciency of dierent phase-eld
models depending on the choice of the underlying themodynamic potential.
In Sect. A.3 I will proceed to discuss some state of the art numerical
approaches to interfacial growth phenomena and thereby give a look at how
advanced numerical techniques can be employed to speed up diuse interface
computation a step which is certainly required to deal with more complex,
challenging growth conditions.

A.1 Relationship Between Physical Variables


and Numerical Parameters
With respect to phase-eld computations in dendritic solidication Kobayashi [187] was the rst to become famous for quite beautiful animations of
3D phase-eld simulations of evolving dendrites. However, this work was still
very qualitative. To use phase-eld simulation for quantitative predictions of
dendritic growth morphologies one has to take a closer look at the relationship
between the physical variables of the sharp interface limit and the numerical
parameters of the phase-eld model equations.
For diusion limited dendritic growth the GibbsThomson relation, the
Stefan condition and the diusion equation for the transport of heat in the
two bulk phases constitute the free boundary problem, which describes the
evolution of the dendritic front. The respective model equations (2.5)(2.7)
are rewritten here by transforming to a dimensionless temperature T via
T =


cp (T  T
)
.
L

(A.1)

This transformation employs the specic heat of the solid given by cp . More
over, T
refers to the dimensional temperature of the undercooled melt far
from the interface. In addition one can introduce a dimensionless supercooling
=



cp (TM
T
)
L

(A.2)

of the liquid phase which is normalized, so that for = 1 the solidication of


one volume element of the liquid phase results in precisely the heat required to


heat up the same volume element from T
to TM
. With these transformations
the set of dimensionless equations corresponding to (2.5)(2.7) reads:

A.1 Relationship Between Physical Variables and Numerical Parameters

T
= D2 T
t
T = d0 ()
v n = D [(T )s (T )l ] n .

147

(A.3)
(A.4)
(A.5)

For large driving forces the consideration of local thermodynamic equilibrium is not valid any longer. This is taken into account by introducing a
kinetic term kin vn into the GibbsThomson relation and thus replacing
the dimensionless equation (A.4) by:
T = d0 () kin v n .

(A.6)

Here kin denotes a kinetic coecient, which is a material parameter (compare to (2.10)). d0 () is the capillary length given by (2.12).
For quantitative application of a derived phase-eld model to a concrete
dendritic growth problem, the capillary length d0 and the kinetic coecient
kin of (A.6) have to be related to the numerical parameters of the phaseeld model. To demonstrate how this can be achieved, I introduce a concrete phase-eld model for dendritic growth of a binary alloy by specifying
G(, X, r) in the remainder of this section. Following [2] I choose
2
4
2
()2 + V0 (

)
(A.7)
2 
4
2

Q2
C2
C
+ Q + 1 + L ()(1 C Q) ,
+ M0
2
2

G(, Q, C) =

with

.
(A.8)
+
2 1 + 2
Here C denotes the solutal eld. To overcome the problems related to phaseeld modeling based on the dimensionless temperature T (see Sect. 4.2) the
dimensionless energy density Q is used instead. It is related to T by

T
: solid phase
Q=
(A.9)
T + 1 : liquid phase .
L

With this choice of G and after retransformation from Q to T the set of


phase-eld model equations I discuss here reads



1 2
2 2
3
, (A.10)

T+
= + V0 ( ) + M0
t
(1 + 2 )2
1+
C
(A.11)
= (DC (C L ())) ,
t
L ()
T
= (DT T )
.
(A.12)
t
t
In (A.10)(A.12), it is the set of parameters , V0 , and M0 for which
relations to the capillary length d0 and the kinetic coecient kin have to be

148

A. Numerical Issues of Diuse Interface Modeling

found. To do so it is convenient to assume that d0 and kin are constants,


which are independent of the temperature eld at the interface [61]. Moreover the TolmanBu eect [226], which describes a correction of the surface
tension originating from curvature, is neglected. Then it is justied to derive
the respective dependencies for a planar front as follows:
Since the capillary length d0 is proportional to the surface tension , d0
is obtained by comparing the Gibbs free energy GSIM of an innite, onedimensional system with sharp interface to the Gibbs free energy GP F M of
the corresponding phase-eld model with nite interface width. Assuming
>
the interface at x = 0, it is the limit 0, i.e. (x) = 1 for x < 0, in
which GSIM follows from GP F M :

2

 2
(x)
(x)4

(x)2
(A.13)
GP F M =
+ V0

4
2
x
2


C(x)2
Q(x)2
+M0
C(x) +
+ Q(x) + 1 + L ()(1 C Q) dx
2
2

  2
CS
Q2S
GSIM = M0
(A.14)
+
CS + QS + 1 dx .
2
2

CL,S and QL,S are thermodynamic equilibrium values of C and Q in liquid


and solid phase, respectively. Since the surface energy is dened as the
dierence of GP F M and GSIM , one obtains an expression for by inserting
the stationary solution of the phase-eld model, which is given by
x


2

(x)

1 + y2

V0 (y 4 1)2 + M0 (y 2 1)2
C(x) = C0 + L ((x)) and
Q(x) = Q0 + L ((x)) .
0

dy ,

(A.15)
(A.16)
(A.17)

Thus
2




(x)
(x)4
1
(x)2
2
+ M0
=
+ V0

L ((x)) dx
x
4
2
2


V0
 1
4
2
2
2

M0 (y 1) + (y 1)
dy .
(A.18)
= 2 M0
1 + y2
0


2
2

Equation (A.18) relates the surface energy to the phase-eld parameters , V0


and M0 . From this the , V0 and M0 dependence of d0 can be obtained as

 1
V0
4
2
2
2
M0 (y 1) + (y 1)
2
dy .
(A.19)
d0 =
1 + y2
M0 0

A.1 Relationship Between Physical Variables and Numerical Parameters

149

A more detailed derivation based on a generalized GibbsThomson relation,


which starts from the specication of liquid and solid chemical potentials
with respect to an idealized phase diagram, is given in [2].
To derive an analogous expression for kin , I will turn to the quasistationary solutions of the phase-eld model equations. For a velocity v 1 they
solve the dierential equations
v

= 2 2
x
x

1 2
(1 + 2 )2
L ((x))
+v
x
L ((x))
+v
,
x

+V0 ( 3 ) + M0
T
x
C
v
x
v

2T
x2
2C
= DC 2
x
= DT


C +T +2

1 + 2


(A.20)
(A.21)
(A.22)

with stationary solutions (v = 0) (A.15) and


T0 (x) = C0 (x) = const .

(A.23)

If one considers the physically relevant sharp interface limit, for which the
diusion lengths lD = 2Dv T and lC = 2Dv C are large compared to the diuse
interface width, in (A.21) and (A.22) the terms with factor v can be neglected
and C as well as T can be assumed to be constant in the framework of
an adiabatic approximation: T (x) = T0 ; C(x) = C0 . From the phase-eld
equation (A.20) one obtains the following expression after integration of in
the interval [1; 1]:


V0

4
2
2
2
2 v 1 M0 (y 1) + (y 1)
T0 + C0 =
dy .
(A.24)
1 + y2
M0 0
Since TInt + CInt = T0 + C0 = kin v, this can be rewritten to yield


V0
 1
4
2
2
2
M0 (y 1) + (y 1)
2
dy .
(A.25)
kin =
1 + y2
M0 0
Moreover because of (A.19) one obtains a relation between the capillary
length and the kinetic coecient, which reads
kin =

d0
.
2

(A.26)

Thus in contrast to the sharp interface model the phase-eld model displays model inherent kinetic eects [192]. Therefore, carrying out quantitative
phase-eld simulations for materials with vanishing kinetic eects requires to

150

A. Numerical Issues of Diuse Interface Modeling

work in parameter regimes, for which the eect of the kinetic term kin vn
is small compared to the curvature term d0 . However, for those parameter
regimes phase-eld modeling is usually inecient due to the large diusion
lengths originating from the low driving forces.
The application of phase-eld models to dendritic growth of vanishing kinetic eects at moderate to large velocities became possible only after the introduction of the isothermal approach proposed by Karma and Rappel [167],
as mentioned already in Chap. 1 and Chap. 5. The basic idea underlying
that approach is to take explicitly into account the dependence of the temperature and solutal elds on the space variables within the interfacial region
|x| <
. Essentially this implies carrying out the asymptotic analysis described in Sect. 5.1 up to the rst order. This yields a solvability condition,
by which the expression for the kinetic coecient kin can be obtained. For
the solidication of a binary alloy it reads:





2
1
1
2 I3
1
kin =
1
.
(A.27)
+
I1 + 2I2
I3 DT
DC
2
M0
In (A.27) the abbreviations I1 , I2 and I3 denote integrals, the specic form
if which depends on the precise version of the phase-eld model equations.
For (A.10)(A.12) they are of the form
 1
(y 1)2

I1 =
dy
(A.28)
V0
4 1)2 + (y 2 1)2
0
(y
M0
 y
 1
1 y2
y

I2 =
d
y dy
(A.29)
2 )2
(1
+
y
V
0
4 1)2 + (
2 1)2
0
0
(
y
y
M0

V0
 1
4
2
2
2
M0 (y 1) + (y 1)
I3 =
)dy .
(A.30)
1 + y2
0

I3
The prefactor 2
in (A.27) corresponds precisely to (A.25), i.e. the sharp
M0
interface limit of the kinetic coecient derived neglecting the spatial changes
of the temperature and solutal eld within the interfacial region. Thus in
(A.27) the expression in parentheses denotes a correction term originating
from the consideration of non-constant behavior of the two diusive elds in
the region |x| <
. For





3
1
1
M0 2
1
1
(A.19)
0
(A.31)
+
=
+
I3
DT
DC
d0 DT
DC

the limit of the correction term is unity, which is the condition for the applicability of the sharp interface limit as given by [168]. Thus in this limit
(A.27) is equal to (A.25).
The great importance of (A.27) results from the fact that it allows us to
minimize the correction term by enlarging or reducing V0 /M0 . Finally for

A.2 Computational Universality of Phase Field Models

2 (DT + DC )
=
DT DC

1
2 I1

I3
+ 2I2

or

d20 M0 (DT + DC )
=
2 DT DC

1
2 I1

I33
+ 2I2

151

(A.32)

it vanishes. Equation (A.32) provides a condition, which allows us to carry


out quantitative phase-eld computations for the solidication of binary alloys with vanishing kinetic coecient. Together with (A.19) and (A.25) it
constitutes the set of expressions relating numerical phase-eld parameters
to physical parameters.
As mentioned above (A.28)(A.30) and as a result (A.32), as well, depend
on the specic choice of phase-eld model equations. Therefore within the
context of the phase-eld
 model given by (A.10)(A.12) a necessary condition

DT DC
for kin 0 is > 2
DT +DC . For other models this is not necessarily the
case. Some of them allow for a vanishing kinetic coecient simply by variation
of M0 [167].

