Computational Materials Science 126 (2017) 139–151

Contents lists available at ScienceDirect

Computational Materials Science

journal homepage: www.elsevier.com/locate/commatsci

Editor’s Choice

Benchmark problems for numerical implementations of phase field

models
A.M. Jokisaari a, P.W. Voorhees a,b, J.E. Guyer c, J. Warren c, O.G. Heinonen d,e,⇑
a
Center for Hierarchical Materials Design, Northwestern University, 2205 Tech Drive, Evanston, IL 60208, USA
b
Department of Materials Science and Engineering, Northwestern University, 2220 Campus Drive, Evanston, IL 60208, USA
c
Material Measurement Laboratory, National Institute of Standards and Technology, 100 Bureau Drive, MS 8300, Gaithersburg, MD 20899-8300, USA
d
Northwestern-Argonne Institute of Science and Engineering, Evanston, IL 60208, USA
e
Materials Science Division, Argonne National Laboratory, Lemont, IL 60439, USA

a r t i c l e i n f o a b s t r a c t

Article history: We present the first set of benchmark problems for phase field models that are being developed by the
Received 8 July 2016 Center for Hierarchical Materials Design (CHiMaD) and the National Institute of Standards and
Received in revised form 13 September Technology (NIST). While many scientific research areas use a limited set of well-established software,
2016
the growing phase field community continues to develop a wide variety of codes and lacks benchmark
Accepted 14 September 2016
problems to consistently evaluate the numerical performance of new implementations. Phase field mod-
eling has become significantly more popular as computational power has increased and is now becoming
mainstream, driving the need for benchmark problems to validate and verify new implementations. We
Keywords:
Phase field model
follow the example set by the micromagnetics community to develop an evolving set of benchmark prob-
Benchmark problem lems that test the usability, computational resources, numerical capabilities and physical scope of phase
Spinodal decomposition field simulation codes. In this paper, we propose two benchmark problems that cover the physics of
Ostwald ripening solute diffusion and growth and coarsening of a second phase via a simple spinodal decomposition model
and a more complex Ostwald ripening model. We demonstrate the utility of benchmark problems by
comparing the results of simulations performed with two different adaptive time stepping techniques,
and we discuss the needs of future benchmark problems. The development of benchmark problems will
enable the results of quantitative phase field models to be confidently incorporated into integrated com-
putational materials science and engineering (ICME), an important goal of the Materials Genome
Initiative.
Ó 2016 Published by Elsevier B.V.

1. Introduction Two general mesoscale modeling approaches exist, with the

primary difference being how interfaces are handled [1–3].
Many important processes in materials microstructural evolu- Sharp-interface approaches, which treat interfaces as mathemati-
tion, such as coarsening, solidification, polycrystalline grain evolu- cally sharp, can be very efficient numerically when simulating
tion, and magnetic and ferroelectric domain formation and motion, the evolution of simple microstructural geometries. However,
occur on mesoscopic length and time scales. The ‘‘mesoscale” is the interface tracking with complex geometries (e.g., during dendritic
scale ‘‘in between;” in this case, in between atomistic scales of the growth) and topology changes, such as particles merging or split-
order of sub-nanometers and femto- to picoseconds, and macro- ting, pose significant numerical challenges [3]. Diffuse-interface
scopic scales of the order of micrometers and microseconds and approaches, in which the interface has a finite width, avoid these
larger. Mesoscale processes can strongly impact materials proper- issues [1–3]. However, they generally require more computational
ties and performance in engineering applications, providing strong resources because the diffuse interface, which often has a width of
motivation to develop accurate mesoscale microstructure evolu- a few nanometers, must be resolved even as other structural fea-
tion models. tures may have length scales in the hundreds of nanometers or
larger.
One popular diffuse-interface technique is the phase field
⇑ Corresponding author at: Materials Science Division, Argonne National approach, which has been used to study dendritic growth, spinodal
Laboratory, Lemont, IL 60439, USA. decomposition, grain growth, ferroelectric domain formation, and
E-mail address: [email protected] (O.G. Heinonen).

http://dx.doi.org/10.1016/j.commatsci.2016.09.022
0927-0256/Ó 2016 Published by Elsevier B.V.
140 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

