Computational Materials Science 193 (2021) 110371

Contents lists available at ScienceDirect

Computational Materials Science

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

Phase field benchmark problems for nucleation

W. Wu a, b, *, D. Montiel c, J.E. Guyer d, P.W. Voorhees a, f, J.A. Warren d, D. Wheeler d,
L. Gránásy g, T. Pusztai g, O.G. Heinonen b, e
a
Center for Hierarchical Materials Design, Northwestern University, 2205 Tech Drive, Suite 1160, Evanston, IL 60208, USA
b
Materials Science Division, Argonne National Laboratory, 9700 South Cass Avenue, Lemont, IL 60439, USA
c
Department of Materials Science and Engineering, University of Michigan, 2300 Hayward Street, Ann Arbor, MI 48109, USA
d
Material Measurement Laboratory, National Institute of Standards and Technology, 100 Bureau Drive, MS 8300, Gaithersburg, MD 20899-8300, USA
e
Northwestern-Argonne Institute of Science and Engineering, 2205 Tech Drive, Suite 1160, Evanston, Illinois 60208, USA
f
Department of Materials Science and Engineering, Northwestern University, 2220 Campus Drive, Evanston, IL 60208, USA
g
Department of Experimental Solid State Physics, Institute for Solid State Physics and Optics, Wigner Research Centre for Physics, 29-33, Konkoly-Thege Miklós út,
Budapest H-1121, Hungary

A R T I C L E I N F O A B S T R A C T

Keywords: We present nucleation phase field model benchmark problems, expanding on our previous benchmark problems
Phase field on diffusion, precipitation, dendritic growth, linear elasticity, fluid flow and electrochemistry. Nucleation is the
Benchmark process in which either a new thermodynamic phase or a new structure is created, such as solidification from the
Nucleation
melt, or self-assembly of particulates. Based on where the nucleation occurs, it can be divided into two main
categories: homogeneous nucleation and heterogeneous nucleation. In the first nucleation benchmark problem,
we focus on homogeneous nucleation for both single seed under different initial conditions and multiple seeds.
The second nucleation benchmark problem focuses on athermal heterogeneous nucleation and nucleation
behavior near the free growth limit with different undercooling driving forces.

1. Introduction enormous practical and theoretical importance for various branches of

science, including physical chemistry, materials science, biophysics,
We continue our series of Benchmark Problems for Phase Field geophysics, and cryobiology, to name a few. Previous work introduced a
models. They are developed by the Center for Hierarchical Materials benchmark problem for dendritic growth describing complex solidifi­
Design (CHiMaD) and the National Institute of Standards and Technol­ cation patterns in an undercooled liquid or solution [2]. This work in­
ogy (NIST) together with considerable support from the phase field troduces a benchmark for nucleation which models the initial stages of
community of developers and modelers. The purpose of the benchmark the solidification process and, thus, complements the solidification
problems is to provide resources to test new algorithms and codes for benchmark, which models the later stages of solidification. It is of in­
numerical accuracy, and to train new researchers. In the publication terest to extend the phase field Benchmark Problems to nucleation, as it
series of the phase field benchmark problems we present the rationale is a process that sets the initial microstructure for solidification, growth,
for selecting and defining the problems, suggest metrics for testing and and coarsening. Phase field modeling of nucleation has a long history
comparing solutions, and also present sample solutions. The problems and is also covered in a number of reviews [4–10]. The problem
are selected from the canon of phase field modeling, and often include formulation of crystallization of an ideal pure liquid cooled below its
multi-physics couplings that may be encountered in typical problems melting point starts with homogeneous nucleation, a process in which
and that may frustrate numerical solutions. Previous problems focused the internal fluctuations of the undercooled liquid lead to the formation
on diffusion of solute and second phase coarsening [1], solidification of of crystal-like seeds able to grow to macroscopic sizes. The nucleation
an undercooled liquid (dendritic growth) and linear elasticity [2], and can be assisted by the presence of surfaces (container walls, foreign
Stokes flow and electrostatics [3]. Solidification, in particular, is of particles, etc.), in which case the process is termed heterogeneous

* Corresponding author.
E-mail addresses: wuw@anl.gov (W. Wu), dmontiel@umich.edu (D. Montiel), guyer@nist.gov (J.E. Guyer), p-voorhees@northwestern.edu (P.W. Voorhees),
james.warren@nist.gov (J.A. Warren), daniel.wheeler@nist.gov (D. Wheeler), granasy.laszlo@wigner.mta.hu (L. Gránásy), pusztai.tamas@wigner.mta.hu
(T. Pusztai), heinonen@anl.gov (O.G. Heinonen).

https://doi.org/10.1016/j.commatsci.2021.110371
Received 3 December 2020; Received in revised form 3 February 2021; Accepted 4 February 2021
Available online 10 March 2021
0927-0256/Published by Elsevier B.V.
W. Wu et al. Computational Materials Science 193 (2021) 110371