A.2 Computational Universality of Phase Field Models


Here I will turn from discussing a freedom in the precise choice of phase-eld
parameters towards a freedom with respect to the precise choice of the potential underlying the model equations by referring to the notion of universality
of phase-eld models. The rst time the term universality of phase-eld models appeared within this text, it was used exactly to denote such a freedom
with respect to the choice of the free energy F underlying the derivation of a
diuse interface model. In this sense it refers to the fact that a large class of
free energies give rise to the same sharp interface equations. Speaking of computational universality exploits this fact and goes a step further: If there is a
freedom in the choice of the precise potential and if there is also some freedom
in the choice of numerical parameters of a phase-eld model as described in
the previous section, then how does this manifest itself in the computational
eciency of a phase-eld model? Are CPU times and convergence rates in the
end factors to distinguish between dierent models and classify them? The
discussion in this section following [182] will reveal that the latter is not the
case. In this sense it will establish the notion of computational universality.
In the remainder of this section time-dependent solutions of dierent
phase-eld models for dendritic solidication in two dimensions are compared.
To this end extensive computations using a specially developed adaptive mesh
renement algorithm [239, 240] are reported. These indicate that when properly used, all phase-eld models give precise results, i.e. not only does each
phase-eld model converge to the steady state predicted by theory, but also
the transient dynamics approach the steady state uniquely. Indeed, once one
has established that there is genuine universal dynamic behavior, the only
remaining question is that of computational eciency. The results discussed
here clearly indicate that the CPU times required for the dierent models are

152

A. Numerical Issues of Diuse Interface Modeling

identical. In particular, there is no advantage for thermodynamically consistent models. Moreover nite discrepancies of the interfacial Peclet number
P eInt encountered between dierent models are shown to be eliminable by
adjusting the phase-eld parameters.
The phase-eld equations considered here are of the form
T
t
(n)

= D2 T +

1 h()
2 t

F (, T )

= (W 2 (n))
+
t

x



W
(n)
+
||2 W (n)
,
y
y

||2 W (n)

W (n)
x

(A.33)


(A.34)

as in [168, 169]. Here the order parameter is denoted by , with = +1 in


the solid and = 1 in the liquid phase. The interface is dened by = 0.
Moreover x = /x and y = /y represent partial derivatives with
respect to x and y.
The function F (, T ) = f ()+T g() is a phenomenological free energy
where f () has the form of a doublewell potential, controls the coupling
between T and , and the relative height of the free energy minima is determined by T and g(). The function h() accounts for the release of latent
heat. Anisotropy has been introduced in (A.34) by dening W (n) = a(n)
and (n) = o a2 (n). o is a time characterizing atomic movement in the
solidliquid interface region, is a length characterizing the width of the
interface, and


4 (x )4 + (y )4
(A.35)
a(n) = (1 3) 1 +
1 3
||4
with a(n) [0, 1]. a is identical with the function a in (2.8), except that here
it is given depending on n rather than on , where n is the normal vector
at the contours of . For phase-eld computations this is convenient, since n
can easily be determined via
+ y y
)/(2x + 2y )1/2 .
n = (x x

(A.36)

The constant parameterizes the deviation of W (n) from and is a measure


of the anisotropy strength as in (2.8).
The asymptotic relationships of Karma and Rappel [168, 169] are used to
map the phase-eld model onto the sharp interface free-boundary problem,
where (A.33) and (A.34)
In terms of a(n), (n) =

reduce2 to (A.3)(A.5).

a(n)
o a(n) and d(n) = do a(n) +
,
where

is
the angle between n and
2
the x-axis, these expressions become (2.8) and d(n) = do (115 cos 4) in the
free-boundary problem (note that tan() = y /x ). The parameters of the
phase-eld model are related to the free-boundary parameters by = a1 /do

A.2 Computational Universality of Phase Field Models

153

and o = 3 a1 a2 /(do D)+ 2 o /do . The positive constants a1 and a2 depend on


the exact form of the phase-eld equations. In choosing to simulate particular
material characteristics, the experimentally measurable quantities do , , and
D are xed, leaving as a free parameter, which determines and o .
Within the simulations reported here fourfold symmetric dendrites are
computed in a quarter-innite space using a nite-element adaptive grid
method as in [239, 240]. Solidication is initiated by a small quarter disk
of radius Ro centered at the origin. The order
parameter is initially set to
its equilibrium value o (x) = tanh((|x| Ro )/ 2) along the interface. The
initial temperature is T = 0 in the solid. Without loss of generality it is
assumed to decay exponentially from T = 0 at the interface to T =
as x , where the far-eld undercooling is given by (A.2). This implies
vanishing in (A.6).

Table A.1. Overview over the phase-eld models under investigation


Model
1
2
3
4
5

g()

1
(1 2 )2
(1 2 )3
(1 2 )4
(1 2 )2

h()

15
( 23 3 + 15 5 )
8

a1
1

2
5

4 2

1.0312
1.1601
4

a2
5
6
47
75

0.52082
0.45448
0.39809

The dierent phase-eld models studied are summarized in Table A.1. To


satisfy the asymptotics, f () is chosen to be an even function, and g() and
h() are odd. All of the models employ f () = 4 /42 /2. For computational
purposes, g() is chosen such that the two minima of F (, T ) are xed at
= 1. Model 1 is a form used by Almgren [11], model 2 by Karma and
Rappel [168, 169], and model 5 is the thermodynamically consistent form
used by Wang et al. [289]. Models 3 and 4 are forms created in particular
for this numerical study. Note that model 1 requires to be less than 1/.
Otherwise the = 1 state becomes linearly unstable.

Table A.2. Overview of simulation parameters

0.45
0.55
0.65
0.65

L
1000
800
800
800

Ro
17
15
15
15

t
0.010
0.016
0.016
0.004

D
3
2
1
2

do
0.5
0.5
0.5
1.5

v
0.00545
0.0170
0.0469
0.0469

P eInt
0.011
0.034
0.094
0.031

154

A. Numerical Issues of Diuse Interface Modeling

In the simulations, the computational domain is an L L square box. Computations are performed at = 0.65, 0.55 and 0.45. A summary of the parameters used for each simulation run is given in Table A.2, where v = vdo /D
is the dimensionless tip velocity predicted by linear solvability theory, x is
the minimum grid spacing of the mesh [239, 240], and t is the simulation
time step. The phase-eld parameters are chosen for each model so that they
all simulated the same free-boundary problem. For all simulations = 0.05,
= 1 and x = 0.39. As long as = 0.45 or = 0.55 these simulation
results reveal that all of the phase-eld models studied produce identical results for the entire temporal evolution of the dendrite and also converge to
steady state solutions that are within a few percent of those predicted by
linear solvability theory.
At = 0.65 (with do = 0.5), on the other hand, signicant quantitative
dierences between the various phase-eld models are encountered. These
discrepancies can be attributed to nite corrections of the interfacial Peclet
number P eInt at higher orders of the asymptotic expansion. I.e., one can
interpret the deviations as an indication that the solutions do not converge
as function of the expansion parameter. Moreover it implies that the phaseeld equations do not operate within the sharp interface limit. The universal
behavior of the dierent models can be recovered via a decrease of P eInt .
The price to pay is a reduction of computational eciency, since the only
parameter, which can be controlled within P eInt is the interfacial width.
This interfacial width, however, always has to be resolved by an appropriate
grid spacing. At this point a detailed investigation of how to push those nite
P eInt -corrections to still higher orders by careful choices of f (), g() and
h(), seems to be a chance to render phase-eld models computationally more
ecient. However, to my knowledge such a study is yet unpublished.

A.3 Selected State-of-the-Art Numerical Approaches


Within this section I will present two examples, which display how enhanced
computational techniques can be applied to phase-eld simulations. The rst
makes use of an adaptive mesh renement algorithm and employs dynamical
data structures [240] to compute dendritic growth in two dimensions. It is
the basis of the second [155], in which the idea of adaptive mesh renement
is extended to 3D simulations of dendritic growth including hydrodynamic
ow in the liquid phase. It employs a semi-implicit approximation projection
method (SIAPM) along with parallelization. Certainly within the 3D code
the precise algorithm for the grid renement runs somewhat dierent from
that in the 2D case. However, for the sake of simplicity I will discuss grid
renement only for the 2D example. Afterwards I will continue to give an
operational description of the additional features concerned with simulating
uid ow in case of the 3D example. These outlines follow [240] and [155],
respectively. Together the discussion of these two examples provides concrete

A.3 Selected State-of-the-Art Numerical Approaches

155

numbers of which performance can be achieved with respect to diuse interface simulations today.
A.3.1 2D Adaptive Mesh Renement Computation
The model employed with respect to the numerics presented here is of the general form of model equations (A.33)(A.34). These equations are solved using
the Galerkin nite element method on dynamically adapting grids of linear,
isoparametric quadrilateral and triangular elements. The grid is adapted dynamically based on an error estimator that utilizes information from both the
and T elds. In the broadest sense, the algorithm performs functions that
can be divided into two classes. The rst deals with the establishment, maintenance and updating of the nite element grids, the second with evolving
and T on these grids according to (A.33) and (A.34).
Maintaining a grid of nite elements on a data structure known as a
quadtree [92, 229, 255] is the main issue of the rst class of functions. The
quadtree is a tree-like structure with branches up to a pre-specied level
creating children elements. These are themselves data structures that contain information analogous to the parent, from which they branch, but one
level down. This quadtree element data structure is depicted in Fig. A.1.
Every entry on the quadtree contains information pertaining to a four-noded
isoparametric quadrilateral nite element. This information includes the following:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.

values of and T at the four nodes,


the nodal coordinates of the element,
the level of renement of the element on the quadtree,
the value of the current error estimate,
the element number, which contains information about the coordinates
of the element and its level of renement,
an array mapping the four nodes of the element onto the entries of a
global solution array,
pointers to the nearest neighbors of the element sharing a common edge
at the same level of grid renement,
a variable that determines whether or not an element contains further
sub-elements termed child elements,
pointers to the child elements of an element,
a pointer to the parent element, from which an element originates.