other phenomena [1,2,4–9]. In a phase field model, a microstruc- of solution accuracy, solver optimizations, and code modifications.
ture is described by one or more continuous fields, uðr; tÞ. The While the phase field community ultimately needs validated
fields change smoothly over the computational domain and across experimental data sets to compare different models, we focus
interfaces. The field variables may be either a physical quantity, our effort here on first developing benchmark problems for numer-
such as composition or density, or a phenomenological descriptor ical implementations, which is a necessary precursor for the com-
[1]. Originally [10,11], the fields were used to denote a local phase parison of model results; a model cannot be validated in a useful
(hence the name phase field), with the value of u at position r and way until questions about the correctness of numerical implemen-
time t indicating the phase. For example, a two-phase system can tations are resolved. The micromagnetics community created
be described by a field u that takes the values ua and ub in the benchmark problems in the late 1990s to early 2000s to address
bulk a and b phases, respectively, while at the a/b interface, the a similar situation of multiple implementations and numerical
value of u changes smoothly over a finite width. The use of phase methods [45], and these problems are still evolving today. Bench-
field methods is now more diverse, with the phase field variable mark problems significantly aided the community in creating accu-
often representing other quantities or properties, such as concen- rate micromagnetics codes [45], such as the Object Oriented
tration or density. The evolution of existing phases within the sys- MicroMagnetics Framework (OOMMF) [46], MuMax3 [47], and
tem is driven by the reduction of the free energy, which is Magpar [48]. To aid in the development, validation, and verifica-
described as a functional of the field variables. Depending on the tion of phase field modeling software, the Center for Hierarchical
physics being modeled, the field variables may be conserved or Design (CHiMaD) and the National Institute of Standards and Tech-
non-conserved. Finally, ‘‘sharp-interface limit” or ‘‘thin-interface nology (NIST) are developing phase field benchmark problems.
limit” analyses have shown that phase field models are equivalent These problems are hosted on the NIST website [49] and are freely
to their analogous sharp-interface models when the interface available. In addition, NIST will also host the solutions to the prob-
width is significantly smaller than the size of other characteristic lems submitted by members of the phase field community so that
length scales (reviewed in Refs. [1,2]). For comprehensive descrip- the results from different implementations may be compared.
tions and reviews of phase field modeling, see Refs. [1,2,4–9]. Phase field benchmark problems for numerical implementa-
Quantitative phase field models have been developed to study tions should exhibit several key features, analogous to those in
technologically important phenomena in real materials systems the micromagnetics benchmark problems. First, the problems
as part of integrated computational materials engineering (ICME) should be nontrivial (i.e., not solvable without a computer) and
[12–14]. In ICME, models at different length scales are linked should exhibit differing degrees of computational complexity, yet
together to design materials for technological applications. A few not require extensive computational resources. Second, simulation
selected references of recent quantitative phase field studies outputs must be defined in such a way that results are easily com-
include solidification in Al alloys [15–17], precipitation in parable. In addition to snapshots or videos of the evolution of the
Ni-based superalloys [18–20], recrystallization in Ti [21] and Mg microstructure itself, the evolution of overall metrics such as the
[22] alloys, quantum dot formation in InAs/GaAs [23], and semicon- total energy of the system or the volume fraction of each phase
ducting core-shell nanoparticles [24]. The phase field approach con- should be quantified. Finally, the problems should test a simple,
tinues to be applied to novel materials systems and phenomena, targeted aspect of either the numerical implementation or the phy-
and a growing number of scientists are adopting the technique. sics. For example, simple physics could be used while complicated
The number of phase field software implementations is prolifer- domain or boundary conditions are tested, or coupled physics
ating with the growing application of phase field techniques, could be tested on a simple domain. Numerical aspects that must
necessitating a means of benchmarking, validating, and verifying be challenged include solver algorithms, mesh geometry, boundary
the numerical behavior of a diverse set of codes. Many research conditions, and time integration. Benchmark problems could be
domains which apply computational modeling have converged especially useful when examining multiphysics coupling, including
around a small number of standard pieces of software and bench- such behaviors as, e.g., diffusion, linear elasticity, fluid flow, aniso-
marking sets (e.g., COMSOL [25] and ABAQUS [26] for engineering tropic interfacial energy, and polarization.
simulations, or VASP [27–30], Quantum ESPRESSO [31], and the In this paper, we present a first set of community-driven,
G3/99 test set [32] for electronic structure calculations1), but this benchmark problems for numerical implementations of phase field
is not the case for the phase field community. A multitude of phase models and the efforts of NIST and CHiMaD to date. This first set of
field software implementations exist, and numerical approaches problems focuses on diffusion of a solute and phase separation; the
abound. Phase field simulations have been performed using open- second problem adds a coupled non-conserved order parameter.
source codes such as MOOSE [33,34], FEniCS [35,36], OpenPhase We discuss our choice of model formulations, parameterizations
[7], DUNE [37,38], FiPy [39,40], as well as with many proprietary and initial conditions so that these considerations may be kept in
codes, such as MICRESS [41,42], PACE 3D [43,44] and other in- mind while developing additional benchmark problems. Further-
house codes. Numerical implementations may employ finite differ- more, we demonstrate the utility of benchmark problems by com-
ence, finite volume, finite element, or spectral methods to solve paring simulation results obtained using two different time
the evolution equations, direct or spectral methods for solid adaptivity algorithms. We also briefly review lessons learned from
mechanics calculations, explicit or implicit time stepping, and adap- the first CHiMaD ‘‘Hackathon,” an event in which different phase
tive or non-adaptive meshing. To confidently incorporate quantita- field codes within the community were challenged against model
tive phase field results obtained from this wide variety of problems. Finally, we discuss the development of additional for-
numerical methods into ICME, both physical models and numerical mulations for the future, and encourage community involvement
implementations must be validated and verified. in the entire process of problem design, development, and report-
A set of standard benchmark problems allows the comparison ing of results.
of models, algorithms, and implementations, as well as the testing

2. Model formulations
1
Certain commercial equipment, instruments, or materials are identified in this
paper to foster understanding. Such identification does not imply recommendation or
endorsement by the National Institute of Standards and Technology, nor does it imply
In phase field models, field variables are evolved using dynam-
that the materials or equipment identified are necessarily the best available for the ics derived from generalized forces. The field variable is often ter-
purpose. med the ‘‘order parameter,” and we adopt that terminology here.
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 141

Most commonly, the time evolution is governed by dissipative where M is the mobility of the solute. For simplicity, both the
dynamics, in which the total free energy of the system decreases mobility and the interfacial energy are isotropic. We choose
monotonically with time (i.e., entropy increases at fixed tempera- ca ¼ 0:3; cb ¼ 0:7; .s ¼ 5; M ¼ 5, and j ¼ 2. Because the interfacial
ture). The order parameter may be locally conserved or non- energy, diffuse interface width, and free energy parameterization
conserved depending on what physical quantity or property the are coupled, we obtain the diffuse interface width of
pffiffiffiffiffiffiffiffiffiffiffi
order parameter represents, and its dynamics are defined by the l ¼ 7:071 j=.s ¼ 4:47 units over which c varies as
response of the system to a generalized force defined by the vari- pffiffiffiffiffiffiffiffi
0:348 < c < 0:652, and an interfacial energy r ¼ 0:01508 j.s
ation in the free energy. Kinetic coefficients, such as mobility or [52].
diffusivity, control how the order parameter responds to the force.
An example of a conserved order parameter is the concentration of 2.2. Ostwald ripening
solute in a matrix, while ferroelectric polarization is an example of
a non-conserved order parameter. The second benchmark problem examines Ostwald ripening in
The first problem in this benchmark set models spinodal a system with an ordered phase and a disordered phase; an exam-
decomposition via conserved dynamics, while the second models ple of this phenomenon in a real materials system is the growth
Ostwald ripening via coupled conserved/non-conserved dynamics. and coarsening of c0 precipitates in a c matrix in nickel-based
In this way, we focus on a single, fundamental aspect of physics superalloys [18,53]. This system is somewhat more complicated
(i.e., diffusion and phase separation) in the first problem, and then than that presented in Section 2.1, in that the microstructural evo-
increase the model complexity in the second problem. We discuss lution is driven by coupled conserved/non-conserved dynamics.
the motivation for each model formulation and the choice of initial However, the formulation is a simple extension of that in the pre-
conditions, boundary conditions, and computational domains. The vious section (note that we neglect elastic energy, an important
problems were formulated to be effectively two-dimensional so factor in c/c0 evolution).
that the essential physical behavior is modeled without making
the test problems unreasonably large or computationally 2.2.1. Free energy and dynamics
demanding. The atomic fraction of solute is again specified by the conserved
variable c, while the phase is indicated by a structural order param-
2.1. Spinodal decomposition eter, g. The structural order parameter is non-conserved and is a
phenomenological phase descriptor, such that the a phase is indi-
Spinodal decomposition is one of the oldest problems in the cated by g ¼ 0, while the b phase is indicated by g ¼ 1. If multiple
phase field canon, and its formulation in terms of continuum fields energetically equivalent orientation variants exist (for example,
goes back to the seminal works by Cahn and Hilliard [50]. The due to crystallographic symmetry considerations or ordered and
Cahn-Hilliard equation thus predates the name ‘‘phase field” in this disordered phases), the model may include p number of structural
context, but the term has subsequently been adopted by the com- order parameters, gp , with one for each orientation variant. We
munity. While spinodal decomposition may be one of the simplest include a nontrivial number of order parameters by setting p ¼ 4,
problems to model, it is highly relevant, as a large number of phase a value commonly used in superalloy models; this will stress the
field models include the diffusion of a solute within a matrix. Fur- numerical solver while not making the problem intractable.
thermore, precipitation and growth may also be modeled with the In this benchmark problem, the free energy of the system is
same formulation if the appropriate initial conditions are chosen. based on the formulation presented in Ref. [19] and is expressed as
For the benchmark problem, we select a simple formulation that !
Z X
p
is numerically tractable so that results may be obtained quickly jc jg
F¼ f chem ðc; g1 ; . . . gp Þ þ jrcj þ
2
jrgi j 2
dV ð4Þ
and interpreted easily, testing the essential physics while minimiz- V 2 2
i¼1
ing model complexity and the chance to introduce coding errors.
where jc and jg are the gradient energy coefficients for c and gi ,
2.1.1. Free energy and dynamics respectively. While the model in Ref. [19] follows the Kim-Kim-
For this benchmark problem of spinodal decomposition in a Suzuki (KKS) formulation for interfacial energy [54], we use the
binary system, a single order parameter, c, is evolved, which Wheeler-Boettinger-McFadden (WBM) [55] formulation for sim-
describes the atomic fraction of solute. The free energy of the sys- plicity. In the KKS model, the interface is treated as an equilibrium
tem, F, is expressed as [50] mixture of two phases with fixed compositions such that an arbi-
Z   trary diffuse interface width may be specified for a given interfacial
j
F¼ f chem ðcÞ þ jrcj2 dV; ð1Þ energy. In the WBM model, interfacial energy and interfacial width
V 2
are linked with the concentration, such that very high resolution
where f chem is the chemical free energy density and j is the gradient across the interface may be required to incorporate accurate inter-
energy coefficient. For this problem, we choose f chem to have a sim- facial energies.
ple polynomial form, The formulation for f chem in Ref. [19] is adapted for our bench-
mark problem as
f chem ðcÞ ¼ .s ðc  ca Þ2 ðcb  cÞ2 ; ð2Þ
a b
f chem ðc; g1 ; . . . gp Þ ¼ f ðcÞ½1  hðg1 ; . . . gp Þ þ f ðcÞhðg1 ; . . . gp Þ
such that f chem is a symmetric double-well with minima at ca and cb ,
and .s controls the height of the double-well barrier. Because f chem þ wgðg1 ; . . . gp Þ; ð5Þ
is symmetric (Fig. 1a), ca and cb correspond exactly with the equilib- a b
where f and f are the chemical free energy densities of the a and b
rium atomic fractions of the a and b phases.
phases, respectively, hðg1 ; . . . gp Þ is an interpolation function, and
Because c must obey a continuity equation – the flux of c is con-
served – the evolution of c is given by the Cahn-Hilliard equation gðg1 ; . . . gp Þ is a double-well function. The function h increases
[50], which is derived from an Onsager force-flux relationship [51]: monotonically between hð0Þ ¼ 0 and hð1Þ ¼ 1, while the function
g has minima at gð0Þ ¼ 0 and gð1Þ ¼ 0. The height of the double well
   barrier is controlled by w. We choose the simple formulation
@c @f chem
¼ r  Mr  jr2 c ð3Þ a
@t @c f ðcÞ ¼ .2 ðc  ca Þ2 ð6Þ
142 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