nucleation. We note that homogeneous nucleation is an idealized an isothermal pure substance with one liquid (ϕ = 0) and one solid (ϕ =
formulation, and it is unlikely homogeneous nucleation occurs, due to 1) phase. The free energy of this system is
impurities present in experimental apparatus. However, creating ( ) ∫[ 2 ( ) ( )]
∊
benchmark problems for homogeneous nucleation is still needed since F ϕ = (∇ϕ)2 + wg ϕ − Δfp ϕ dV, (1)
2
the focus in nucleation research often lies on the underlying homoge­
neous nucleation, as it is the basis for advancing theoretical approaches
to the much more complex phenomena governing heterogeneous set g(ϕ) = ϕ2 (1 − ϕ)2 , a simple double well function with minima at ϕ =
nucleation. 0 and ϕ = 1, w controls the double-well barrier height, ∊2 is the gradient
We present two Benchmark Problems on nucleation, with the first energy coefficient, let p(ϕ) = ϕ3 (10 − 15ϕ + 6ϕ2 ), which ensures that
one targeting homogeneous nucleation. There are two main modeling p(0) = p′ (0) = p′ (1) = 0 and p(1) = 1, and Δf is the driving force for
approaches to introduce nuclei into a metastable system: the Langevin solidification at the simulation temperature (Δf is positive below the
noise method [4,11] and the explicit nucleation method [12,13]. We melting point). The time evolution of ϕ is given by the Allen–Cahn
focus on the explicit method, where the critical nucleus size and the equation [25]
nucleation energy are determined by the classical nucleation theory. In ∂ϕ δF [ ( ) ( )]
contrast with the Langevin noise method, the explicit nucleation method = − M = M ∊2 ∇2 ϕ − wg′ ϕ + Δfp′ ϕ (2)
∂t δϕ
allows for consideration of nucleation events that may occur whenever
there is a sufficient driving force. We break down our consideration of where M is the mobility parameter. We will restrict the problem to two
homogeneous nucleation into three parts. The first part considers the dimensions (2D). For a planar interface, one can show that the equi­
simple case of single-seed homogeneous nucleation. We explore how the librium (Δf=0) solid–liquid interface profile with the interface centered
particle size influences the evolution of the nucleus when the thermo­ at x = x0 is given by
dynamic driving force is close to the critical value where we can observe ( ) [ ( )]
whether the particle grows, dissolves or remains stationary. In the sec­ 1
ϕ x = 1 − tanh √̅̅̅
x − x0
, (3)
ond and third parts we consider multiple-seeds homogeneous nucle­ 2 2ℓ
ation, with nuclei appearing at fixed time t = 0 or at random times
where x is the perpendicular distance from the interface. We use this
distributed uniformly, respectively, with the latter part illustrating a
expression as the initial condition when we introduce a nucleus. We can
scenario with constant nucleation rate. The Benchmark Problem probes
also obtain the width ℓ and the free energy of the interface, γ, as
the differences between transformation kinetics of the two cases and
√̅̅̅̅̅
summarizes them using the Avrami plots based on the John­ ∊2
son–Mehl–Avrami-Kolmogorov (JMAK) theory [14–18]. The second ℓ= (4)
w
Benchmark Problem focuses on athermal heterogeneous nucleation as
proposed by Quested and Greer [19]. Athermal nucleation can explain and
the performance of inoculants in grain-refining of commercial √̅̅̅̅̅̅̅̅
aluminum alloys [20,21]. The particles remain dormant at and below a γ=
∊2 w
√̅̅̅ . (5)
critical undercooling, at which the radius of the particles is equal to that 3 2
of the critical radius for the homogeneous nucleus. Further free growth
Choosing a characteristic length scale ξ = ℓ and a characteristic time
is possible if the degree of undercooling is increased. The Benchmark
scale τ = 1/(Mw), we can obtain a nondimensional form of the problem
Problem explores the nucleation behavior around the free growth limit
(the nondimensional quantities are denoted by tildes),
with different undercooling driving forces. We use a simple example for
( ) ∫[ ( ) ( )]
the heterogeneous nucleation. In more complex realistic scenarios, the 1 ̃ 2
phase field method could remove some of the critical assumptions
̃ ϕ =
F (∇ϕ) + g ϕ − Δf ̃ p ϕ dV ̃ (6)
2
involved in classical approaches to heterogeneous nucleation existing in
the literature and predict significantly different critical nuclei shapes and
and activation energy barriers [22,23]. ( ) ( )
∂ϕ ̃ 2
The mathematical and algorithmic implementation of these Bench­ ̃ p′ ϕ ,
= ∇ ϕ − g′ ϕ + Δf (7)
mark Problems is relatively simple compared to other Benchmark
∂̃t
Problems, e.g., the coupled Cahn–Hilliard-linear elastic problem, or the
coupled Cahn–Hilliard-Poisson Benchmark Problem [2,3]. However, with Δf
̃ = Δf/w.

this does not make solving the problems numerically trivial, and we
deliberately chose undercoolings in some problem formulations that
2.2. The properties of the classical nucleus
stress convergence of the solver. Nevertheless, the solutions should be
relatively straightforward, and we believe that the pedagogical aspect of
The explicit nucleation method that we employ introduces nuclei
introducing nucleation Benchmark Problems is an important one. As
into a metastable system, with their radii and nucleation barrier deter­
with the previous Benchmark Problems, we present suggested metrics
mined from classical nucleation theory. Classical nucleation theory
for how to evaluate solutions for the nucleation benchmark problem. In
views crystallite fluctuations appearing in the undercooled liquid as
addition to our own solutions presented here, The PFHub website [24]
small spherical domains of the bulk crystalline phase bounded by a
(https://pages.nist.gov/pfhub/) hosts our solutions as well as solutions
mathematically sharp solid–liquid interface. For a 2D system, the free
by others using a variety of codes. We encourage readers to upload their
energy of a circular solid particle of radius r is
solutions and to explore the additional resources there, and to partici­
( )
pate in discussions around the Benchmark Problems. ΔG r = 2πrγ − πr2 Δf (8)

2. Model formulations where Δf is the nucleation driving force that is used in our phase field
model, and γ is the free energy of the interface. The free energy of the
2.1. The phase field model in classical nucleation theory particle is a balance between the energy cost in forming the solid–liquid
interface, and the free energy from the driving force which is released
For these benchmark problems we use the simplest possible phase when the crystalline particle forms. Once the rate of change of free en­
field model with a single non-conserved phase field ϕ, which describes ergy with respect to particle size becomes negative, the particle can

2
W. Wu et al. Computational Materials Science 193 (2021) 110371

grow. Taking the derivative of ΔG(r) we get the rate of change of free 2D where r is the vector that starts from the central origin to point (x,y).
energy with respect to radius as It has been shown that the phase field order parameter profile in a
critical nucleus can deviate from that assumed in the classical nucleation
dΔG
= 2πγ − 2πrΔf . (9) theory, especially when the critical radius is close to the interface width
dr
[4,7,10,26]. In this problem, we carefully choose the driving force so
By setting dΔG/dr to zero, we obtain the critical radius r* as that the critical radius r* = 5 is much larger than the interface width ℓ =
1. Therefore, the phase field critical nuclei should behave as described
γ
r* = . (10) by the classical sharp-interface model. Problem I.1 is then defined by the
Δf
Allen–Cahn equation
The corresponding critical nucleation free energy is ( ) ( )
∂ϕ
= ∇2 ϕ − g′ ϕ + Δfp′ ϕ (15)
πγ2 ∂t
ΔG* = . (11)
Δf
for the phase field ϕ, and the three different initial conditions: in the first
Using the same units as before, the nondimensional forms of these case with the seed radius r0 corresponding exactly to the critical one
quantities are (critical nucleus), and in the other two cases, slightly below and above
the critical radius (subcritical and supercritical nuclei):
1 1
r̃* = √̅̅̅ (12) ( )
1
[ (
||r|| − r0
)]
3 2 Δf̃ ϕ r = 1 − tanh √̅̅̅ , (16)
2 2
and
where r0 = {r* ,0.99r* ,1.01r* }. We obtain Eq. 16 from Eq. 3 by imposing
̃* =
ΔG
π 1
, (13) circular symmetry on the nuclei and use the isotropy of the free energy
̃
18 Δf density. When r = (x, 0), Eq. 16 reduces to the 1D case Eq. 3. Fig. 1
shows the computational domain with an initial seed of radius r0 = r* .
respectively. This seed is the diffuse-interface approximation of the classical sharp
interface nucleus corresponding to the given Δf, and therefore it should
2.3. Avrami plots be fairly close to (an unstable) equilibrium. The time evolution of the
system is then followed for times up to t = 100 units, and the solid
Avrami plots describe how solids transform from one phase to fraction Y and the total free energy F are plotted as functions of time.
another at constant temperature. In particular, they can describe the Finally, the problem also includes a convergence test with respect to
kinetics of nucleation. The Avrami plots come from the JMAK theory mesh (spatial resolution). The closer the initial radius r0 is to the critical
[14–18], which makes a number of assumptions and simplifications: (i) radius r* , the more sensitive the numerical integration will be to round-
nucleation occurs randomly and homogeneously over the entire un- off errors. Therefore, a convergence check should be performed for r0 =
transformed portion of the material, (ii) the growth rate is constant 1.01r* for times up to 200 units. This convergence check can simply be
and does not depend on the extent of transformation, (iii) the particles done by halving the average spatial mesh size Δx between consecutive
have convex shape with the same orientation, and (iv) the size of the runs until plots of the solid fraction Y(t) vs. time t for two consecutive
system is infinite and so is the time period in which the transformation runs become indistinguishable by some measure. For uploads to the
occurs.
The basis of the Avrami plots is the transformed fraction (solid
fraction) Y(t) vs. time. According to the JMAK theory,
Y(t) = 1 − exp( − Ktn ), (14)