A parent element and its four child elements are referred to as a family.
Renement produces a ner mesh within the connes of the original parent
grid by bisecting each side. Unrenement, which consists of fusing the four
child elements back into the parent, has the opposite eect, locally creating a coarser mesh. Both renement and unrenement proceed via dynamic
memory allocation, making the code scalable. Unrenement can occur only
if the child elements do not possess further children of their own. Also, in

156

A. Numerical Issues of Diuse Interface Modeling

order to avoid regions of dierent renement bordering each other, the restriction that any two neighboring quadrilateral elements may be separated
by no more than one level of renement is imposed. Moreover one denes a
level of renement of an element le , so that a uniform grid at renement level
le would contain 2le 2le grid points in a physical domain LB LB .

(b)

(a)

(d)

(e)

(c)

(f)

Fig. A.1. An illustration of the quadtree element data structure. (a) An element
and four child elements. (b) Splitting of the children. (c) Splitting of the childrens
children. (d)(e) Evolution of branching in the quadtree. Branches with triangles
indicate square elements which are bridged with triangular or rectangular elements.
Figure following [240].

Cases, in which an element has no children, a missing neighbor, or no


parent are handled by null pointers. The latter case occurs only for the root
of the quadtree. All elements at a given level of renement on the quadtree
are strung together by a linked list of pointers, referred to as the G-list.
There are as many G-lists as there are levels of renement in the quadtree.
Each pointer in the G-list accesses the location in memory assigned to one
element of the quadtree. The purpose of the G-list is to allow traversing of
the quadrilateral elements sequentially rather than recursively. The latter
procedure would be memory intensive and relatively slow.
In addition the code maintains two independent grids representing special
linear isoparametric triangular and rectangular elements. These elements are
used to connect the extra nodes, which arise when two or more quadrilateral
elements of dierent renement levels border each other. These element types
are referred to as bridging elements. They are maintained as two linked lists of
derived data types, one containing information about triangular elements, the
other rectangular. Elements of both grids include the following information:
1. the values of and T at the three nodes (four for rectangles) of the
element,

A.3 Selected State-of-the-Art Numerical Approaches

157

2. the nodal coordinates,


3. node numbers that map the nodes of the element onto the global solution
array.
The main set of operations performed on the grids described above concern renement of the nite element mesh as a whole. The renement process
is performed only on the quadrilateral mesh. The triangular and rectangular
grids are established after this process is completed. To rene the grid the
code traverses the elements of the quadtree, rening (unrening) any element,
for which the error estimate is above (below) a critical value h (l ). Fusion
of four quadrilateral elements can only occur, if all four error estimates are
below the critical value l , where l < h . Simulations show [240] that for
l = h the grid starts to oscillate, i.e. a large numbers of elements become
alternatively rened at one time step, then unrened at the next.
The processes described so far are grouped into modules that encapsulate
various related tasks. They can cross-reference the data and instructions of
each other. The module highest in hierarchy contains the denition of the
quadtree data structure as well as routines that construct the initial uniform
grid, rene and unrene individual quadrilateral elements and impose the initial conditions. Another module constructs the G-lists. It contains routines
that construct the initial G-list from an initially uniform quadtree data structure. Moreover it adds (deletes) element pointers from the G-list as elements
are created (deleted) from the quadtree. Another module accessing the data
structure of the previous ones has the role of creating the triangular and
rectangular element grids. It contains denitions for creating triangular and
rectangular elements data structures and routines that search the quadtree,
building the linked lists of triangles and rectangles that make up these grids.
The second class of functions is concerned with updating of the nite
element grids. With respect to this point the nal modul runs the integration
of (A.33) and (A.34). This module performs four main processes:
1. It maps the internal element node numbers to the indices of a global
solution vector.
2. It advances the T and eld vectors by Nr time steps on the nite
element grids dened above.
3. It calculates an error estimate for each element of the quadtree, based on
an error estimate of the quadrilateral elements.
4. It invokes the routines to rene the grid according to the error estimator.
Steps (1)-(4) are repeated until a sucient time evolution of the microstructure is established. The variable Nr is set, so that the interface remains within the regions of ne meshing between regriddings. The latter is
typically done every 100 time steps. Step (1) involves searching for all elements and the respective neighbors, as well as assigning each node a unique
number, if it has a counterpart on a global solution vector.

158

A. Numerical Issues of Diuse Interface Modeling

The nite element discretization of (A.33) and (A.34) is done using


Galerkins weighted residual method [78]. The method begins by assuming
that and T are interpolated within an element as
N

e =

ei Ni (x, y)

i=1

Te =

Tie Ni (x, y)

(A.37)

i=1

where ei and Tie are the eld values at the N nodes of the element e and
their interpolated values in its interior. The functions Ni (x, y) are standard
linear interpolation functions appropriate for the used element [303]. They
satisfy
Ni (xj , yj ) = i,j ,
(A.38)
where i,j is the Kronecker delta. Rewriting the dierential equations for
in (A.33) and (A.34) as L = 0, as well as of the T -equation as LT T = 0,
the Galerkin method requires that

(A.39)
Ni (x, y)L e (x, y)dxdy = 0
e

Ni (x, y)LT T e (x, y)dxdy = 0 ,
e

for i = 1, 2, 3, . . . , N , where e represents the area of an element e. Substituting (A.37) into (A.39), one obtains two linear algebraic equations for i
and Ti , i = 1, 2, 3, . . . , N in the element e.
The next step is to dene e = (1 , 2 , 3 , , N )T and Ye = (T1 , T2 , T3 ,
, TN )T , where the superscript T denotes the transpose, making e and Te
column vectors. Then the linear algebraic statement of the nite element
form of (A.33) and (A.34) reads


e
+ E en + Fe ()
e ) d = A
C(
dt
dTe
1 de
= DATe + C
,
C
dt
2 dt

(A.40)

A, A
and E and the vector Fe () are given by
where the matrices C, C,

C=
NT Ndxdy ,
(A.41)


=
C


A=
=
A


e

NT NA2 ((e ))dxdy ,



NT Nx + NT Ny dxdy ,

(A.43)


NT Nx + NT Ny A2 ((e ))dxdy ,

(A.44)

(A.42)


E=

A.3 Selected State-of-the-Art Numerical Approaches


NT Nx NT Ny A((e ))((e ))dxdy ,
Fe () =


e

NT f (e , U e ; )dxdy .

159

(A.45)
(A.46)

Here Nx , Ny denote the partial derivatives of the vector of shape functions


with respect to x and y, respectively. The function A((e )) is given by
(A.35). () is proportional to the derivative of A(). It reads:
((e )) = 16

tan (1 tan2 )
.
(1 + tan2 )2

(A.47)

[78] is employed. In
A lumped formulation for the matrices C and C
this procedure, the row vector of shape functions N in (A.41) is replaced by
the identity row vector I = (1, 1, 1, ). The resulting matrix C consists of
identical columns, each of which contains the element Ni (x, y) at the position
of the ith row. A lumped term is dened as a diagonal matrix of entries which
take the values
nodes

1
Lc =
Ni (x, y)dxdy .
(A.48)
nodes i=1 e
The use of a lumped matrix for C enables the assembly of a diagonal matrix
for the left hand side of (A.40), stored as a one-dimensional vector. (Twodimensional arrays would be required if the consistent formulation for the
assembly of the C matrices were used. However, microstructures evolving at
low undercooling can produce interfaces with over 2 105 elements, making
the storing of 2 105 2 105 matrices impossible.)
The global (obtained after assembly of the element equations in (A.40))
is time-stepped using a forward dierence (explicit) time scheme. For each
time step of the eld, the global T eld is solved iteratively using a Crank
Nicholson scheme. Convergence of T is obtained within a few iterations.
Regridding is based on an error estimator function discussed next. It
is obtained following Zienkiewicz and Zhu [303], based on the dierences
between calculated and smoothed gradients of the and T elds. In particular
a composite eld
= + T
(A.49)
is dened, for which is a constant. This denition enables a regridding in
accordance to the requirements of both the and T eld, as opposed to using
merely the gradients of the -eld in establishing the grid [39]. Since and
T have to be determined rather than their gradients, one does not expect the
gradient of to be continuous across element boundaries, due to the order
of the interpolation used. Thus the dierence between the calculated and the
smoothed gradients, which are both continuous across element boundaries,

160

A. Numerical Issues of Diuse Interface Modeling

can be expected to provide a reasonable estimate of error. This method appropriately meshes regions of both steep gradients and regions where the
and T elds change rapidly.
The error estimator function e is dened as
e = qs qc ,

(A.50)

where qc and qs are the calculated and smoothed gradients of , respectively.


To determine qs , one assumes it to be interpolated in the same way as the
and T elds, namely
qs = NQs ,
(A.51)
where N is the row vector of element shape functions and Qs a 4 2 matrix,
which columns represent the nodal values of uxes of in the x and y direction, respectively. To nd Qs , Galerkins method is employed, minimizing
the weighted residual


(A.52)
NT (NQs qc )d = 0 .
NT ede =
e

This calculation is simplied by lumping the left hand side of (A.52), leading
to



e

NT Id Qs =

NT qc d .

(A.53)

Assembling (A.53) for all quadrilateral elements yields an equation for the
smoothed gradients Qg of the global eld , at all element nodes, of the form
DQg = b ,

(A.54)

where D is a diagonal matrix, due to masslumping, and Qg is a N 2


matrix for the global, smoothed ux.
For the actual error updating on the elements of the quadtree one employs
the normalized error

|(qs qc )|2 de
2
e

.
(A.55)
Ee = 
2
e |qs | d
The domain of integration in the denominator denotes the entire domain
of the problem. Thus Ee2 gives the contribution of the local element error
relative to the total smoothed error calculated over the entire grid.
The CPU time of this approach scales with the arc length of the problem
under investigation. For most computations this implies an increase in computational eciency compared to computations on a uniform grid, which scale
with the size of the required grid itself. E.g., for = 0.70, D = 2, dt = 0.016,
chosen to simulate = 0 and system size 800 800 with xmin = 0.4,
simulations of 105 time steps take approximately 15 CPU-hours on a Sun
UltraSPARC 2200 workstation with 512 MHz.