Fig. 1. The free energy density surfaces for (a) the spinodal decomposition problem, and (b) the Ostwald ripening problem for ðc; gÞ (not shown: ðgi ; gj Þ surface). The free
energy density surfaces are defined for all real values of c (both problems) and gi (Ostwald ripening problem) and not only over the intervals of interest, necessitating care in
choosing initial conditions (Section 2.3).

f ðcÞ ¼ .2 ðcb  cÞ2

b
ð7Þ almost every phase field model: the diffusion of solute (Section 2.1)
and the coupling of composition with a structural order parameter
X
p (Section 2.2). All of the model parameters chosen here are within a
hðg1 ; . . . gp Þ ¼ g3i 6g2i  15gi þ 10 ð8Þ few orders of unity, improving numerical performance. In addition,
i¼1
the structural order parameter in the second model is phenomeno-
p h
logical and varies within the interval of [0, 1]. This interval is cho-
X i X
p X
p
gðg1 ; . . . gp Þ ¼ g2i ð1  gi Þ2 þ a g2i g2j ; ð9Þ sen because multiphysics coupling that relies on the phase of the
i¼1 i¼1 j–i material is often incorporated by way of a structural order param-
eter (e.g., misfit strain of a precipitate phase with respect to a
a b
where f and f have minima at ca and cb ; .2 controls the curvature matrix phase).
of the free energies, and a controls the energy penalty incurred by Several trade-offs were considered between the free energy for-
the overlap of multiple non-zero gi values at the same point. mulations and the initial conditions. The free energy could be cho-
Because the energy values of the minima are the same (Fig. 1b), sen to be realistic, for example by using the CALPHAD method, or
ca and cb correspond exactly with the equilibrium atomic fractions to be more simplistic while still representing the main physics
of the a and b phases. and being numerically tractable. The CALPHAD method is a semi-
The time evolution of c is again governed by the Cahn-Hilliard empirical method of formulating free energies of mixing using
equation [50,56], known thermodynamic data and equilibrium phase diagrams
   [58,59]. While CALPHAD free energies are extremely useful, their
@c @f chem
¼ r  Mr  jc r2 c : ð10Þ functional form generally contains natural logarithms, which pose
@t @c several numerical and mathematical challenges for incorporation
The Allen-Cahn equation [57], which is based on gradient flow, into phase field models. Therefore, simple polynomial free energy
governs the evolution of gi , density formulations are chosen because they are numerically
  tractable and straightforward to implement. Ideally, the energy
@ gi dF @f chem
¼ L ¼ L  jg r2 gi ; ð11Þ formulation should be robust such that the system will tend to
@t dgi @ gi the equilibrium values of the phases no matter the initial condi-
where L is the kinetic coefficient of gi . We choose M ¼ 5 and L ¼ 5 tion. This behavior may be ensured by fixing the global minimum
so that the transformation is diffusion-controlled, and as in Sec- of the free energy density within the interval of interest. However,
tion 2.1, the kinetic coefficients and gradient energy coefficients many formulations do not exhibit these global minima, including
are isotropic. In addition, we again choose ca ¼ 0:3 and cb ¼ 0:3, the one in Section 2.2 (Fig. 1b). In addition, the characteristics of
pffiffiffi the free energy density surface are sensitive to the parameteriza-
and further specify kc ¼ kg ¼ 3; . ¼ 2; w ¼ 1, and a ¼ 5. For
tion of the model. For example, the local minima present at
these values, the diffuse interface between 0:1 < g < 0:9 has a
g ¼ 0; c ¼ 0:3 and g ¼ 1; c ¼ 0:7 become shallower as w
width of 4.2 units.
decreases. For certain values of w and a (e.g., w ¼ 0:1 and a ¼ 1),
the lowest energy occurs when all of the structural order parame-
2.3. Reasons for choices of models and parameters ters assume a value of approximately 0.9 in the b phase. This
behavior is due to the g2i g2j term in g. Finally, transient solute
The two benchmark problems presented here are simplified for-
depletion in the a phase, which may cause c to decrease below 0,
mulations designed to focus on fundamental aspects common to
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 143