where K is a constant depending on the nucleation and growth rates, n =

d +1 for continuous nucleation and n = d if nucleation happens only at
t = 0, and d is the number of spatial dimensions. If we plot log[ −
log(1 − Y(t) ) ] vs. log(t), then for the JMAK kinetics (Eq. 14) we get a
straight line with slope n.

3. Nucleation benchmark problems

We present two nucleation benchmark problems in this set of

benchmarks, the first for homogeneous nucleation and the second for
athermal heterogeneous nucleation. From here on, we will use only the
nondimensional forms of the phase field (Eqs. 6, 7, and 3) and nucle­
ation (Eqs. 12 and 13) equations, but we will drop the tildes.

3.1. Problem I.1 – Homogeneous nucleation, single seed

In this problem we examine the morphology change of the nucleus

for different initial radii. We consider a 2D simulation domain of size
100 × 100 units centered at x = y = 0 with zero normal-derivative
Neumann boundary conditions, ∇ϕ⋅ ̂ n = 0, on all boundaries. The
√̅̅̅ Fig. 1. Illustration of the 100 × 100 computational domain in 2D with zero
driving force is set to Δf = 2/30, which corresponds to a critical radius normal-derivative Neumann boundary conditions on all boundaries. A diffuse-
of r* = 5 (Eq. 12). Next, we incorporate a diffuse-interface seed at the interface circular seed with profile ϕ(r) and radius r0 = r* is placed in the center
center of the domain using the profile ϕ(r) given by modifying Eq. 3 to at t = 0.