A.3 Selected State-of-the-Art Numerical Approaches

161

A.3.2 3D Simulations Including Fluid Flow


The phase-eld model underlying the computational approach presented
within this section employs the same phase-eld model equations as the previous section. To extend it to hydrodynamic ow in the liquid phase a method
proposed by Beckermann et al. [26] is employed. Within this approach a phase
average k of a variable for phase k over volume V is dened as

2 V Xk dV
k =
,
(A.56)
1
where Xk {0, 1} is an existence function. The formulation ensures that the
uid velocity is extinguished in the solid, and further that the shear stress at
the liquidsolid interface is handled correctly.
The governing equations for simulating dendritic growth with uid ow
are the mixture continuity equation:


1

u =0,
(A.57)
2
where u is the velocity vector, the averaged momentum equation:



 

1
1
1

u +u
u +
p
t
2
2
2


h2 (1 2 )(1 + )
1
= 2
u
u,
(A.58)
2
8 2

where t is time, p the pressure, the kinematic viscosity, = W0 / 2 the


characteristic interface width, and h a constant (=2.757), which ensures that
the interface shear stress is correct for a simple shear ow (see Beckermann
et al. [26]), and the averaged energy conservation equation in terms of
the dimensionless temperature T :


T
1
1
+
u T = D2 T +
.
(A.59)
t
2
2 t
The phase averages of velocity and pressure are used for deriving the mixture continuity equation, averaged liquid momentum equation and averaged
energy conservation equation. These are solved together with the 3D phaseeld evolution equation, which is a straightforward extension of (A.34).
Here the 3D ow equations are solved using the SIAPM by Gresho [136].
SIAPM is a predictor-corrector method, which can solve (A.57) and (A.58)
eectively, especially for large 3D problems, because it uses relatively small
amounts of memory. The velocity degrees of freedom are solved in a segregated form. Pressure is updated using a projection method. For a detailed
discussion of the algorithm, the reader is referred to the original paper [136].
The algorithm consists of three steps, which are summarized briey in the
following:

162

A. Numerical Issues of Diuse Interface Modeling

n+1 from
1. Compute an intermediate velocity u




1
1
1
1
n+1
M K+F u
M
+
K
uni
=
i
t
2
t
2
A(un )uni Gi pn ,

(A.60)

n+1
is the vector of nodal values of the intermediate velocity
where u
i
component i at time step n + 1, uni is the corresponding vector at time
step n, and pn is the vector of nodal pressures at time step n. The
coecient matrices are dened in terms of the velocity shape functions
N as follows:

(1 ) T
(A.61)
M =
N Nd
2




( 1) NT N
NT N
d
(A.62)
K =

+I
x x
xk xk
2


h(1 2 )(1 + ) T
F =
N Nd
(A.63)
8 2


(1 ) T n N
N uk
A(un ) =
d
(A.64)
2
xk


(1 ) T N
Gi =
d .
(A.65)
N
2
xi

2. The velocity eld found in the rst step is generally not divergencefree. The next step corrects the pressure to obtain an approximately
divergence-free velocity eld by solving a Poisson equation for
pn+1 = pn+1 pn :
Lpn+1 =
where

 n+1

1

D u
un ,
t

(A.66)

(1 ) NT N
I
d
2
xk xk

(1 ) T N
N
d .
D=
2
x

L =

(A.67)
(A.68)

pn+1 is then updated according to


pn+1 = pn + pn+1 .

(A.69)

3. Finally, the projected velocity un+1 is computed in a corrector step by


solving
n+1 tML 1 Gpn+1 ,
un+1 = u
(A.70)
where


ML =

1
Nd .
2

(A.71)

A.3 Selected State-of-the-Art Numerical Approaches

163

The computations are started using an initial velocity eld u0 determined


from the boundary conditions. In order to obtain a eld u0 , whose discrete
divergence is close to zero, the pressure is adapted by iteration. The initial
pressure is zero everywhere.
Equations (A.60) and (A.66) are solved by the conjugate gradient (CG)
method with diagonal preconditioning [217]. The SIAPM can calculate the
velocity eld for large 3D problems much faster than fully implicit timestepping methods, because convergence for (A.60) is reached within a few
iterations. Moreover, the number of degrees of freedom of (A.66) is one. The
CG iteration for (A.66) converges more slowly, typically within 50-200 iterations.
The averaged energy equation (A.59) is solved using the CG method with
diagonal preconditioning as well. Streamline upwind schemes [45] are employed for the convection terms in (A.60) and (A.59). The 3D phase-eld
equation is a nonlinear system. In order to solve the system implicitly, an
iterative method such as the NewtonRaphson method is required. Alternatively an explicit time-stepping scheme can be employed. Stable solutions are
obtained from the explicit scheme, since the variation of the eld exists
only in the interface region and a suciently small time increment t can
be used.
In analogy to the previous section the algorithm is implemented along
with adaptive grid renement. Moreover it is parallelized on the basis of the
Charm++FEM framework [158].
Based on this implementation in [155] for a set of problem parameters
corresponding to SCN, i.e. t = 0.016, W = 1, 0 = 1, = 6.383, d0 (=
a1 W
) = 0.139, = 92.4 and P r = 23.1 computations of 100 time steps
of a growing 3D dendrite are reported to require 190 hours of CPU for an
underlying minimum grid spacing of xmin = 0.4. 80% to 90% of the CPU
time were spent for the update of the velocity and the pressure eld [155].
To my knowledge these 3D simulations of SCN including uid ow are the
most extensive computations of dendritic growth so far. Everything beyond
still remains an open challenge from point of view of numerics as well as
from point of view of modeling, which might help to simplify numerics.

References

1. T. Abel, E. Brener, H. M
uller-Krumbhaar: Phys. Rev. E 55, 7789 (1997)
2. T. Abel: Phasenfeldmodelle f
ur Wachstumsprozesse in Multikomponenten- und
Multiphasensystemen. PhD Thesis, Research Center J
ulich, J
ulich (1998)
3. Handbook of Mathematical Funktions ed. by M. Abramowitz, I.A. Stegun, 10th
ed. (Wiley, New York 1972)
4. G. Adam, O. Hittmair: W
armetheorie, 3rd ed. (Vieweg, Braunschweig 1988)
5. S. Akamatsu, G. Favier, T. Ihle: Phys. Rev. E 51, 4751 (1995)
6. M. Albrecht, S. Christiansen, J. Michler, W. Dorsch, H.P. Strunk, P.O. Hansson, E. Bauser: Appl. Phys. Lett. 67, 1232 (1995)
7. S.M. Allen, J.W. Cahn: Acta Metall. 27, 1085 (1979)
8. F.J. Almgren, J.E. Taylor, L. Wand: SIAM J. Cont. Opt. 31, 386 (1993)
9. R. Almgren, W. S. Dai, V. Hakim: Phys. Rev. Lett. 71, 3461 (1993)
10. R. Almgren: J. Comp. Phys. 106, 337 (1993)
11. R. Almgren: SIAM J. Appl. Math. 59, 2086 (1999)
12. D.M. Anderson, G.B. McFadden, A.A. Wheeler: Phys. Fluids 9, 1870 (1997)
13. D.M. Anderson, G.B. McFadden, A.A. Wheeler: Ann. Rev. Fluid Mech. 30,
139 (1998)
14. D.M. Anderson, G.B. McFadden, A.A. Wheeler: NISTIR 6237 (1998)
15. D.M. Anderson, G.B. McFadden, A.A. Wheeler: Physica D 135, 175 (2000)
16. D.M. Anderson, G.B. McFadden, A.A. Wheeler: Physica D 144, 154 (2000)
17. D.M. Anderson, G.B. McFadden, A.A. Wheeler: In Interfaces for the TwentyFirst Century, ed. by M.K. Smith, M.J. Miksis, G.B. McFadden, G.P. Neitzel,
D.R. Canright (Imperial College Press, London 2002), p. 131
18. D.M. Anderson, G.B. McFadden, A.A. Wheeler: Physica D 151, 305 (2001)
19. R. Aris: Vector, Tensors, and the Basic Equations of Fluid Dynamics (Dover,
New York 1962), p. 84
20. G.S. Bales, A. Zangwill: Phys. Rev. B 41, 5500 (1990)
21. A.-L. Barabasi, H. E. Stanley: Fractal Concepts in Surface Growth, 1st
ed. (Cambridge University Press, Cambridge 1995)
22. M.Barber, A.Barbieri, J.S. Langer: Phys. Rev. A 36, 3340 (1987)
23. M.N. Barber, D. Singleton: J. Math. Soc. B 36, 325 (1994)
24. W. Blanke (ed.): Thermophysikalische Stogr
oen, 1st ed. (Springer-Verlag,
Berlin 1989), p. 155
25. J. Bechhoefer: Directional Solidication at the Nematicisotropic Interface.
PhD Thesis, The University of Chicago, Chicago (1988)
26. C. Beckermann, H.-J. Diepers, I. Steinbach, A. Karma, X. Tong: J. Comp.
Phys. 154, 468 (1999)
27. M. Ben Amar, B. Moussalam: Phys. Rev. Lett. 60, 317 (1988)
28. M. Ben Amar, P. Pelce: Phys. Rev. A 39, 4263 (1989)
29. E. Ben-Jacob, N. Goldenfeld, J.S. Langer, G. Sch
on: Phys. Rev. A 29, 330
(1984)