may occur during the first several time steps of the simulation as 2.4.1. Spinodal decomposition
the system quickly relaxes from its initial conditions. Furthermore, The initial conditions for the first benchmark problem are cho-
Gibbs-Thomson-induced composition shift of the b phase may sen such that the average value of c over the computational
result in a composition greater than 1. Both behaviors are non- domain is approximately 0:5. The initial value of c for the square
physical if c is the atomic fraction of solute, but can occur within and ‘‘T” computational domains is specified by
the formulations in this paper because the free energy function is
cðx; yÞ ¼ c0 þ ½cosð0:105xÞ cosð0:11yÞ
defined even for non-physical solute concentrations. To avoid
these issues, the compositions of the a and b phases are chosen þ ½cosð0:13xÞ cosð0:087yÞ2 þ cosð0:025x  0:15yÞ
as intermediate values within the atomic fraction interval, and  cosð0:07x  0:02yÞ; ð12Þ
the initial conditions presented in Section 2.4 are formulated such
that the system will not exit the interval of 0 6 c 6 1 and where c0 ¼ 0:5 and  ¼ 0:01. In addition, the initial value of c for the
0 6 g 6 1. spherical computational domain is specified by

cðh; /Þ ¼ c0 þ sphere ½cosð8hÞ cosð15/Þ þ ðcosð12hÞ cosð10/ÞÞ2

2.4. Initial conditions, boundary conditions, and domain geometries
þ cosð2:5h  1:5/Þ cosð7h  2/Þ; ð13Þ
Several important factors were considered in determining the where sphere ¼ 0:05, and h and / are the polar and azimuthal angles,
initial conditions and computational domains of the benchmark
respectively, in a spherical coordinate system. These angles are
problems. First, the initial conditions for spinodal decomposition
translated into a Cartesian system as h ¼ cos1 ðz=rÞ and
and precipitation simulations are typically created with a pseudo-
/ ¼ tan1 ðy=xÞ dependent upon angle. The initial conditions speci-
random number generator. However, the initial conditions must be
fied by Eqs. (12) and (13) are shown in Fig. 2.
repeatable from implementation to implementation in a bench-
mark problem, precluding the use of pseudorandom number gen-
2.4.2. Ostwald ripening
eration. Therefore, we choose trigonometric functions to provide
The initial conditions for the Ostwald ripening problem are
smoothly varying, relatively disordered fields that are
qualitatively similar to those given for the spinodal decomposition
implementation-independent. Furthermore, the average composi-
problem, but the magnitude of the fluctuations around c ¼ 0:5 are
tion and the amplitude and width of the fluctuations must be cho-
greater, and fluctuations between [0, 1] are applied to the struc-
sen such that phase separation will occur, as opposed to the
tural order parameter fields. The initial condition for c is again
formation of a uniformly under- or supersaturated a phase. Finally,
given by Eq. (12) for the square and ‘‘T” domains, with c0 ¼ 0:5
the computational domain sizes and shapes are chosen to stress
and  ¼ 0:05, while it is given by Eq. (13) for the spherical domain,
the software implementation, because a wide variety of numerical
with c0 ¼ 0:5 and sphere ¼ 0:05. The initial condition for gi in the
methods are currently in use. The domain sizes and interface reso-
square and ‘‘T” domains is given as
lution requirements must be large enough that runtime should be
improved by parallel computing and mesh and time adaptivity, yet gi ðx; yÞ ¼ g fcosðð0:01iÞx  4Þ cosðð0:007 þ 0:01iÞyÞ
not so large as to require significant resources on a high- þ cosðð0:11 þ 0:01iÞxÞ cosðð0:11 þ 0:01iÞyÞ
performance computing cluster. We also anticipate that the use
þw½cosðð0:046 þ 0:001iÞx þ ð0:0405 þ 0:001iÞyÞ cosðð0:031
of non-rectilinear domains will become commonplace as new
applications of the phase field method are investigated, such as þ0:001iÞx  ð0:004 þ 0:001iÞyÞ2 g2 ; ð14Þ
nano-fabricated structures and cracking. Several phase field inves-
where g ¼ 0:1 and w ¼ 1:5, while for the spherical domain, it is
tigations (e.g., Refs. [36,60]) have already been performed with
given as
spherical domains.
Several boundary conditions, initial conditions, and computa- gi ðh; /Þ ¼ sphere
g fcosðih  4Þ cosðð0:7 þ iÞ/Þ
tional domain geometries are used to challenge different aspects þ cosðð11 þ iÞhÞ cosðð11 þ iÞ/Þw½cosðð4:6 þ 0:1iÞh
of the numerical solver implementation. For both benchmark prob-
þð4:05 þ 0:1iÞ/Þ cosðð3:1 þ 0:1iÞh  ð0:4 þ 0:1iÞ/Þ2 g2 ð15Þ
lems, we test four combinations that are increasingly difficult to
solve: two with square computational domains with side lengths with sphere
g ¼ 0:1 and i ¼ 1; . . . ; 4 enumerates the order parameters
of 200 units, one with a computational domain in the shape of a corresponding to the different phase variants. The initial conditions
‘‘T”, with a total height of 120 units, a total width of 100 units, for the Ostwald ripening simulations are shown in Fig. 3 for c; g1 ,
and horizontal and vertical section widths of 20 units (Fig. 2), and g2 .
and one in which the computational domain is the surface of a
sphere with a radius of r ¼ 100 units. While most codes readily
3. Numerical methods
handle rectilinear domains, a spherical domain may pose prob-
lems, such as having the solution restricted to a two-dimensional
To provide example solutions to the benchmark problems, the
curved surface. The coordinate systems and origins are given in
MOOSE computational framework is used. MOOSE [61,62] is an
Fig. 2. Periodic boundary conditions are applied to one square
open-source finite element framework and is the basis for several
domain, while no-flux boundaries are applied to the other square
other phase field applications, including Marmot [33] and Hyrax
domain and the ‘‘T”-shaped domain. Periodic boundary conditions
[63,64]. To avoid computationally expensive fourth-order deriva-
are commonly used with rectangular or rectangular prism domains
tive operators, the Cahn-Hilliard equation is split into the two
to simulate an infinite material, while no-flux boundary conditions
second-order equations [33,56], given by
may be used to simulate an isolated piece of material or a mirror
plane. As the computational domain is compact for the spherical @c
¼ r  ðM rlÞ ð16Þ
surface, no boundary conditions are specified for it. Note that the @t
same initial conditions are used for the square computational
and
domains with no-flux and periodic boundary conditions (Sections
2.4.1 and 2.4.2), such that when periodic boundary conditions @f chem
l¼  jr2 c; ð17Þ
are applied, there is a discontinuity in the initial condition at the @c
domain boundaries.
144 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

Fig. 2. The computational domains and initial conditions for the spinodal decomposition benchmark problem. The origin for the coordinate system of the sphere is at its
center.