3
W. Wu et al. Computational Materials Science 193 (2021) 110371

PFHub website, the phase field ϕ at the last time step can be uploaded 3.4. Problem II – Athermal heterogeneous nucleation
together with the real-space grid points (x, y) for the mesh used in the
simulation. A script on the website can then be used to interpolate so­ For a benchmark problem on athermal heterogeneous nucleation, a
lutions on different meshes and calculate a convergence error. surface with good wetting properties but of limited size is needed. In the
usual realistic scenario, these nucleation surfaces represent small, flat,
foreign particles which serve as nucleation sites. In order to simplify the
3.2. Problem I.2 – Homogeneous nucleation, multiple seeds at t = 0 problem and to focus it on the athermal nucleation process rather than
on technical issues with introducing foreign flat particles, we avoid
The second part of the homogeneous nucleation benchmark problem adding additional boundaries (particles) to the problem formulation,
focuses on the kinetics of nucleation using Avrami plots and compares and instead use one of the sides of a rectangular domain by making parts
them to the JMAK theory. To obtain reasonable statistics for these two of it a good wetting surface. We use a parameter ϕ0 in Dirichlet
parts, the simulation volume needs to be larger so it can encompass a boundary conditions on ϕ (ϕ = ϕ0 ) to determine the wetting properties
larger number of smaller nuclei. We therefore increase the domain size of the bounding surface – this corresponds to the “Model B” approach
to 500 × 500 units of length while keeping the same Neumann boundary described by Warren et al. [8]. This “Model B” approach has two ad­
conditions on all boundaries. We then increase the driving force to Δf = vantages: (i) it is very simple to implement, and (ii) if we set ϕ0 large
√̅̅̅
2/6, which corresponds to a critical radius of r* = 1. Random initial enough, a surface spinodal would occur, i.e., a solid phase will auto­
positions of 25 supercritical seeds r1 , …, r25 are generated with r0 = matically appear at the surface and grow, so we can eliminate the effort
2.2 = 2.2r* drawn from a uniform distribution on the 500 × 500 of inserting an appropriate seed in the beginning of the simulation. Zero
domain. The distribution of the phase field ϕ is the sum of the phase normal-derivative Neumann boundary conditions (∇ϕ⋅ ̂ n = 0) are
fields ϕi with profiles applied to the other three sides of the computational domain.
( )
1
[ (
||r − ri || − r0
)] The goal of this problem is to explore the behavior around the free
ϕi r = 1 − tanh √̅̅̅ . (17) growth limit, which, in our 2D setting, is the half-circular configuration
2 2
of solid on top of a flat surface of length 2r* , where r* is the radius of the
from the different seeds i. Overlaps between different seeds are managed homogeneous nucleus corresponding to the given driving force (Eq. 12).
√̅̅̅
by setting ϕ = 1 in all regions where the sums of ϕi > 1.1 After placing To define the problem, the undercooling is set to Δf0 = 2/60, which
initial seeds and adjusting the initial phase field ϕ so that ϕ ≤ 1 every­ corresponds to a critical radius r* = 10. We represent the strongly
where, the simulation is run up to total time t = 200, at which time the wetting surface of a particle of this size by setting ϕ0 = 0.9 in the |x| ≤
whole domain is transformed to a solid (ϕ = 1). This is repeated four 10 inner part, and ϕ0 = 0 in the |x| > 10 outer part of the bottom
times, each time with different random seeds, and the total free energy, boundary in a simulation domain of 40 units by 20 units (Fig. 2). Please
time evolution of the solid fraction Y, discrete particle count N (the note that the discontinuities in ϕ0 at |x| = 10 will result in singularities
number of disjoint regions with ϕ = 1) are plotted for the five simula­ in ϕ around these points with very high local free energy density. This
tions, and Avrami plots are generated from the five simulations; from the slightly changes the total free energy and the critical undercooling.
Avrami plot the exponent n is estimated and compared to the JMAK However, the incurred changes are small enough that they do not
theory. A snapshot of the phase field at t = 40 is taken for one of the five change the end result in a discernible way, as whether or not the system
runs. passes the free growth limit will depend on the relative energies of
Both Problem I.2 and Problem I.3 involve a random number gener­ configurations near the semi-circular critical configuration. We have, in
ator. Because different executions will in general generate different fact, checked that eliminating these singularities by smearing out the
random numbers, specific individual solutions will in general not be steps in ϕ0 at |x| = 10 in the boundary condition by using tanh profiles
reproduced. However, statistics such as the average slope of Avrami slightly increases the critical undercooling by about 1 %. For the purpose
plots can be meaningfully compared. of this problem, this difference is acceptable as the aim of the problem
formulation is to produce a critical undercooling close to unity which is
more easily realizable with a sharp boundary profile than with a diffuse
3.3. Problem I.3 – Homogeneous nucleation, multiple seeds at random
one. Starting from ϕ = 0 everywhere in the domain, the simulation is
times
run for times up to t = 6500, and the transformed fraction Y is plotted as
a function of time t. The simulation is repeated with the same boundary
For the third part of the homogeneous benchmark problem, the
conditions but varying undercooling Δf = 0.98Δf0 , Δf = 1.02Δf0 , and
computational domain is first expanded to 1000 × 1000. Instead of
Δf = 1.04Δf0 , respectively. We add an additional convergence test with
inserting nuclei at fixed time t = 0 as in part two of the problem, in this
respect to spatial resolution for the solid fraction change when Δf =
case 100 random nucleation times are generated, t1 ,…,t100 , drawn from
1.1Δf0 by halving the average linear dimension of the mesh elements
a uniform distribution in the interval ti ∈ [0, 600) with centers ri drawn
until the plots of solid fraction Y(t) vs. time of two consecutive mesh
from a uniform distribution on the 1000 × 1000 domain. We decrease
√̅̅̅ sizes become indistinguishable.
the driving force to Δf = 2/12, which corresponds to a critical radius
*
of r = 2, but keep the magnitude of initial radii of the inserted nuclei 4. Numerical methods
the same as those in the second part of the problem, i.e., Eqs. 17 with
r0 = 2.2 = 1.1r* . Again, ϕ is set to unity in regions of overlaps of nuclei. We use an application within the MOOSE framework [27,28] for our
The simulations are run up to times t = 600, and repeated for four more simulations, as we have done for previous benchmark problems. Solu­
random initial conditions. The total free energy, the time evolution of tions obtained with other codes, e.g., FiPy [29] or PRISMS [30], have
the solid fraction X, and the discrete particle count N are plotted as been uploaded to the PFHub website https://pages.nist.gov/pfhub/
functions of time, and an Avrami plot is generated for the five runs. [24].
Again, a snapshot of the phase field at t = 100 is taken for one of the five The computational domains created by the MOOSE application are
runs. meshed using square, four-node quadrilateral elements. First-order
Lagrange shape functions are employed for the order parameter ϕ.
The system of nonlinear equations is solved with the full Newton method
1
Strictly speaking, this may cause ϕ to become non-differentiable at inter­ that uses the single-matrix preconditioning with the additive Schwarz
section points of the overlapping seeds, however that does not affect numerical preconditioner and ILU sub-preconditioning. The second backward
convergence when the Laplacian of ϕ is evaluated according to some stencil.

4
W. Wu et al. Computational Materials Science 193 (2021) 110371

Fig. 2. Illustration of the 40 × 20 2D computational domain with Dirichlet boundary conditions on the bottom side and Neumann boundary conditions on the other
sides. The red line represents the middle part of the boundary of length 20 where ϕ0 = 0.9, ϕ0 = 0 elsewhere.

differentiation formula (BDF2) time integration scheme is applied. The not. We choose the total free energy of the system and the solid fraction
simulations are solved with a nonlinear relative tolerance of 10− 6 and a as metrics with which to compare simulation results.
nonlinear absolute tolerance of 10− 9 . A maximum of ten nonlinear it­ Fig. 3 shows the total free energy of the system, and Fig. 4 shows how
erations per solve, and a maximum of 50 linear iterations, are specified. the solid fraction changes over time. When the initial radius of the seed
To improve computational efficiency, adaptive meshing and adaptive is smaller than the critical radius, e.g., r0 = 0.99r* , based on classical
time stepping are used. The initial time step Δt is set to 0.01, and the nucleation theory, the quantity dΔG/dr in Eq. 9 is positive. That means
time step is allowed to grow by 5 % per step. For Problem I.3, if Δt grows that it is energetically more favorable for the system to decrease the size
too large then many nucleation events may occur in one time step. In of the nucleus. As a result, the total free energy (black dashed line in
order to reduce the probability of this and force nucleation events to Fig. 3) and the solid fraction (black dashed line in Fig. 4) both decrease
occur at different times, Δt is capped at unity and is also reduced to 0.1 if over time. The inserted seed shrinks and at t ≈ 90, it fully dissolves into
a nucleus has just been added to the system, from which Δt is allowed to the liquid. The solid fraction drops to zero and the total free energy of
grow again. In addition, this scheme can also help with the convergence the system stops decreasing. Nucleation does not happen in this case. On
of solvers. For the single-seed homogeneous nucleation part of the first the other hand, when the initial radius of the seed is larger than the
problem, the mesh contains 64 × 64 elements. This translates to an critical radius (r0 = 1.01r* ), dΔG/dr in Eq. 9 is negative, and it is
element size of Δx = 1.5625. A single element can be refined up to four energetically favorable for the system to increase the size of the nucleus,
times for adaptive meshing. Thus the finest part of the mesh has ele­ i.e., the system lowers its free energy by growing the solid fraction from
ments with size Δx = 0.09765. The convergence test for the solid the inserted seed. We note that with no fluxes entering or leaving the
fraction change in the r0 = 1.01r* case is done by four simulations with system, the total energy is necessarily monotonically non-increasing.
different fixed spatial resolutions (Δx = 0.8,0.4,0.2,0.1), and compared When the initial radius of the seed is equal to the critical radius (r0 =
with applying the adaptive mesh. For the second and third parts of the 1.00r* ), dΔG/dr in Eq. 9 is equal to zero, and the system is in an unstable
first problem, the mesh contains 160 × 160, and 320 × 320 elements, equilibrium state. Therefore, the total free energy (red solid line in
respectively, which result in the same spatial resolution, Δx = 3.125, Fig. 3) stays approximately constant. Ideally, the solid fraction would
and the finest part of the mesh has elements with size Δx = 0.1953. For also remain constant. The small changes in the calculated solid fraction
the second problem, on athermal nucleation, the mesh contains 100 × (red solid line in Fig. 4) when t is large are due to numerical errors that
50 elements, which gives an element size of Δx = 0.4. With the adaptive accumulate with time. Because the system is unstable, small errors will
meshing, the size of the smallest element is Δx = 0.025. Five different
spatial resolutions (Δx = 0.4, 0.2, 0.1, 0.05, 0.025) are used for the
convergence test when plotting the solid fraction versus time for the
Δf = 1.1Δf0 case. We also compare the results with the simulation using
adaptive meshing.