166

References

30. E. Ben-Jacob, N. Goldenfeld, J.S. Langer, G. Sch


on: Phys. Rev. Lett. 51, 1930
(1983)
31. C.M. Bender, S.A. Orszag: Advanced Mathematical Methods for Scientists and
Engineers (McGraw-Hill, New York 1978)
32. K.W. Benz: private communication (Freiburg 2002)
33. J.J. Binney: The Theory of Critical Phenomena: An Introduction to the Renormalization Group, 1st repr. with corr. (Clarendon, Oxford 1999)
34. P. Bouissou, B. Perrin, P. Tabeling: Phys. Rev. A 40, 509 (1989)
35. W.J. Boettinger, J.A. Warren: Met. Trans. A 27, 657 (1996)
36. W.J. Boettinger, J.A. Warren: J. Cryst. Growth 200, 583 (1999)
37. A. B
osch: Phasenfeld-Modelle f
ur Mehrkomponentensysteme. Diploma Thesis,
Research Center J
ulich, J
ulich (1993)
38. A. B
osch, H. M
ullerKrumbhaar, O. Shochet: Z. Phys. 97, 367 (1995)
39. R.J. Braun, B.T. Murray: J. Cryst. Growth 174, 41 (1997)
40. R.J. Braun, J.W. Cahn, G.B. McFadden, H.E. Rushmeier, A.A. Wheeler: Acta
Mater. 46, 1 (1998)
41. A. J. Bray: Adv. Phys. 32, 357 (1994)
42. E. Brener, V. I. Melnikov: Adv. Phys. 40, 53 (1991)
43. E. Brener: Phys. Rev. Lett. 71, 3653 (1993)
44. E. Brener, D. Temkin: Phys. Rev. E 51, 351 (1995)
45. A.N. Brooks, T.J.R. Hughes: Comp. Methods Appl. Mech. Eng. 32, 199 (1982)
46. R. Brower, D. Kessler, J. Koplik, H. Levine: Phys. Rev. Lett. 51, 1111 (1983)
47. W.K. Burton, N. Cabrera, F.C. Frank: Phil. Trans. Roy. Soc. 243, 299 (1951)
48. G. Caginalp, P. Fife: Phys. Rev. B 33, 7792 (1986)
49. G. Caginalp: Arch. Rat. Mech. Anal. 92, 205 (1986)
50. G. Caginalp: SIAM J. Appl. Math. 48, 506 (1988)
51. C. Caginalp, J. Jones: App. Math. Lett. 4, 79 (1991)
52. G. Caginalp, X. Chen: In On The Evolution Of Phase Boundaries, ed. by M.E.
Gurtin and G.B. McFadden (Springer, New York, 1992), p. 1
53. G. Caginalp, W. Xie: Phys. Rev. E 48, 1897 (1993)
54. J.W. Cahn, J. E. Hilliard: J. Chem. Phys. 28, 258 (1958)
55. J.W. Cahn, J.E. Hilliard: J. Chem Phys. 31, 688 (1959)
56. J.W. Cahn: Acta Metall. 8, 554 (1960)
57. J.W. Cahn: Acta Metall. 9, 795 (1961)
58. J.W. Cahn, D.W. Homann: Acta Met. 22, 1205 (1974)
59. J.W. Cahn, W.C. Carter: Met. Trans. 27A, 1431 (1996)
60. J.W. Cahn, P. Fife, O. Penrose: Acta Mater. 45, 4397 (1997)
61. R.W. Cahn, P. Haase: Physical Metallurgy, 4th ed. (North Holland, Amsterdam 1996)
62. R.W. Cahn, P. Haase, E.J. Kramer: Materials and Technology, 1st ed. (VCH,
Weinheim 1991)
63. D. Caraballo: A Variational Scheme for the Evolution of Polycrystals by
Curvature. PhD Thesis, Department of Mathematics, Princeton University
(Princeton, 1997)
64. B. Caroli, C. Caroli, B. Roulet: In Solids Far From Equilibrium, ed. by G.
Godr`eche (Cambridge University Press, Cambridge 1992), p. 155
65. M. Carrard, M. Gremaud, M. Zimmermann, W. Kurz: Acta Metall. 40, 983
(1992)
66. Y.C. Chang, T.Y. Hou, B. Merriman, S. Osher: J. Comput. Phys. 124, 449
(1996)
67. Ch. Charach, P.C. Fife: SIAM J. Appl. Math. 58, 1826 (1998)
68. Ch. Charach, P.C. Fife: Open Syst. Inf. Dyn. 5, 99 (1998)
69. Ch. Charach, C.K. Chen, P.C. Fife: J. Stat. Phys. 95, 1141 (1999)

References

167

70. S. Chen, B. Merriman, S. Osher, P. Smereka: J. Comput. Phys. 135, 8 (1997)


71. S. Chen, B. Merriman, M. Kang, R.E. Caisch, C. Ratsch, L.-T. Cheng, M.
Gyure, R.P. Fedkiw, C. Anderson, S. Osher: J. Comput. Phys. 167, 475 (2001)
72. D.L. Chopp: J. Comput. Phys. 162, 104 (2000)
73. S. Clarke, D.D. Vvendensky: Phys. Rev. Lett. 58, 2235 (1987)
74. J.B. Collins, H. Levine: Phys. Rev. B 31, 6119 (1985)
75. M. Conti: Phys. Rev. E 56, 3197 (1997)
76. M. Conti: Phys. Rev. E 55, 701 (1997)
77. M. Conti: Phys. Rev. E 55, 765 (1997)
78. R. Cook, D. Malkus, M. Plesha: Concepts and Applications of Finite Element
Analysis, 3rd ed. (Wiley, New York 1989)
79. J. B. Collins, H. Levine: Phys. Rev. B 31, 6119 (1985)
80. S.R. Coriell, D. Turnbull: Acta Metall. 30, 2135 (1982)
81. S.R. Coriell, R.S. Sekerka: J. Cryst. Growth 61, 499 (1983)
82. S.R. Coriell, R.F. Boisvert, G.B. McFadden, L.N. Brush, J.J. Favier: J. Cryst.
Growth 140, 139 (1994)
83. S. R. Coriell, G. B. McFadden, R. F. Sekerka and W. J. Boettinger: J. Cryst.
Growth 191, 573 (1998)
84. D.P. Corrigan, M.B. Koss, J.C. LaCombe, K.D. de Jager, L.A. Tennenhouse,
M.E. Glicksman: Phys. Rev. E 60, 7217 (1999)
85. M.E. Curtin, D. Polignone, J. Vinals: Math. Models. Methods Appl. Sci. 6,
815 (1996)
86. A.G. Cullis, D.J. Robbins, A.J. Pidduck, P.W. Smith: Mater. Res. Soc. Symp.
Proc. 280, 383 (1993)
87. A.G. Cullis: Mater. Res. Soc. Symp. Proc. 399, 303 (1996)
88. J.A. Dantzig, L.S. Chao: In Proc. 10th U.S. Nat. Cong. of Appl. Mech., ed.
by J. Lamb (1986), p. 249
89. S.H. Davis: Theory of Solidication, 1st ed. (Cambridge University Press,
Cambridge 2001)
90. S. de Cheveigne, C. Guthmann, M.M. Lebrun: J. Phys. Paris 47, 2095 (1986)
91. S.R. de Groot, P. Mazur: Non-Equilibrium Thermodynamics, 1st ed. (Dover,
New York 1984)
92. P.R.B. Devloo: An H-P Adaptive Finite Element Method for Steady Compressible Flow. Ph.D Thesis, The University of Texas at Austin, Austin, TX
(1987)
93. C. Domb, M.S. Green: Phase Transitions and Critical Phenomena, Vol. 5
(Academic Press, London 1976)
94. W. Dorsch, S. Christiansen, M. Albrecht, P.O. Hansson, E. Bauser, H.P.
Strunk: Surf. Sci. 896, 331 (1994)
95. W. Dorsch, B. Steiner, M. Albrecht, H.P. Strunk, H. Wawra, G. Wagner: J.
Cryst. Growth 183, 305 (1998)
96. F. Drolet, K. R. Elder, Martin Grant, J. M. Kosterlitz: Phys. Rev. E 61, 6705
(2000)
97. D.J. Eaglesham, M. Cerullo: Phys. Rev. Lett. 64, 1943 (1990)
98. Ch. Eck, P. Knabner, S. Korotov: A two-scale model for the computation
of solid-liquid phase transitions with dendritic microstructure. Preprint No.
278, Institute for Applied Mathematics, University Erlangen-N
urnberg (2001)
99. W. Eckhaus: Matched Asymptotic Expansions and Singular Perturbations
(North Holland, Amsterdam 1973)
100. K R. Elder, F. Drolet, J.M. Kosterlitz, M. Grant: Phys. Rev. Lett. 72, 677
(1994)
101. K.R. Elder, J.D. Gunton, M. Grant: Phys. Rev. E 54, 6476 (1996)

168

References

102. K.R. Elder, M. Grant, N. Provatas, J.M. Kosterlitz: Phys. Rev. E 64, 021604
(2001)
103. The appendix of this chapter follows closely the appendices of Elder et al. [102].
104. G. Ehrlich: Surface Sci. 63, 422 (1977)
105. H.F. ElNashar, H.A. Cerdeira: Physica A 283, 6 (2000)
106. H. Emmerich, K. Kassner, C. Misbah, T. Ihle: J. Phys.: Condens. Matter 11,
9985 (1999)
107. H. Emmerich, D. Schleussner, K. Kassner, T. Ihle: J. Phys.: Condens. Matter
11, 8981 (1999)
108. H. Emmerich, K. Kassner, T. Ihle, A. Wei: J. Cryst. Growth 211, 43 (2000)
109. H. Emmerich: Phys. Rev. B 65, 233406 (2002)
110. H. Emmerich: accepted for publication in Continuum Mech. Therm.
111. H. Emmerich, M. Jurgk, R. da Silva de Siquieri: in preparation
112. M.A. Eshelman, V. Seetharaman, R. Trivedi: Acta Metall. 36, 1165 (1988)
113. M. Fabbri, V.R. Voller: J. Comp. Phys. 130, 256 (1997)
114. R. Fedkiw, T. Aslam, B. Merriman, S. Osher: J. Comput. Phys. 152, 457
(1999)
115. P.C. Fife: In CBMS-NSF Regional Conference Series in Applied Mathematics
53 (SIAM, Philadelphia 1988), p. 11
116. P.C. Fife, O. Penrose: Electron. J. Di. Equations 1, 1 (1995)
117. A detailed investigation of the classical problem of density change ow without generalized boundary conditions was carried out by T. Fischalek, Ottovon-Guericke University Magdeburg (2000). Results presented at the TMR
network Pattern, Noise and Chaos meeting Leiden, July 2000. The calculations presented in this section are based on this work.
118. M. Fisher: Rev. Mod. Phys. 46, 597 (1974)
119. G. Fix: Research Notes in Mathematics 2, (1983)
120. G.J. Fix: In Free Boundary Problems: Theory and Applications, Vol. II, ed. by
A. Fasano and M. Primicerio (Pitman, Boston 1983), p. 580
121. E. Fried, M.E. Gurtin: Physica D 72, 287 (1994)
122. E. Fried, M.E. Gurtin: Physica D 91, 143 (1996)
123. H. Gao: J. Mech. Phys. Solids 42, 741 (1994)
124. W. Gebhardt, U. Krey: Phasen
uberg
ange und kritische Ph
anomene, 1st
ed. (Vieweg, Braunschweig 1980)
125. A. Ghazali, C. Misbah: Phys. Rev. A 46, 5026 (1992)
126. J.W. Gibbs: T. Conn. Acad. 2, 382 (1873)
127. M.E. Glicksman: Mater. Sci. Eng. 65, 45 (1984)
128. M.E. Glicksman, S.P. Marsh: In Handbook of Crystal Growth, Vol. 1b, ed. by
D.T.J. Hurle (Elsevier, Amsterdam 1993), p. 1075
129. M.E. Glicksman, M.B. Koss, E.A. Winsa: Phys. Rev. Lett. 73, 573 (1994)
130. M. E. Glicksman: Microgravity News, NASA 4, 4 (1997)
131. M. E. Glicksman: private communication (Aachen 2002)
132. R. Glowinski, T.-S. Pan, J. Periaux: Comput. Methods Appl. Mech. Eng. 111,
283 (1994)
133. G.H. Godunov, V.S. Ryabenkii: Dierence Schemes: An Introduction to the
Underlying Theory (North Holland, Amsterdam 1987)
134. H. Goldstein: Klassische Mechanik, 11th ed. (Aula-Verlag, Wiesbaden 1978)
135. G.H. Golub, J.M. Ortega: Scientic Computing (Academic Press, Boston
1993)
136. P.M. Gresho, S.T. Chan, M.A. Christon, A.C. Hindmarsh: Int. J. Numer.
Methods Fluids 21, 837 (1995)
137. M. Griebel, T. Dornseifer, T. Neunhofer: Numerische Simulation in der
Str
omungsmechanik, 1st ed. (Vieweg, Braunschweig 1995)