Fig. 3. The computational domains and initial conditions for the Ostwald ripening benchmark problem. Top row: Initial conditions for the atomic fraction. Middle row: Initial
conditions for g1 . Bottom row: Initial conditions for g2 . Not shown: initial conditions for g3 and g4 .
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 145

where f chem and j are given in Sections 2.1.1 and 2.2.1 for the spin- the different time steppers may be directly compared at given
odal decomposition and Ostwald ripening problems, respectively. times. Figs. 4 and 5 show the total free energy of the different sim-
The square computational domains are meshed with square, four- ulations of spinodal decomposition and Ostwald ripening, respec-
node quadrilateral elements by the mesh generator within MOOSE, tively. In all cases, the total free energy decreases rapidly, then
while the ‘‘T”-shaped and spherical domains are meshed with trian- asymptotically approaches the local energy minimum of the sys-
gular three-node elements using CUBIT [65]. Linear Lagrange shape tem (which varies given the initial and boundary conditions).
functions are employed for c; gi , and l. For computational effi- While the starting and final free energies are the same for each
ciency, the system of nonlinear equations are solved with the full set of simulations (e.g., spinodal decomposition in the square
Newton method for the first problem, and the preconditioned Jaco- domain with no-flux boundary conditions), the evolution of the
bian Free Newton-Krylov (PJFNK) method for the second problem. energy is affected by the choice of the adaptive time stepper. In
The second backward differentiation formula (BDF2) [66] time inte- the spinodal decomposition problem, the differences in energy as
gration scheme is applied in all cases. The simulations are solved a result of the different time steppers are small for the square com-
with a nonlinear relative tolerance of 1  108 and a nonlinear abso- putational domains, but are more significant for the spherical and
lute tolerance of 1  1011 . T-shaped computational domains. As shown in Fig. 6, which pre-
To improve computational efficiency, adaptive meshing and sents microstructure snapshots for spinodal decomposition at
adaptive time stepping are used. Each simulation is performed t ¼ 200; t ¼ 2000, and the end of the simulation, obvious
twice, once with the aggressive ‘‘SolutionTimeAdaptive” time step- microstructural differences are discernible. Small variations early
per designed to finish the simulation as rapidly as possible [33], in the simulations strongly affect the microstructural evolution at
and once with the more conservative ‘‘IterationAdaptive” time later times, even in some cases affecting the final lowest-energy
stepper, which attempts to maintain a constant number of nonlin- structure, as seen for the T-shaped computational domain.
ear iterations and a fixed ratio of nonlinear to linear iterations. We An example microstructure for the Ostwald ripening problem is
choose a target of five nonlinear iterations, plus or minus one, and shown in Fig. 7, illustrating the solute and structural order param-
a linear/nonlinear iteration ratio of 100. Both time adaptivity algo- eter fields at t ¼ 20 for no-flux boundary conditions. The effect of
rithms allow a maximum of 5% increase per time step. In addition, the choice of time stepper is less evident in the free energy evolu-
gradient jump indicators [67] for c and l are used to determine tion (Fig. 5), but in some cases, coarsening kinetics are impacted.
mesh adaptivity, and the diffuse interface width spans at least five Fig. 8 shows parallel snapshots of the microstructure when the
elements in all simulations. conservative and aggressive time steppers are applied to a simula-
tion with periodic boundary conditions. In both simulations, a
smaller particle in the center of the computational domain is
4. Results and discussion shrinking; however, the particle has completely dissolved by
t ¼ 4111 when the conservative time stepper is used (Fig. 8a),
In this section, we present lessons learned from the first Hacka- while it has not quite disappeared by t ¼ 4131 when the aggressive
thon, the results of the two benchmark problems simulated with time stepper is used (Fig. 8b).
two different time adaptivity algorithms, and the needs that As illustrated in Fig. 5, the shrinkage of the central particle and
should be addressed with future benchmark problems. As dis- the concomitant coarsening of the surrounding particles is slower
cussed in the Introduction, many different software implementa- when the simulation is performed using the aggressive time step-
tions exist for phase field models, including bespoke software per, which is likely due to the fact that a particle shrinks faster as
developed in-house. While several phase field codes are designed its radius decreases. The conservative time stepper naturally cuts
to be applied to multiple types of problems, the possible multi- the time step size as the rate of particle shrinkage increases, while
physics couplings are so varied that it may be impossible to the aggressive time stepper typically tends to increase (or at least
develop a single phase field modeling framework to suit all phase maintain) the time step size until the solver is unable to converge
field modeling needs. Benchmark problems will help the phase to a solution. In multiple instances, particle dissolution is delayed,
field community in assessing the accuracy and performance of particularly in the later stages of the simulations. This is likely due
individual software implementations. to the fact that at later stages of the simulation, there is a greater
Several lessons were learned from the first Hackathon hosted by disparity in the radii of the shrinking and growing particles. In
CHiMaD, influencing the current benchmark problems as well as the case of late-stage particle shrinkage and dissolution, then, the
our design considerations for future problems. The Hackathon is aggressive time stepper may choose a time step size that is inap-
a twenty-four hour event in which teams of two participants each propriate for the physics of the system, increasing the error in
use their phase field software of choice to simulate a specified set the simulation. These results highlight the fact that adaptive time
of phase field problems with whatever software and computational steppers must be carefully assessed and chosen to minimize their
resources are available to them, including over the Internet. The impact on the simulated microstructural evolution.
goal of the Hackathon is to understand how different numerical For these benchmark problems, we are interested in the
implementations handle a set of phase field model problems of microstructural evolution all the way to the lowest energy state,
increasing difficulty with respect to accuracy and speed. We found although for other problems, evolving to equilibrium or local
that the original problem statements needed additional specifica- energy minimum may be unrealistic or even uninteresting. While
tions for participants to successfully run the simulations without the evolution of the total system energy provides important infor-
guesses or assumptions. Furthermore, the free energy functional, mation to assess simulation results, we find that it is difficult to
which was chosen from the literature, did not produce the phase determine a proper simulation exit condition. We originally tried
compositions that were indicated. Finally, we needed standardized a relative differential norm from one time step to the next with a
outputs for direct, quantitative comparison of the results. tolerance of 5  108 , but found that the simulations would some-
For these benchmark problems, we choose the total free energy times exit significantly before equilibrium was reached. Therefore,
of the system and microstructural snapshots as the metrics to com- we ran the simulations without exit parameters and relied on
pare simulation results. Because we use time adaptivity, we choose human intervention. We found that once a simulation has visibly
several synchronization times (t = 1, 5, 10, 20, 100, 200, 500, 1000, reached equilibrium (e.g., planar interfaces), the system free
2000, 3000, and 10,000) so that simulation results obtained from energy continues to decrease slowly, presumably due to equilibra-
146 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

Fig. 4. The total free energy evolution of the different variations of the spinodal decomposition benchmark problem simulated with two different time steppers, for (a) the
square computational domain with no-flux boundary conditions, (b) the square computational domain with periodic boundary conditions, (c) the T-shaped computational
domain, and (d) the spherical surface domain. ‘‘IA” indicates the conservative ‘‘IterationAdaptive” time stepper within MOOSE, and ‘‘STA” indicates the aggressive
‘‘SolutionTimeAdaptive” time stepper.