5. Results and discussion

In this section, we discuss the example solutions and metrics ob­

tained from them. We also study error and convergence behavior for the
r0 = 1.01r* case in the single seed homogeneous nucleation problem and
the Δf = 1.1Δf0 case in the athermal heterogeneous nucleation problem.
Furthermore, we discuss some peculiar issues that may arise in solving
these benchmark problems.

5.1. Problem I.1 – Homogeneous nucleation, single seed

For this problem we mainly look into how the initial radius of the Fig. 3. The total free energy evolution of the single seed homoge­
inserted seed determines whether the nucleation is going to happen or neous nucleation.

5
W. Wu et al. Computational Materials Science 193 (2021) 110371

5.2. Problem I.2 – Homogeneous nucleation, multiple seeds at t = 0

For this part of the homogeneous nucleation benchmark problem we

mainly look into the nucleation behavior when multiple seeds are
inserted at t = 0. Besides the total free energy of the system and the solid
fraction, we also add the number of discrete particles (calculated by
using FeatureFloodCount postprocessor in MOOSE) and the Avrami plot
as metrics to illustrate the transformation kinetics. Figs. 6–9 depict the
calculated total free energy of the system, solid fraction, number of
discrete particles, and the Avrami plot for the multiple-seed homoge­
neous nucleation that happens at t = 0, respectively. Different colors in
the figures represent runs with different random seeds. A snapshot of the
phase field domain at t = 40 is taken and shown in Fig. 10.
In all five cases with different random nuclei positions, the total free
energy is decreasing monotonically, as shown in Fig. 6, as should be the
Fig. 4. The solid fraction of the single seed homogeneous nucleation. case. Since the initial radius for the nuclei is larger than the critical
radius, we expect to see an increase for the solid fraction over time, and
then eventually lead to growth or shrinkage of the nucleus. the nuclei should grow and eventually fill up the entire simulation
The convergence test for the solid fraction change for the case r0 = volume. This is indeed the case, as shown in Fig. 7. The solid fraction
increases with time in an S-shaped curve with a short left tail. The short
1.01r* is shown in Fig. 5. We use an L2 metric as convergence check. The
tail means that the early stage of nucleation with a low transformation
L2 error, eL2 , between two consecutive runs i − 1 and i is defined as
rate is short, since all the nuclei are inserted and ready to grow at time
∫
t = 0. For most cases, the number of discrete particles starts at 25 since
eL2 = |ϕi (r, t = 200) − ϕi− 1 (r, t = 200)|2 dr. (18)
we insert 25 seeds at t = 0. However, the number of discrete particles
can be smaller, e.g., 24 or even less, if two or more inserted seeds
The ability to calculate the L2 error involves interpolating solutions to overlap. The number of discrete particles also decreases monotonically
different meshes. Because this can be a bit challenging technically, we and finally drops to unity at some time between t = 75 and 100, which
did not explicitly include calculating the L2 error in the problem defi­ indicates that the nuclei have merged into one large particle. The
nition. We choose r0 = 1.01r* for the convergence test because the snapshot in Fig. 10 is used as an illustration, showing twelve discrete
system is very sensitive when r0 = 1.00r* , and small numerical errors particles at t = 40 which is the outcome of particles merging. Notice that
can result in exponential growth of error. On the other hand, if r0 is far the seeds in Fig. 10 have equal size since they are inserted at the same
enough away from r* , numerical integration is rather robust and time t = 0. The nucleation kinetics is compared with the JMAK theory
convergence is readily achieved. Fig. 5 clearly shows that the mesh by making the Avrami plots in Fig. 9 and extracting the exponent n. Note
resolution has a significant effect on the time-evolution of the solid that we could also fit the solid fraction Y(t) directly to the functional
fraction. When the mesh is of low resolution, i.e., Δx = 0.8, only a small form given by the JMAK theory, Y(t) = 1 − exp[ − Ktn ]. Extracting the
increase in solid fraction is observed before t = 200, and the nucleus power n in JMAK theory from the Avrami plots vs. direct fit to the
grows slowly. When we refine the mesh to improve the spatial resolu­ functional form tends to give slightly different results as different points
tion, the solid fraction increases more rapidly. The calculated L2 error are weighted differently, and the Avrami plots tend to enhance the error
between simulations with consecutive meshes Δx = 0.4 and Δx = 0.2 is in the early and late stages of the transformation. However, it is more
2.78, and that between mesh sizes of Δx = 0.2 and Δx = 0.1 is 0.69, common to use the Avrami plots in experiments to obtain the kinetic
which indicates that the results from the latter pair of meshes are coefficient; therefore, we use the Avrami plots here.
similar. Further increases in the mesh resolution would not significantly When all the nucleation events happen at t = 0, the slope of the
improve the accuracy of the simulation results. This is also indicated in Avrami plot should be close to the dimension of the simulation domain,
Fig. 5 where the curves of Δx = 0.2 and Δx = 0.1 are indistinguishable. which is 2 in this case. Ideally the size of the system should be infinite for
The adaptive mesh behaves similarly to the refined fixed mesh. We thus the JMAK theory to hold, but this is not realistic in our simulation. We fit
treat the solution generated with Δx = 0.1 as a gold standard solution a straight line to the data ranging from 1.1 < logt < 2.1. The average
for comparison with the results of other numerical solution methods and slope of the fitted line for the five runs is 〈n〉 = 2.01, and the standard
implementations.

Fig. 6. The total free energy evolution of the multiple seed homogeneous
Fig. 5. Convergence test for the solid fraction change for r0 = 1.01r * . nucleation with all seeds nucleating at t = 0.