References

169

138. B. Grossmann, K.R. Elder, M. Grant, J. M. Kosterlitz: Phys. Rev. Lett. 71,
3323 (1993)
139. J.D. Gunton, M. San Miguel, P. Sahni: In Phase Transitions and Critical
Phenomena, Vol.8, ed. by C. Domb, J.L. Lebowitz (Academic Press, London
1983), p. 267
140. I. Gyrmati: Non-Equilibrium Thermodynamics, 1st ed. (Springer, New York
1970)
141. Proceedings of the Workshop on Virtual Molecular Beam Epitaxy, ed. by M.F.
Gyure, J.J. Zinck, Comp. Mater. Sci. 6, 113 (1996)
142. R. Haase, H. Schoenert: SolidLiquid Equilibrium, 1st ed. (Pergamon, Oxford
1969)
143. B.I. Halpering, P.C. Hohenberg, S. Ma: Phys. Rev. B 10, 139 (1974)
144. S.C. Hardy: J. Cryst. Growth 69, 456 (1984)
145. D.W. Homann, J.W. Cahn: Surf. Sci. 31, 368 (1972)
146. P.C. Hohenberg, B.I. Halperin: Rev. Mod. Phys. 49, 435 (1977)
147. G. Horvay, J. W. Cahn: Acta Metall. 9, 695 (1961)
148. T. Hou, Z. Li, S. Osher, H.-K. Zhao: J. Comput. Phys. 134, 236 (1997)
149. S.-C. Huang, M. Glicksman: Acta Metall. 29, 1697 (1981)
150. T. Ihle, H. M
uller-Krumbhaar: Phys. Rev. E 49, 2972 (1994)
151. T. Ihle: Eur. Phys. J. B 16, 337 (2000)
152. S.E. Ingle, F.H. Horne: J. Chem. Phys. 59, 5882 (1973)
153. G. P. Ivantsov: Dokl. Akad. Nauk USSR 58, 1113 (1947)
154. D. Jasnow: Rep. Prog. Phys. 47, 1061 (1984)
155. J.H. Jeong, N. Goldenfeld, J. Dantzig: Phys. Rev. E 64, 041602 (2001)
156. R.A. Johnson, D.M. Belk: AIAA J. 33, 2305 (1995)
157. D. Juric, G. Tryggvason: J. Comput. Phys. 123, 127 (1996)
158. L.V. Kale, S. Krishnan: In Parallel Programming using C++, ed. by G.V.
Wilson, P. Lu (MIT Press, Massachusetts 1996), p. 175
159. N.S. Kalospiros, B.J. Edward, A.N. Beris: Int. J. Heat Mass Trans. 36, 1191
(1993)
160. N.S. Kalospiros, R. Ocone, G. Astarita, J.H. Meldon: Ind. Eng. Chem. Res.
30, 851 (1991)
161. S. Kaplun: In Fluid Mechanis and Singular Perturbations, ed. by P.A. Lagerstrom, L.N. Howard, C.S. Liu (Academic Press, New York 1967)
162. A. Karma, J.S. Langer: Phys. Rev. A 30, 3147 (1984)
163. A. Karma, B.G. Kotliar: Phys. Rev. A 31, 3266 (1985)
164. A. Karma, A. Sarkissian: Phys. Rev. E 47, 513 (1993)
165. A. Karma: Phys. Rev. E 48, 3441 (1993)
166. A. Karma: Phys. Rev. E 49, 2245 (1994)
167. A. Karma, W. Rappel: Phys. Rev. Lett. 77, 4050 (1996)
168. A. Karma, W.-J. Rappel: Phys. Rev. E 53, R3017 (1996)
169. A. Karma, W.-J. Rappel: Phys. Rev. E 57, 4323 (1998)
170. A. Karma, M. Plapp: Phys. Rev. Lett. 81, 4444 (1998)
171. A. Karma: Phys. Rev. Lett. 87, 115701 (2001)
172. K. Kassner, C. Misbah, J. M
uller, J. Kappe, P. Kohlert: Phys. Rev. E 63,
036117 (2001)
173. K. Kawasaki, T. Ohta: Prog. Theor. Phys. 67, 147 (1982)
174. K. Kawasaki, T. Ohta: Prog. Theor. Phys. 68, 129 (1982)
175. K. Kawasaki, T. Ohta: Physica 118A, 175 (1983)
176. K. Kawasaki, M.C. Yalabik, J.D. Gunton: Phys. Rev. A 17, 455 (1978)
177. L.J.T.M. Kemper: J. Chem. Phys. 115, 6330 (2001)

170

References

178. J. Kepler: Strena seu de nive sexangula, ed. by G. Tampach (Frankfurt 1611);
translated in A New Years Gift or on the SixCornered Snowake, ed. by C.
Hardie (Clarendon, Oxford 1966), p. 74
179. D. Kessler, J. Koplik, H. Levine: Phys. Rev. A 30, 3161 (1984)
180. D. Kessler, H. Levine: Europhys. Lett. 4, 215 (1987)
181. D. Kessler, J. Koplik, H. Levine: Adv. Phys. 37, 255 (1988)
182. Y.T. Kim, N. Provatas, N. Goldenfeld, J. Dantzig: Phys. Rev. E 59, R2546
(1999)
183. Y.T. Kim, N. Goldenfeld, J. Dantzig: Phys. Rev. B 62, 2471 (2000)
184. M.B. King: Phase Equilibrium Mixtures, 1st ed. (Pergamon Press, Oxford
1969)
185. J.S. Kirkaldy, D.J. Young: Diusion in the Condensed State (The Institute of
Metals, London 1987)
186. R. Kobayashi: Bull. Jpn. Soc. Ind. Appl. Math. 1, 22 (1991)
187. R. Kobayashi: Physica D 63, 410 (1993)
188. R. Kobayashi, J.A. Warren, W.C. Carter: Physica D 140, 141 (2000)
189. P. Kopczynski, W.-J. Rappel, A. Karma: Phys. Rev. Lett. 77, 3387 (1996)
190. D.J. Korteweg: Arch. Neerl. Sci. Exactes Nat. Ser. II 6 (1901)
191. M. Kruskal, H. Segur: Asymptotics Beyond all Orders in a Model of Dendritic
Crystals. In Aero. Res. Ass. of Princeton Tech. Memo (Princeton 1985)
192. R. Kupfermann: Morphology, Coexistence and Selection in interfacial Pattern
Formation. PhD Thesis, Tel Aviv University, Tel Aviv (1991)
193. J.C. LaCombe, M.B. Koss, M.E. Glicksman: Phys. Rev. Lett. 83, 2997 (1999)
194. J.C. LaCombe, M.B. Koss, J.E. Frei, C. Giummarra, A.O. Lupulescu, M.E.
Glicksman: Phys. Rev. E 65, 031604 (2002)
195. B. Lafaurie, S. Zaleski, G. Zanetti: J. Comp. Phys. 113, 134 (1994)
196. P.A. Lagerstrom, D.A. Reinelt: SIAM J. Appl. Math. 44, 451 (1984)
197. P.A. Lagerstrom: Matched Asymptotic Expansions, 1st ed. (Springer, New
York 1988)
198. Landoldt-B
ornstein, Vol. IV/5 D (Springer, New York 1993)
199. J.S. Langer: Rev. Mod. Phys. 52, 1 (1980)
200. J. Langer: In Directions in Condensed Matter Physics, ed. by G. Grinstein,
G. Mazenko (World Scientic, Singapore 1986), p. 164
201. J.S. Langer: Phys. Rev. A 36, 3350 (1987)
202. J. Langer: In Chance and Matter, Les Houches Session XLVI, ed. by J. Souletie,
J. Vannenimus, R. Stora (North Holland, Amsterdam, 1987), p. 629
203. Y.-W. Lee, R. Ananth, W.N. Gill: Chem. Eng. Comm. 153, 41 (1996)
204. F.K. LeGoues, M.C. Reuter, J. Terso, H. Hammar, R.M. Tromp: Phys. Rev.
Lett 73, 300 (1994)
205. M.-A. Lemieux, G. Kotliar: Phys. Rev. A 36, 4975 (1987)
206. R.J. LeVeque, Z. Li: SIAM J. Numer. Anal. 31, No. 4, 1019 (1994)
207. F. Liu, H. Metiu: Phys. Rev. E 49, 2601 (1994)
208. J. Lowengrub, J. Truskinovsky: Proc. Roy. Soc. Ser. A 454, 2617 (1998)
209. S. Luckhaus: Solidication of alloys and the GibbsThomson law. Preprint
No. 335, Institute of Mathematics, University of Bonn (1993)
210. B.Y. Lyubov: Kinetic Theory of Phase Transitions (Metallurgia Publishers,
Moskau 1978)
211. A. Madhukar, S. V. Ghaisas: CRC Crit. Rev. Solid State Mater. Sci. 14, 1
(1988)
212. G.B. McFadden, S.R. Coriell: J. Cryst. Growth 74, 507 (1986)
213. G.B. McFadden, A.A. Wheeler, R.J. Braun, S.R. Coriell, R.F. Sekerka: Phys.
Rev. E 48, 2016 (1993)
214. G.B. McFadden, A.A. Wheeler, D.M. Anderson: Physica D 144, 154 (2000)

References
215.
216.
217.
218.
219.
220.
221.
222.
223.
224.
225.
226.
227.
228.
229.
230.
231.
232.
233.
234.
235.
236.
237.
238.
239.
240.
241.
242.
243.
244.
245.
246.
247.
248.
249.
250.
251.
252.