tion of very small solute gradients. Eventually the free energy stops energy (and thus V1 dF
dt
) that can occur when the system stops evolv-
evolving to six or seven significant figures, at which point we chose ing is evident. As an example, we find that a value of
to end the simulations. In some simulations, however, the value of 1 dF
¼ 1  1014 in dimensionless units generally appears to be suf-
V dt
the total free energy fluctuates in the sixth or seventh significant
ficient for these benchmark problems to indicate that the final con-
digit, indicating that the solver has reached its limits in terms of
figuration has been achieved while avoiding the descent into
numerical accuracy, and the simulations were again ended.
numerical noise. For other systems, the free energy and the rate
To determine a more useful criterion for ending these simula-
of change of the average free energy density may need to be nor-
tions, we calculate the rate of change of the volume-averaged free
malized to reasonable values for evaluation purposes.
energy density, V1 dF
dt
, an example of which is plotted in Fig. 9 for The proposed benchmark problems presented in this paper
both benchmark problems simulated with the conservative time model only a small subset of physics that have been incorporated
stepper. The rate of change allows for a more direct comparison into phase field formulations. Future benchmark problems should
of the evolution of the different within the different computational test additional key aspects, such as anisotropic linear elasticity
domains. While the total free energies of the different simulations with inhomogeneous moduli, anisotropic diffusivities and interfa-
vary by several orders of magnitude because of their differing com- cial energies, solidification, and CALPHAD-based thermodynamics.
putational domain size, V1 dFdt
is similar, as shown in Fig. 9. The rate of In addition, benchmark problems may benefit from being formu-
change varies by about ten orders of magnitude throughout the lated with a parameter that controls the numerical difficulty of
course of the simulation, highlighting the need for accurate adap- the problem, where possible. The solvers and the interpretation
tive time stepping algorithms when studying long-term of the problem statement may be verified for the ‘‘easy” problem,
microstructural evolution. As shown for the T-shaped spinodal while the software may be stress-tested when the parameter value
decomposition simulation in Fig. 9c, the numerical noise in the free makes the problem ‘‘difficult.” Furthermore, perturbation studies,
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 147

Fig. 5. The total free energy evolution of the different variations of the Ostwald ripening benchmark problem simulated with two different time steppers, for (a) the square
computational domain with no-flux boundary conditions, (b) the square computational domain with periodic boundary conditions, (c) the T-shaped computational domain,
and (d) the spherical surface domain. ‘‘IA” indicates the conservative ‘‘IterationAdaptive” time stepper within MOOSE, and ‘‘STA” indicates the aggressive ‘‘SolutionTimeA-
daptive” time stepper.

in which the initial conditions or problem parameters are slightly within different software; however, the governing physics are cap-
varied, may also be useful in determining how much simulation tured. Furthermore, the initial conditions are formulated such that
results differ as a result of numerical solvers versus any inherent they are implementation-independent, yet still disordered, similar
instability in the problem. Finally, while the quantities of interest to initial conditions often used in the literature. Multiple computa-
in each benchmark problem may vary (e.g., volume fraction of tional domains and boundary conditions are given so that the
solidified material, polarization field in a ferroelectric material), numerical implementations may be challenged and to address
we propose that the total free energy evolution is a standard, quan- future needs of phase field applications. We also discuss the need
titative output that may be used for every problem. Additional rel- to produce tractable output, i.e., data formats that allow simulation
evant comparison metrics should be identified and utilized on a results to be directly compared from different implementations,
per-problem basis. Community discussion and feedback is essen- and identify the total free energy evolution as a metric that should
tial for the development of relevant, useful problem sets, and we be used for every problem. We demonstrate the utility of the
urge individual researchers to contribute. benchmark problems by studying the effect of different time step-
pers on the microstructural evolution: small variations between
the simulations at earlier times become amplified at later times.
5. Conclusion
Given the deviation in our own results, we note that variations in
results between different implementations does not necessarily
In this paper, we propose two benchmark problems for numer-
imply an incorrect implementation. We also describe the use of
ical implementations of phase field models that capture essential
the normalized rate of change of the total free energy to halt the
physical behavior present in a vast majority of models: solute dif-
simulations appropriately. Finally, the problems presented in this
fusion and second-phase growth and coarsening. The model for-
paper test only a small subset of the physics often incorporated
mulations are simplified to make the tests easier to implement
148 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

Fig. 6. Snapshots of the microstructure evolution for spinodal decomposition simulated with different time steppers. (a-c), (g-i): conservative time stepper; (d-f), (j-l):
aggressive time stepper. (a), (d), (g), (j): t ¼ 200; (b), (e), (h), (k): t ¼ 2000; (c), (f), (i), (l): end time. Note the clearly visible differences in microstructure between the two time
steppers in (b) and (e), and (h) and (k).

into phase field models by design. Further benchmark problems problems should allow the validation of models with standard
are needed to model additional physics, such as linear elasticity, experimental data sets by ensuring that the differences in simula-
anisotropic diffusion and interfacial energies, solidification, tion results are not merely due to variations in numerical
and other phenomena. Ultimately, numerical benchmark implementations.
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 149

Fig. 7. The solute (a) and structural order parameter (b–e) fields at t ¼ 20 for the Ostwald ripening problem simulated with no-flux boundary conditions (no appreciable
difference in results is observed between the conservative and aggressive time steppers). Second-phase particles of differing gi in contact with each other do not coalesce.

Fig. 8. A comparison of the composition fields of Ostwald ripening simulations performed with periodic boundary conditions illustrating the effect of the choice of time
stepper on coarsening behavior. A shrinking particle has (a) completely dissolved by t ¼ 4111 when the simulation is performed with the conservative time stepper, while the
particle has not yet completely dissolved by (b) t ¼ 4131 when the simulation is performed with the aggressive time stepper.
150 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151

Fig. 9. The calculated V1 dF

dt
for the benchmark problems simulated with the conservative time stepper, which allows for a direct comparison of the evolution within the
different computational domains: (a) spinodal decomposition, (b) Ostwald ripening. (c) The calculated V1 dFdt
for the T-shaped spinodal decomposition simulation as the
simulation approaches equilibrium. Note the numerical noise in the free energy of the system as it stops evolving. ‘‘NF” and ‘‘PBC” indicate square computational domains
with no-flux boundary and periodic boundary conditions, respectively, and ‘‘T” and ‘‘Sphere” indicate the T-shaped and spherical surface computational domains.