6
W. Wu et al. Computational Materials Science 193 (2021) 110371

Fig. 7. The solid fraction of the multiple seed homogeneous nucleation with all
seeds nucleating at t = 0.

Fig. 10. The phase field domain of ϕ at t = 40 of the multiple seed homoge­
neous nucleation with all seeds nucleating at t = 0.

deviation is 0.0376. In addition, we see that the Avrami plots of the five
runs with different initial random positions overlap with each other very
well. This indicates that the initial positions of the nuclei have little
influence on the nucleation kinetics.

5.3. Problem I.3 – Homogeneous nucleation, multiple seeds at random

times

For this part of the homogeneous nucleation benchmark problem we

examine the nucleation behavior when multiple seeds are inserted at
random times. We use the same metrics as in the previous problem,
(Fig. 11–14) with also a snapshot of the phase field domain at t = 100 as
shown in Fig. 15. Just as in Problem I.2, different colors in the figures
Fig. 8. The number of discrete particles of the multiple seed homogeneous represent runs with different random seeds.
nucleation with all seeds nucleating at t = 0. Compared to what we observed for homogeneous nucleation with
multiple seeds at the same time, the total free energy curve shows a
similar decreasing trend but it is now not necessarily monotonic. This is
because every time we insert a nucleus there is an increase in the total
free energy. However, this energy increase is so small, about four orders
of magnitude less than the total energy change in the entire time range,
so one cannot really notice corresponding glitches in the energy plot.
The solid fraction increases monotonically. The main difference is

Fig. 9. The Avrami plots of the multiple seed homogeneous nucleation with all
seeds nucleating at t = 0. The average slope fitted within the range of 1.1 <
logt < 2.1 is 2.01, and the standard deviation is 0.0376.

Fig. 11. The total free energy evolution of the multiple seed homogeneous
nucleation with seed nucleating at random times.

7
W. Wu et al. Computational Materials Science 193 (2021) 110371

Fig. 12. The solid fraction of the multiple seed homogeneous nucleation with
seed nucleating at random times.

Fig. 15. The phase field domain of ϕ at t = 100 of the multiple seed homo­
geneous nucleation with seed nucleating at random times.

related to the shape change of the solid-fraction curve. Transformations

are observed to follow a characteristic sigmoidal profile (as shown in
Fig. 12), where the transformation rates are low at the beginning and the
end of the transformation but rapid in between. (In contrast, for ho­
mogeneous nucleation with multiple seeds at t = 0, the Avrami plots are
approximately straight lines after the initial slow growth, Fig. 9.) The
initial slow rate can be attributed to the time required for forming a
significant number of nuclei of the solid phase.2 During the intermediate
period the transformation is rapid as the nuclei grow into particles and
consume the old liquid phase while nuclei continue to form in the
remaining liquid phase. Once the transformation nears completion the
liquid phase in the domain has almost all been consumed: there is not
much space left for nuclei to form and the production of new particles
slows down. Furthermore, the particles already existing begin to over­
lap, forming a boundary where growth of solid phase stops. Fig. 15 is an
Fig. 13. The number of discrete particles of the multiple seed homogeneous
example showing the growth of the particles at t = 100. Notice that in
nucleation with seed nucleating at random times.
contrast with the snapshot shown in Fig. 10, this time the particles have
different sizes since the seeds are inserted at different times. The number
of discrete particles is 0 at t = 0. As the time increases up to around t =
200, more new nuclei are added to the system at rate that outpaces
nuclei merging events, resulting in an overall increase in the number of
discrete particles. After that, the merging rate is greater than the
insertion rate. Therefore, the number of discrete particles starts to
decrease and finally drops to unity as in the homogeneous nucleation
with seeds at t = 0. For a quantitative comparison of the nucleation
kinetics, we calculate the slope of the Avrami plots. This time when
nucleation happens at different random times, and the slope of the
Avrami plots should be around 3.0, according the JMAK theory (two
dimensions of growth plus one, a constant growth rate). We fit a straight
line to the data ranging from 1.5 < logt < 2.5. The average slope of the
fitted line for the five runs is 〈n〉 = 3.07, and the standard deviation is
0.180. Although the Avrami plots of the five runs with different initial
random positions do not overlap as well as for the case with all seeds at
t = 0, the average slope is close to the ideal value 3.0 and the variance is
small. This again indicates that the nucleation kinetics do not depend on

Fig. 14. The Avrami plots of the multiple seed homogeneous nucleation with
seed nucleating at random times. The average slope fitted within the range of 2
The initial near-vertical lines for a few of the plots is an artifact of an
1.5 < logt < 2.5 is 3.07, and the standard deviation is 0.180.
aggressive adaptive time stepper: if the time step has grown too large before the
initial seed is inserted, the code does not tag the insertion time accurately and
the growth between the first and second time step after insertion is inaccurate.

8
W. Wu et al. Computational Materials Science 193 (2021) 110371

the choice of the pseudorandomly generated seed.