171

B. Merriman, J. Bence, S. Osher: J. Comput. Phys. 112, 334 (1994)


H. Metui, Y.-T. Lu, Z.Y. Zhang: Science 255, 1088 (1992)
X. Mikic, E.C. Morse: J. Comp. Phys. 61, 154 (1985)
Y.-W. Mo, D.E. Savage, B.S. Schwartzentruber, M.G. Lagally: Phys. Rev.
Lett 65 1020 (1990)
P. Morse, H. Feshbach: Methods of Theoretical Physics, 1st ed. (McGraw-Hill,
New York 1953)
J. M
uller, M. Grant: Phys. Rev. Lett. 82, 1736 (1999)
H. M
uller-Krumbhaar, H. Emmerich, E. Brener, M. Hartmant: Phys. Rev. E
63, 026304 (2001)
W.W. Mullins: In Metal Surfaces: Structure, Energetics, and Kinetics,
ed. American Society for Metals (1962), p. 17
W.W. Mullins, R.F. Sekerka: J. Appl. Physics 35, 444 (1964)
M. Muschol, D. Lui, H. Z. Cummins: Phys. Rev. A 64 (1992) 1038
T. Ohta, D. Jasnow, K. Kawasaki: Phys. Rev. Lett. 49, 1223 (1982)
S. Ono, S. Kondo: Encyclopaedia of Physics 10, Springer (1960)
L. Onsager: Phys. Rev. 37, 405 (1931)
S. Osher, J.A. Sethian: J. Comput. Phys. 79, 12 (1988)
N. Palle, J. A. Dantzig: Met. Trans. A 27A, 707 (1996)
R.B. Pember, J.B. Bell, P. Colella: J. Comput. Phys. 120, 278 (1995)
O. Penrose, P.C. Fife: Physica D 43, 44 (1990)
O. Penrose, P.C. Fife: Physica D 69, 107 (1993)
C.S. Peskin: J. Comput. Phys. 25, 220 (1977)
R. Peters, J.S. Langer: Phys. Rev. Lett. 56, 1948 (1986)
O. PierreLouis, C. Misbah, Y. Saito, J. Krug, P. Politi: Phys. Rev. Lett. 80,
4221 (1998)
Y. Pomeau, M. Ben Amar: In Solids far from Equilibrium, ed. by Godr`eche
(Cambridge University Press, Cambridge 1991), p. 365
I. Prigogine: Non-Equilibrium Statistical Mechanics, 2nd ed. (Wiley, New York
1966)
N. Provatas, E. Elder, M. Grant: Phys. Rev. B 53, 6263 (1996)
N. Provatas, N. Goldenfeld, J. Dantzig: Phys. Rev. Lett. 80, 3308 (1998)
N. Provatas, N. Goldenfeld, J. Dantzig: J. Comp. Phys. 148, 265 (1999)
J.J. Quirk: ICASE Report No. 927, NASA Langley Research Center, Hampton, VA (1992)
C. Ratsch, P. Ruggerone, M. Scheer: In Surface Diusion: Atomistic and
Collective Processes, ed. by M.C. Tringides (Plenum, New York 1997), p. 83
W.H.S. Rouse: Great Dialogues of Plato (transl.) A Mentor Book, 1st reissue
ed. (Signet Classic, New York 1999)
J. S. Rowlinson: J. Stat. Phys. 20, 197 (1979) (an english translation of the
original Ph.D. thesis by van der Waals)
Y. Saito, G. Goldbeck-Wood, H. M
ullerKrumbhaar: Phys. Rev. A 38, 2148
(1988)
A. Schmidt: J. Comp. Phys. 125, 293 (1996)
L.Z. Schlitz, S. Garimella: J. Appl. Phys. 76, 4863 (1994)
D. Schwabe, A. Scharmann: Phys. Bl. 42, 352 (1986)
V. Seetharaman, M.A. Eshelman, R. Trivedi: Acta Metall. 36, 1175 (1988)
R.F. Sekerka, Z. Bi: Physica A 261, 95 (1998)
R.F. Sekerka: In Proceedings of the Microgravity Materials Science Conference
(Huntsville, Alabama 2000), p. 533
R.F. Sekerka, Z. Bi : In Interfaces for the Twenty-First Century ed. by M.K.
Smith, M.J. Miksis, G.B. McFadden, G.P. Neitzel, D.R. Canright (Imperial
College Press, London 2002), p. 147

172

References

253. J.A. Sethian: Proc. Nat. Acad. Sci. 93, 1591 (1996)
254. J.A. Sethian: Level Set Methods and Fast Marching Methods, 2nd ed. (Cambridge University Press, Cambridge 1999)
255. M. S. Shephard, P. L. Baehmann, K. R. Grice: Comm. App. Num. Meth. 4,
379 (1988)
256. W. Shyy: Computational Modeling for Fluid Flow and Interfacial Transport
(Elsevier, Amsterdam 1994)
257. W. Shyy, H.S. Udaykumar, M.M. Rao, R.W. Smith: Computational Fluid
Dynamics with Moving Boundaries (Taylor & Francis, Washington, DC, 1996)
258. A.J. Simon, J. Bechhoefer, A. Libchaber: Phys. Rev. Lett. 61, 2574 (1988)
259. J. Smith: J. Comp. Phys. 39, 112 (1981)
260. K. Somboonsuk, J.T. Mason, R. Trivedi: Metall. Trans. 15A, 967 (1984)
261. B.J. Spencer, S.H. Davis, P.W. Voorhees: Phys. Rev. B 47, 9760 (1993)
262. D.J. Srolovitz: Acta Metallogr. 37, 621 (1988)
263. H.E. Stanley: Phase Transitions and Critical Phenomena, 1st ed. (Oxford
Univ. Press, New York 1971)
264. J.L. Steger: In Proceedings of the Computational Fluid Dynamics Symposium on Aeropropulsion, NASA CP3078 (Natl. Aeronautics & Space Admin.,
Washington, DC, 1991), p. 1
265. Z. Suo: Advances in Appl. Mechs. 33, 193 (1997)
266. M. Sussman, P. Smereka, S. Osher: J. Comput. Phys. 114, 146 (1994)
267. J.E. Taylor: Acta Met. 40, 1475 (1992)
268. J.E. Taylor, J.W. Cahn: Acta Met. 42, 1045 (1994)
269. J.E. Taylor, J.W. Cahn: J. Stat. Phys. 77, 183 (1994)
270. D. E. Temkin: Dokl. Akad. Nauk SSSR 132, 1307 (1960)
271. D. ter Haar: Collected Papers of Landau (Gordon and Breach, Washington,
DC, 1967)
272. R. T
onhardt, G. Amberg: J. Cryst. Growth 194, 406 (1998)
273. J.C. Tolendano, P. Tolendano: The Landau Theory of Phase Transitions
(World Scientic, Singapore 1987)
274. R. Trivedi: Metall. Trans. 15A, 977 (1984)
275. R. Trivedi, J.A. Sekhan, V. Seetharaman: Metall. Trans. A20, 769 (1989)
276. J.Y. Tsao: Materials Fundamentals of Molecular Beam Epitaxy (Academic
Press, Boston 1993)
277. H. S. Udaykumar, W. Shyy: Numer. Heat Transfer B 27, 127 (1995)
278. H.S. Udaykumar, W. Shyy: Int. J. Heat Mass Transfer 38, 2057 (1995)
279. H.S. Udaykumar, W. Shyy, M M. Rao: Int. J. Numer. Methods Fluids 22, 691
(1996)
280. H.S. Udaykumar, H.-C. Kan, W. Shyy, R. Tran-Son-Tay: J. Comput. Phys.
137, 366 (1997)
281. H.S. Udaykumar, R. Mittal, W. Shyy: J. Comput. Phys. 153, 535 (1999)
282. A. Umantsev, A.L. Roitburd: Sov. Phys. 30 651 (1988)
283. A.P. Umantsev: J. Chem. Phys. 96, 606 (1992)
284. A. Umantsev, S.H. Davis: Phys. Rev. A 45 7195 (1992)
285. S.O. Unverdi, G. Tryggvason: J. Comput. Phys. 100, 25 (1992)
286. V. Venkatakrshnan: AIAA J. 34, 533 (1996)
287. D.D. Vvendensky, A. Zangwill, C.N. Luse, M.R. Wilby: Phys. Rev. E 48, 852
(1993)
288. J. Villain: J. Phys. I 1, 19 (1991)
289. S.-L. Wang, R.F. Sekerka, A.A. Wheeler, B.T. Murray, S.R. Coriell, R.B.
Braun, G.B. McFadden: Physica D 69, 189 (1993)
290. S.-L. Wang, R.F. Sekerka: Phys. Rev. E 53, 3760 (1996)
291. J.A. Warren, W.J. Boettinger: Acta Metall. Mater. A 43, 689 (1995)