For standard benchmark problems to become successful, the [11] J. Langer, Models of pattern formation in first-order phase transitions, in: G.
Grinstein, G. Mazenko (Eds.), Directions in Condensed Matter Physics, Series
community must provide feedback about and input into the cur-
on Directions in Condensed Matter Physics, World Scientific, 1986, pp. 165–
rently proposed problems and future problems. These standard 186.
benchmark problems are hosted on the NIST website, [12] D.U. Furrer, Application of phase-field modeling to industrial materials and
https://pages.nist.gov/chimad-phase-field/, along with the simula- manufacturing processes, Curr. Opin. Solid State Mater. Sci. 15 (2011) 134–
140.
tion data for the models presented here for download and compar- [13] A.A. Luo, Material design and development: from classical thermodynamics to
ison. The website will also serve as a repository for community- CALPHAD and ICME approaches, Calphad 50 (2015) 6–22.
submitted results. We encourage the community to contact the [14] G. Schmitz, B. Böttger, M. Apel, Microstructure modeling in ICME settings, in:
Proceedings of the 3rd World Congress on Integrated Computational Materials
authors directly or via the website for additional discussion. Engineering (ICME), John Wiley & Sons, 2015, p. 165.
[15] R. Qin, E. Wallach, A phase-field model coupled with a thermodynamic
Acknowledgments database, Acta Mater. 51 (2003) 6199–6210.
[16] H. Kobayashi, M. Ode, S.G. Kim, W.T. Kim, T. Suzuki, Phase-field model for
solidification of ternary alloys coupled with thermodynamic database, Scr.
The work by A.M.J., P.W.V., and O.G.H. was performed under Mater. 48 (2003) 689–694.
financial assistance award 70NANB14H012 from U.S. Department [17] B. Böttger, A. Carré, J. Eiken, G. Schmitz, M. Apel, Simulation of microstructure
formation in technical aluminum alloys using the multiphase-field method,
of Commerce, National Institute of Standards and Technology as Trans. Indian Inst. Met. 62 (2009) 299–304.
part of the Center for Hierarchical Material Design (CHiMaD). We [18] J. Zhu, Z. Liu, V. Vaithyanathan, L. Chen, Linking phase-field model to
gratefully acknowledge the computing resources provided on CALPHAD: application to precipitate shape evolution in Ni-base alloys, Scr.
Mater. 46 (2002) 401–406.
Blues and Fission, high-performance computing clusters operated
[19] J. Zhu, T. Wang, A. Ardell, S. Zhou, Z. Liu, L. Chen, Three-dimensional phase-
by the Laboratory Computing Resource Center at Argonne National field simulations of coarsening kinetics of c0 particles in binary Ni–Al alloys,
Laboratory and the High Performance Computing Center at Acta Mater. 52 (2004) 2837–2845.
[20] T. Kitashima, H. Harada, A new phase-field method for simulating c
Idaho National Laboratory, respectively. Finally, A.M.J. thanks
precipitation in multicomponent nickel-base superalloys, Acta Mater. 57
J.R. Jokisaari for constructive writing feedback. (2009) 2020–2028.
[21] S. Gentry, K. Thornton, Simulating recrystallization in titanium using the phase
References field method, in: IOP Conference Series: Materials Science and Engineering,
vol. 89, IOP Publishing, 2015, p. 012024.
[22] M. Wang, B. Zong, G. Wang, Grain growth in AZ31 Mg alloy during
[1] N. Moelans, B. Blanpain, P. Wollants, An introduction to phase-field modeling recrystallization at different temperatures by phase field simulation,
of microstructure evolution, Calphad 32 (2008) 268–294. Comput. Mater. Sci. 45 (2009) 217–222.
[2] H. Emmerich, Advances of and by phase-field modelling in condensed-matter [23] L. Aagesen, L. Lee, P.-C. Ku, K. Thornton, Phase-field simulations of GaN/InGaN
physics, Adv. Phys. 57 (2008) 1–87. quantum dot growth by selective area epitaxy, J. Cryst. Growth 361 (2012) 57–
[3] R. Duddu, D.L. Chopp, P. Voorhees, B. Moran, Diffusional evolution of 65.
precipitates in elastic media using the extended finite element and the level [24] J. Mangeri, O. Heinonen, D. Karpeyev, S. Nakhmanson, Influence of elastic and
set methods, J. Comput. Phys. 230 (2011) 1249–1264. surface strains on the optical properties of semiconducting core-shell
[4] C. Shen, Y. Wang, Phase-field microstructure modeling, in: L.S. Semiatin, D.U. nanoparticles, Phys. Rev. Appl. 4 (2015) 014001.
Furrer (Eds.), Fundamentals of Modeling for Metals Processing, vol. 22A, ASM [25] Comsol, Introduction to COMSOL Multiphysics: Version 5.1, Comsol, 2015.
International, 2009, pp. 297–311. [26] Simulia Advantage Support, 2016. <http://www.3ds.com/products-services/
[5] W. Boettinger, J. Warren, C. Beckermann, A. Karma, Phase-field simulation of simulia/support/documentation/> (accessed 11 May 2016).
solidification 1, Annu. Rev. Mater. Res. 32 (2002) 163–194. [27] G. Kresse, J. Hafner, Ab initio molecular dynamics for liquid metals, Phys. Rev.
[6] L.-Q. Chen, Phase-field models for microstructure evolution, Annu. Rev. Mater. B 47 (1993) 558.
Res. 32 (2002) 113–140. [28] G. Kresse, J. Hafner, Ab initio molecular-dynamics simulation of the liquid-
[7] I. Steinbach, Phase-field models in materials science, Modell. Simul. Mater. Sci. metal–amorphous-semiconductor transition in germanium, Phys. Rev. B 49
Eng. 17 (2009) 073001. (1994) 14251.
[8] B. Nestler, A. Choudhury, Phase-field modeling of multi-component systems, [29] G. Kresse, J. Furthmüller, Efficiency of ab-initio total energy calculations for
Curr. Opin. Solid State Mater. Sci. 15 (2011) 93–105. metals and semiconductors using a plane-wave basis set, Comput. Mater. Sci. 6
[9] I. Steinbach, Phase-field model for microstructure evolution at the mesoscopic (1996) 15–50.
scale, Annu. Rev. Mater. Res. 43 (2013) 89–107. [30] G. Kresse, J. Furthmüller, Efficient iterative schemes for ab initio total-energy
[10] G.J. Fix, Free Boundary Problems: Theory and Applications, vol. 2, Pitman, calculations using a plane-wave basis set, Phys. Rev. B 54 (1996) 11169.
Boston, USA, 1983, p. 580.
 A.M. Jokisaari et al. / Computational Materials Science 126 (2017) 139–151 151