5.4. Problem II – Athermal heterogeneous nucleation

In this benchmark problem we study how the driving force for so­
lidification influences the free-growth condition of athermal heteroge­
neous nucleation. Fig. 16 shows what the system looks like at t = 6500
under different driving forces.
When the applied undercooling driving force Δf = 1.00Δf0 , the
particle will reach its free growth limit, which is the half-circular
configuration, and its radius is equal to the radius of the homogeneous
nucleus corresponding to the given driving force. When Δf = 0.98Δf0 ,
the corresponding critical radius increases, the newly formed nucleus on
the surface cannot grow freely, its final shape could be treated as part of
a circle with larger radius compared to the previous case. The solid
fraction of these two cases increases at the beginning but would reach a
plateau very soon (black solid line and red dashed line in Fig. 17). When Fig. 17. The solid fraction of the athermal heterogeneous nucleation for Δf =
0.98Δf0 , Δf = 1.00Δf0 , Δf = 1.02Δf0 , and Δf = 1.04Δf0 .
we increase the undercooling driving force to Δf = 1.02Δf0 , the cor­
responding critical radius decreases, the particle is able to grow freely
and almost takes up the space of the entire system. The corresponding
solid fraction would keep increasing until it becomes close to 1. We
remark that the system is very sensitive to numerical errors around Δf =
1.00Δf0 ; for some methods, free growth cannot be observed when Δf =
1.02Δf0 . Further increasing the driving force to Δf = 1.04Δf0 may help
the free growth to occur. Unsurprisingly, for those methods that evi­
denced free growth at Δf = 1.02Δf0 , a faster growth rate is observed
when Δf = 1.04Δf0 .
We have performed a convergence test for the solid fraction change
for an undercooling driving force of Δf = 1.1Δf0 , as shown in Fig. 18. As
we refined the mesh resolution from Δx = 0.4 to Δx = 0.025, the solid
fraction curve would gradually shift to the right – with decreasing
element size, it takes longer for the free growth to occur. Our L2 error at
t = 800 is 0.42 between mesh sizes of Δx = 0.05 and Δx = 0.025, the L2
error is 0.23 between mesh sizes of Δx = 0.025 and Δx = 0.0125.
Fig. 18 shows that visually, the difference between mesh size of Δx =
0.05 and Δx = 0.025 is already barely discernible. It is interesting to Fig. 18. The convergence test for the solid fraction change in the case Δf =
note that the result for a mesh size of Δx = 0.0125 (not shown) is 1.1Δf0 .
indistinguishable from using adaptive meshing – and at least in our case
the run-time for adaptive meshing was smaller by orders of magnitudes nucleation problem illustrates that an initial particle will shrink and
than the run-time for a fixed mesh with Δx = 0.0125. Please note that dissolve when its initial radius is smaller than the critical radius, and
the undercooling Δf = 1.1Δf0 is safely above the critical undercooling, that it will grow when its initial radius is larger than the critical radius.
shown by the fact that the transition happens rapidly, without any initial The second and third parts of this problem provide simple illustrations of
period of slow growth or “hesitation”. As the system approaches the nucleation kinetics and connect the results to the JMAK theory. When
critical undercooling, the system becomes more sensitive to small de­ the nucleation happens at t = 0, the slope of the Avrami plot is
tails, such as the solver type and the numerical resolution used; there­ approximately equal to d, where d is the domain dimension. When the
fore, one should expect slower or worse convergence and also reduced nucleation happens at random times, the slope of the Avrami plot is
reproducibility between solvers when closer to the critical value. approximately d + 1. The second problem demonstrates the free growth
condition for the athermal heterogeneous nucleation. Free growth will
6. Conclusion only occur when the driving force of solidification is larger than the
critical driving force corresponding to the critical radius.
The benchmark problems on homogeneous and heterogeneous As benchmark problems are aimed at probing numerical imple­
nucleation presented in this work are an addition to our ongoing work to mentations, the problems highlight several potential issues. Numerical
develop benchmark problems for phase field modeling [1–3]. It is our solutions are sensitive to any perturbations or numerical noise when
hope that the discussions of these problems are of instructional value to initial radii are close to the critical radius r* . This is evident in Problem
practitioners, teachers, and students. On a pedagogical level, the first

Fig. 16. The snapshot for the athermal heterogeneous nucleation at t = 6500 when (a) Δf = 0.98Δf0 , (b) Δf = 1.00Δf0 , (c) Δf = 1.02Δf0 .