References

173

292. J.A. Warren, B.T. Murray: Mod. Simul. Mater. Sci. Eng. 4, 215 (1996)
293. J.D. Weeks, G.H. Gilmer: Adv. Chem. Phys. 40, 157 (1979)
294. H. Wenzel: In Physikalische Grundlagen metallischer Werkstoe, 13th IFFSpring School at Research Center Juelich (1982)
295. A.A. Wheeler, W.J. Boettinger, G.B. McFadden: Phys. Rev. A 45, 7424 (1992)
296. A.A. Wheeler, W.J. Boettinger, G.B. McFadden: Phys. Rev. E 47, 1893 (1993)
297. A.A. Wheeler, B.T. Murray, R.J. Schaefer: Physica D 66, 243 (1993)
298. A.A. Wheeler, G.B. McFadden, W.J. Boettinger: Proc. Royal Soc. London A
452, 495 (1996)
299. D.J. Wollkind, R.N. Maurer: J. Cryst. Growth 42, 24 (1977)
300. G. Wul: Z. Kristallogr. Mineral 34, 449 (1901)
301. L. Zhang, S. Garimella: J. Appl. Phys. 74, 2494 (1993)
302. H.K. Zhao, T. Chan, B. Merriman, S. Osher: J. Comput. Phys. 127, 179
(1996)
303. O. C. Zienkiewicz, J. Z. Zhu: Int. J. Numer. Meth. Eng. 24, 337 (1987)

Index

G-list 156
H 1 norm 48
L2 norm 48
triangular elements

156

ADI 121, 133


adiabatic approximation 133, 149
advective term 102
anities 33
AllenCahn equation 5, 40
alloy solidication 142, 143
anisotropic surface tension 9, 47, 104,
106
anti-phase domain coarsening 40
asymptotic analysis 4, 59
asymptotic matching method 59
attachment kinetics 73, 102
balance law formulations 49
binary alloy 55, 147
binary mixture 21
Boltzmann constant 75
boundary-value problem 60
bulk energies 55
bulk modulus 132
bulk region 2
buoyancy 120, 123
CahnHilliard equation 5, 39, 54
generalized anisotropic 48
viscous 40
CahnHomann capillary vector 47,
51
capillary action 3
capillary length 3, 37, 75, 108, 147
capillary tensor 98
chemical potential 21, 72, 84, 149
chemical rate constants 31
child element 155
coalescence 143
coarsening 135

coexistence regime 24, 79


computation 143
computational eciency 154
computational techniques 154
computational universality 151
conuent hypergeometric functions
114
conodal line 55
conservation of mass 97
conserved variable approach 28
constant mobility 39
constants of integration 64
constitutive relations 31, 52
continuity equation 33
continuity of tangential components
109
continuity of temperature 106, 110
continuum description 72
continuum equations 69
continuum eld models 2
convection limited growth 99
correlation length 2
CPU time 151, 160, 163
cross coupling eects 37
crystalline anisotropy 71
crystalline lms 69
crystallization 143
crystallization temperature 9, 106
curvature 9, 47, 67, 78, 102, 106, 110
curvilinear coordinates 84
cusp like solution 135
cylindric coordinates 111
dendritic growth 9, 78, 97
dendritic selection theory 104
dendritic solidication 3
dendritic tip velocity 142
density change ow 103, 104
depending variables 19
deposition time 140
desorptionless epitaxial growth 136
diuse interface 1, 8

176

Index

diuse interface model 39, 59, 78


for hydrodynamically inuenced
dendritic growth 98
diusion constant 9
diusion limited dendritic growth
104, 142
diusion limited growth 46, 99, 142
diusionrelaxation process 71
directional solidication 10
displacement eld 134
dissipative boundary force 120
dividing surface 7
doubletangent construction 27
doublewell potential 41, 74, 100, 152
doublon morphology 10
doublondendrite transition 10
Dufour eect 36
EhrlichSchwoebel eect 71
elastic coecients 31
elastic energy 132
elastic interaction 69
electrodeposition 99, 142
energy
conservation 110
density 28, 74, 100, 147
ux 73
entropic connement 70
entropy 32
density 45, 74, 100
ux 73
functional 45, 68, 143
production 49, 74, 99
epitaxial growth 5, 68, 99, 131
equilibrium thermodynamics 19, 31
error estimator function 159
Euler scheme 141
EulerLagrange equation 34, 53
EulerianLagrangian methods 14
eutectic alloy solidication 78, 143
extended domains 62
extension theorem 61
extensive quantities
conserved 33
non-conserved 35
extremal principle 21, 31
Fast Marching Method 16
nite surface 2
nite-size eects 123
rst principle considerations 39
rst-principle calculations 72
uctuations, thermal 122

free energy density 39, 132


free energy functional 2, 78
free-boundary problem 152
two-sided 75
Galerkin nite element method 155
Galilean transformation 98
generalized analysis 4, 59, 78, 81
Gibbs free energy 19, 21, 32, 99, 148
Gibbs surface 84
GibbsDuhem relation 9
GibbsThomson relation 9, 76, 91, 97,
102, 106, 110, 143, 146
governing variables 19
gradient expansion 33
gradient ow method 47
grain boundaries 144
grain growth 55
gravity 140
Greens function 81, 83, 115
growth rate 38
Helmholtz free energy 19, 44
Hermite polynomials 114
high performance computing 141
homogeneous heat conduction 113
homogenization 144
Hookes law 132
image processing 142
implementation 145
inhomogeneous heat conduction 115
inhomogeneous states 42
initial value problem 63
inner energy density 46
inner energy functional 68, 73
inner entropy functional 73
inner product 47
inner solution 67, 87
instantaneous relaxation 39
interfacial Peclet number 81, 152
irregular perturbation analysis 104
irregular perturbation expansion 118
irreversible thermodynamics 2, 31, 49,
140
island growth 69
island morphology 134
isothermal asymptotics 3
kinematic viscosity 106
kinetic attachment 71
kinetic coecient 147, 149
Kinetic Monte Carlo 69
kinetic principles 31

Index

177

Korteweg stresses 98
Kummers equation 113, 117

numerical issues 5, 145


numerically tractable models

Lagrangian parameter 33
LandauGinzburgWilson functional
42
LandauKhalatnikov ansatz 39
latent heat 9, 100, 106
lattice model 42
layer-by-layer growth 69
level set method 15
limit of absolute stability 12
liquid crystals 4
liquidus line 12, 23, 136
LPE 135
Lyapunov function 31

Onsager coecients 33
Onsager reciprocity relations 33
Onsager theory 31
order parameter 1, 39, 99, 144, 152
orderdisorder phase boundaries 144
order/disorder transitions 78
organic materials 127
Ostwald ripening 78, 143
outer solution 67, 87
overlap domains 64

MAC 122
Markovian system 33
mass balance 72, 76, 105, 109
masslumping 160
matching 59, 63, 118
MBE 135
mean-eld approximation 43
measurable quantities 153
melting temperature 100
mesh renement 155
metastable phase 55, 81
method of matched asymptotic
expansion 59, 67, 142
microgravity experiment 105, 120, 142
microscopic derivation 43
microsegregation 143
midpoint scheme 133
miscibility gap 12, 87
mist strain 132
mobility coecient 74
Model C 44
model systems 21
moving boundary problems 12, 31
Mullins surface diusion equation 48
multi-grid algorithm 145
multi-scale approach 144
multicomponent material 99
multicomponentmultiphase system
23
natural variables 19
NavierStokes equations 72, 97
Newtonian liquid 98
no-slip condition 97, 102, 106, 120, 122
non-classical ux 74
non-conserved variable approach 28
nucleation 1

131

parabolic coordinates 108


parabolic cylinder functions 114
paradigmatic problem 142
parallelization 141, 154
parasite current 122
parent element 155
Peclet number 108
per unit mass formula 100
per unit volume formula 100
periodic perturbation 37
perturbative analysis 126
phase diagram 19, 79
phase separation 55
phase transition phenomena 1
phase transitions 39, 143
phase variable 40
phase-eld 78
approach 132
equation 78
model 2, 28
parameters 148
variable 32, 38
phenomenological theory 5, 31, 45
Pochhammer symbol 114
Poisson equation 48, 162
Poisson ratio 132
polymer systems 142
Prandtl matching condition 62
Prandtl number 101, 107
preconditioning 163
predictor-corrector method 161
principle of minimal energy dissipation
33
projection operator 78, 115
quadtree element data structure 155
quasi-incompressible system 99
quench 79
rectangular elements

156

178

Index

reduced gravity 123


reduced phase diagram 24
regions of uniform validity 60
regular perturbation analysis 104
regular perturbation expansion 112
Reynolds transport theorem 50
rippled morphology 134
roughening 69
scalable algorithm 155
SCN 104
selected tip velocity 118
selection problem of dendritic growth
97
separation of variables 113
shape function 160, 162
sharp interface 1, 8
sharp interface approach 131
sharp interface limit 78, 89
sharp interface models 59
SIAPM 154, 161
side branching eects 143
silicon 104
smooth tip condition 113
solidliquid interface 97
solidication 47, 99
solidus line 12, 23, 136
solute 99, 136
solute trapping 143
solvability condition 77, 89, 118, 150
solvability theory 118
Soret eect 36
specic heat 9, 105, 146
spinodal decomposition 1, 40, 54, 78
spinodal line 55
state variables 19
steady state solutions 110
Stefan condition 9, 92, 97, 102, 146
Stefan problem 68
step ow growth 69
strain energy relaxation 134
strain tensor 132
strained surfaces 131
sub-element 155
subsidiary conditions 19
superconducting materials 142

supercooling 146
surface energy 55, 148
surface excess 7
surface quantity 1
surface tension 111, 118
terracestepkink model 7
thermocapillary convection 124
thermodynamic consistency 2, 28, 31,
45, 47, 73, 78, 145
thermodynamic equilibrium 19, 31,
33, 53
thermodynamic functional 31
thermodynamic potential 19, 28, 33,
38, 53
thermodynamic state 19
thermodynamic variables 53
thin lm growth 99
thin interface analysis 73, 81
thin interface limit 3, 89, 143
tip radius 107
TolmanBu eect 148
transparent materials 10
triple junction 57
two-sided epitaxial growth problem
131
undercooled melt 3, 146
universality 43, 46, 145, 151
validation 2, 75, 141
variational principles 2, 49
variational problem 53
velocity selection 103
virtual displacement 34
viscous stress tensor 52
vorticity
equation 107
scalar 107
weakly contaminated melt
Webers equation 114
WKB formalism 118
Wronskian 116
Youngs modulus

132

24

You might also like