[31] P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, [48] W. Scholz, J. Fidler, T. Schrefl, D. Suess, H. Forster, V. Tsiantos, et al., Scalable
G.L. Chiarotti, M. Cococcioni, I. Dabo, A.D. Corso, S. de Gironcoli, S. Fabris, G. parallel micromagnetic solvers for magnetic nanostructures, Comput. Mater.
Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Sci. 28 (2003) 366–383.
Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. [49] Chimad Phase Field Website, 2016. <https://pages.nist.gov/chimad-phase-
Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A.P. Seitsonen, A. Smogunov, P. field/> (accessed 15 April 2016).
Umari, R.M. Wentzcovitch, Quantum espresso: a modular and open-source [50] J.W. Cahn, On spinodal decomposition, Acta Metall. 9 (1961) 795–801.
software project for quantum simulations of materials, J. Phys.: Condens. [51] R.W. Balluffi, S. Allen, W.C. Carter, Kinetics of Materials, John Wiley & Sons,
Matter 21 (2009) 395502. 2005.
[32] L.A. Curtiss, K. Raghavachari, P.C. Redfern, J.A. Pople, Assessment of Gaussian-3 [52] J.W. Cahn, J.E. Hilliard, Free energy of a nonuniform system. I. Interfacial free
and density functional theories for a larger experimental test set, J. Chem. energy, J. Chem. Phys. 28 (1958) 258–267.
Phys. 112 (2000). [53] T.M. Pollock, S. Tin, Nickel-based superalloys for advanced turbine engines:
[33] M.R. Tonks, D. Gaston, P.C. Millett, D. Andrs, P. Talbot, An object-oriented finite chemistry, microstructure and properties, J. Propul. Power 22 (2006) 361–374.
element framework for multiphysics phase field simulations, Comput. Mater. [54] S.G. Kim, W.T. Kim, T. Suzuki, Phase-field model for binary alloys, Phys. Rev. E
Sci. 51 (2012) 20–29. 60 (1999) 7186.
[34] P.C. Millett, M.R. Tonks, K. Chockalingam, Y. Zhang, S. Biner, Three dimensional [55] A.A. Wheeler, W. Boettinger, G. McFadden, Phase-field model for isothermal
calculations of the effective Kapitza resistance of UO2 grain boundaries phase transitions in binary alloys, Phys. Rev. A 45 (1992) 7424.
containing intergranular bubbles, J. Nucl. Mater. 439 (2013) 117–122. [56] C.M. Elliott, D.A. French, F.A. Milner, A second order splitting method for the
[35] M. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Cahn-Hilliard equation, Numer. Math. 54 (1989) 575–590.
Ring, M. Rognes, G. Wells, The FEniCS project version 1.5, Arch. Numer. Softw. [57] S.M. Allen, J.W. Cahn, A microscopic theory for antiphase boundary motion and
3 (2015). its application to antiphase domain coarsening, Acta Metall. 27 (1979) 1085–
[36] M.J. Welland, D. Karpeyev, D.T. O’Connor, O. Heinonen, Miscibility gap closure, 1095.
interface morphology, and phase microstructure of 3D LixFePO4 nanoparticles [58] N. Saunders, A.P. Miodownik, CALPHAD (Calculation of Phase Diagrams): A
from surface wetting and coherency strain, ACS Nano 9 (2015) 9757–9771. Comprehensive Guide, vol. 1, Elsevier, 1998.
[37] P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, M. Ohlberger, O. Sander, [59] H.L. Lukas, S.G. Fries, B. Sundman, Computational Thermodynamics: The
A generic grid interface for parallel and adaptive scientific computing. Part I: Calphad Method, vol. 131, Cambridge University Press, 2007.
Abstract framework, Computing 82 (2008) 103–119. [60] C.M. Funkhouser, F.J. Solis, K. Thornton, Dynamics of coarsening in
[38] P. Bastian, M. Blatt, A. Dedner, C. Engwer, R. Klöfkorn, R. Kornhuber, M. multicomponent lipid vesicles with non-uniform mechanical properties, J.
Ohlberger, O. Sander, A generic grid interface for parallel and adaptive Chem. Phys. 140 (2014) 144908.
scientific computing. Part II: Implementation and tests in DUNE, Computing 82 [61] D. Gaston, J. Peterson, C. Permann, D. Andrs, A. Slaughter, J. Miller, Continuous
(2008) 121–138. integration for concurrent computational framework and application
[39] J.E. Guyer, D. Wheeler, J.A. Warren, FiPy: partial differential equations with development, J. Open Res. Softw. 2 (2014).
Python, Comput. Sci. Eng. 11 (2009) 6–15. [62] D.R. Gaston, C.J. Permann, J.W. Peterson, A.E. Slaughter, D. Andrš, Y. Wang, M.P.
[40] D. Wheeler, J.A. Warren, W.J. Boettinger, Modeling the early stages of reactive Short, D.M. Perez, M.R. Tonks, J. Ortensi, Physics-based multiscale coupling for
wetting, Phys. Rev. E 82 (2010) 051601. full core nuclear reactor simulation, Ann. Nucl. Energy 84 (2015) 45–54.
[41] I. Steinbach, F. Pezzolla, B. Nestler, M. Seeßelberg, R. Prieler, G. Schmitz, J. [63] A. Jokisaari, K. Thornton, General method for incorporating CALPHAD free
Rezende, A phase field concept for multiphase systems, Physica D: Nonlinear energies of mixing into phase field models: application to the a-zirconium/d-
Phenom. 94 (1996) 135–147. hydride system, Calphad 51 (2015) 334–343.
[42] M. Mecozzi, J. Eiken, M. Santofimia, J. Sietsma, Phase field modelling of [64] A. Jokisaari, C. Permann, K. Thornton, A nucleation algorithm for the coupled
microstructural evolution during the quenching and partitioning treatment in conserved–nonconserved phase field model, Comput. Mater. Sci. 112 (2016)
low-alloy steels, Comput. Mater. Sci. 112 (2016) 245–256. 128–138.
[43] B. Nestler, H. Garcke, B. Stinner, Multicomponent alloy solidification: phase- [65] T. Blacker, S. Owen, M. Staten, R. Quadros, B. Hanks, B. Clark, R. Meyers, C.
field modeling and simulations, Phys. Rev. E 71 (2005) 041609. Ernst, K. Merkley, R. Morris, C. McBride, C. Stimpson, M. Plooster, S. Showman,
[44] B. Stinner, B. Nestler, H. Garcke, A diffuse interface model for alloys with CUBIT: Geometry and Mesh Generation Toolkit 15.1 User Documentation,
multiple components and phases, SIAM J. Appl. Math. 64 (2004) 775–799. Sandia National Laboratory, Albuquerque, New Mexico, SAND2016-1649 R,
[45] lMAG - Micromagnetic Modeling Activity Group, 2016. <http://www.ctcms. 2016.
nist.gov/rdm/mumag.org.html> (accessed 1 April 2016). [66] A. Iserles, A First Course in the Numerical Analysis of Differential Equations,
[46] M.J. Donahue, D.G. Porter, OOMMF User’s Guide, Version 1.0, volume NISTIR Cambridge University Press, 2009.
6376, US Department of Commerce, Technology Administration, National [67] B.S. Kirk, J.W. Peterson, R.H. Stogner, G.F. Carey, libMesh: a C++ library for
Institute of Standards and Technology, Gaithersburg, Maryland, 1999. parallel adaptive mesh refinement/coarsening simulations, Eng. Comput. 22
[47] A. Vansteenkiste, J. Leliaert, M. Dvornik, M. Helsen, F. Garcia-Sanchez, B. Van (2006) 237–254.
Waeyenberge, The design and verification of MuMax3, AIP Adv. 4 (2014)
107133.