9
W. Wu et al. Computational Materials Science 193 (2021) 110371

I.1 (single-seed nucleation). Depending on the numerical implementa­ Conceptualization, Methodology, Software, Investigation, Data cura­
tion and the specific code, the solid fraction at an initial radius r0 = r* tion, Writing - original draft, Writing - review & editing, Visualization,
could either increase or diminish (Fig. 4). Similarly, for the athermal Supervision.
nucleation (Problem II), different numerical implementations and codes
may give different results close to critical driving force (Fig. 17), with Declaration of Competing Interest
the steep onset of growth occurring at later or earlier times.
A related issue is that when r0 is close to r* in single-seed homoge­ The authors declare that they have no known competing financial
neous nucleation, or close to critical driving force in single-seed athe­ interests or personal relationships that could have appeared to influence
rmal nucleation, the convergence with respect to mesh size can be slow the work reported in this paper.
and care has to be taken to ensure that the results are well converged
with respect to mesh size. This is especially the case for the athermal Acknowledgments
nucleation close to critical driving force (Δf = 1.1f0 ), where we had to
reduce the mesh size to Δx = 0.025 for well-converged solution. It is This work was performed under financial assistance award
interesting – and reassuring – to note that in both cases, adaptive 70NANB19H005 from the U.S. Department of Commerce, National
meshing led to well-converged solutions at much smaller execution time Institute of Standards and Technology as part of the Center for Hierar­
than for a fine uniform mesh. This clearly demonstrates the power of chical Material Design (CHiMaD). DM was supported by the U.S.
adaptive meshing schemes for these problems. Department of Energy, Office of Basic Energy Sciences, Division of
The slopes of the Avrami plots, especially for multiple seeds at t = 0, Materials Sciences and Engineering under Award #DE-SC0008637 as
are usually found to be larger than the ideal value d. The reason for this part of the Center for PRedictive Integrated Structural Materials Science
is subtle: the JMAK model assumes circular particles, sometimes over­ (PRISMS Center) at University of Michigan. TP and LG were supported
lapping. At the points of overlap, there will be cusps in the boundary of by the National Agency for Research, Development, and Innovation,
the merged particles (see, e.g., Fig. 10). Because the Cahn–Hilliard Hungary (NKFIH contract No. KKP-126749). We gratefully acknowledge
equations correctly incorporate the Gibbs–Thomson effect, the time- the computing resources provided on Bebop and Blues, high-
evolution of the systems seeks to smooth out cusps and make them performance computing clusters operated by the Laboratory
straighter. This leads to a faster growth of the solid fraction than pre­ Computing Resource Center at Argonne National Laboratory. We
dicted by the JMAK theory. This is particularly evident at smaller particularly thank the participants of the Phase Field Workshops held in
driving forces where contributions to the free energy from the Gibbs–­ Evanston, IL, to whom we are deeply indebted for invaluable feedback
Thomson effect are relatively larger. We initially ran simulations for and comments. We also thank Dr. T. Keller (NIST) and Dr. M. Donahue
√̅̅̅
Problem I.2 (multiple seeds at t = 0) with a driving force Δf = 2/12. for helpful and insightful comments.
This resulted in slopes in the Avrami plots of close to 2.3. Increasing the
√̅̅̅
driving force by a factor of 2 to Δf = 2/6 has the effect of decreasing References
the interfacial energy contribution relative to the bulk free energy,
which diminishes the Gibbs–Thomson effect and slows down the growth [1] A.M. Jokisaari, P.W. Voorhees, J.E. Guyer, J.A. Warren, O.G. Heinonen,
of the solid fraction so the slope in the Avrami plot approaches the ideal Benchmark problems for numerical implementations of phase field models,
Comput. Mater. Sci. 126 (2017) 139–151.
value. We confirmed that for the system sizes used here (500 × 500), the [2] A.M. Jokisaari, P.W. Voorhees, J.E. Guyer, J.A. Warren, O.G. Heinonen, Phase field
slopes in the Avrami plots were not affected by finite-size effects. On the benchmark problems for dendritic growth and linear elasticity, Comput. Mater. Sci.
other hand, Problem I.3 (multiple seeds at random times) gave slopes 149 (2018) 336–347.
[3] A.M. Jokisaari, W. Wu, P.W. Voorhees, J.E. Guyer, J.A. Warren, O.G. Heinonen,
very close to the ideal value of d +1 without increasing the driving force Phase field benchmark problems targeting fluid flow and electrochemistry,
√̅̅̅
from Δf = 2/12. We believe this is because the constant-rate nucle­ Comput. Mater. Sci. 176 (2020) 109548.
[4] L. Gránásy, T. Börzsönyi, T. Pusztai, Nucleation and bulk crystallization in binary
ation yields a better average over microscopic processes thereby
phase field theory, Physical Review Letters 88 (2002) 206105.
diminishing the effect of cusps. [5] M. Castro, Phase-field approach to heterogeneous nucleation, Phys. Rev. B 67
The development of these benchmark problems, as the ones before (2003) 035412.
them, has relied very heavily on comments and feedback from the [6] J.P. Simmons, Y. Wen, C. Shen, Y. Wang, Microstructural development involving
nucleation and growth phenomena simulated with the phase field method, Mater.
community. It is a great experience to work in this way with an enthu­ Sci. Eng.: A 365 (2004) 136–143.
siastic and engaged community. In order to make these benchmark [7] L. Gránásy, T, T. Pusztai, D. Saylor, J.A. Warren, Phase field theory of
problems as useful as possible, we urge the community to continue to heterogeneous crystal nucleation, Physical Review Letters 98 (2007) 035703.
[8] J.A. Warren, T. Pusztai, L. Környei, L. Gránásy, Phase field approach to
provide feedback for existing and possible additional benchmark prob­ heterogeneous crystal nucleation in alloys, Physical Reviews B 79 (2009) 014204.
lems at https://pages.nist.gov/pfhub/, and to implement these bench­ [9] T.W. Heo, L. Chen, Phase-field modeling of nucleation in solid-state phase
mark problems and upload completed results. transformations, J. Minerals, Metals Mater. Soc. 66 (2014) 1520–1528.
[10] L. Gránásy, G. Tóth, J.A. Warren, F. Podmaniczky, G. Tegze, L. Rátkai, T. Pusztai,
Phase-field modeling of crystal nucleation in undercooled liquids – a review,
CRediT authorship contribution statement Progress in Materials Science 106 (2019) 100569.
[11] R. Kubo, The fluctuation-dissipation theorem, Reports Progess Phys. 29 (1966)
255–284.
W. Wu: Conceptualization, Methodology, Software, Validation,
[12] J.P. Simmons, C. Shen, Y. Wang, Phase field modeling of simultaneous nucleation
Investigation, Data curation, Writing - original draft, Writing - review & and growth by explicitly incorporating nucleation events, Scripta Mater. 43 (2000)
editing, Visualization. D. Montiel: Conceptualization, Methodology, 935–942.
[13] R. Shi, S. Khairallah, T.W. Heo, M. Rolchigo, J.T. McKeown, M.J. Matthews,
Software, Validation, Writing - review & editing. J.E. Guyer: Concep­
Integrated simulation framework for additively manufactured Ti-6Al-4V: melt pool
tualization, Methodology, Software, Validation, Writing - review & dynamics, microstructure, solid-state phase transformation, and microelastic
editing, Supervision. P.W. Voorhees: Conceptualization, Methodology, response, J. Miner., Metals Mater. Soc. 71 (2019) 3640–3655.
Writing - review & editing, Supervision. J.A. Warren: Conceptualiza­ [14] W.A. Johnson, R.F. Mehl, Reaction kinetics in processes of nucleation and growth,
Am. Inst. Min. Metal. Petro. Eng. 135 (1939) 416–458.
tion, Methodology, Writing - review & editing, Supervision. D. [15] M. Avrami, Kinetics of phase change. I General theory, J. Chem. Phys. 7 (1939)
Wheeler: Data curation, Writing - review & editing, Visualization. L. 1103–1112.
Gránásy: Conceptualization, Methodology, Software, Validation, [16] M. Avrami, Kinetics of phase change. II Transformation-time relations for random
distribution of nuclei, J. Chem. Phys. 8 (1940) 212–224.
Investigation, Writing - review & editing, supervision. T. Pusztai: [17] M. Avrami, Granulation, phase change, and microstructure kinetics of phase
Conceptualization, Methodology, Software, Validation, Investigation, change. III, J. Chem. Phys. 9 (1941) 177–184.
Writing - review & editing, Supervision. O.G. Heinonen: [18] A.N. Kolmogorov, On the statistical theory of the crystallization of metals, Bull.
Acad. Sci. USSR 1 (1937) 355–359.

10
W. Wu et al. Computational Materials Science 193 (2021) 110371

[19] T.E. Quested, A.L. Greer, Athermal heterogeneous nucleation of solidification, Acta [26] C. Shen, J. Li, Y. Wang, Finding critical nucleus in solid-state transformations,
Mater. 53 (2005) 2683–2692. Metall. Mater. Trans. A 39 (2008) 976–983.
[20] A.L. Greer, A.M. Bunn, A. Tronche, P.V. Evans, D.J. Bristow, Modelling of [27] D. Gaston, J. Peterson, C. Permann, D. Andrs, A. Slaughter, J. Miller, Continuous
inoculation of metallic melts: application to grain refinement of aluminium by integration for concurrent computational framework and application development,
Al–Ti–B, Acta Materialia 48 (2000) 2823–2835. J. Open Res. Softw. 2 (2014).
[21] A.L. Greer, Overview: Application of heterogeneous nucleation in grain-refining of [28] D.R. Gaston, C.J. Permann, J.W. Peterson, A.E. Slaughter, D. Andrš, Y. Wang, M.
metals, J. Chem. Phys. 145 (2016) 211704. P. Short, D.M. Perez, M.R. Tonks, J. Ortensi, et al., Physics-based multiscale
[22] R. Shi, C. Shen, S.A. Dregia, Y. Wang, Form of critical nuclei at homo-phase coupling for full core nuclear reactor simulation, Ann. Nucl. Energy 84 (2015)
boundaries, Scripta Mater. 146 (2018) 276–280. 45–54.
[23] L. Zhang, L.-Q. Chen, Q. Du, Morphology of critical nuclei in solid-state phase [29] J.E. Guyer, D. Wheeler, J.A. Warren, FiPy: Partial differential equations with
transformations, Phys. Rev. Lett. 98 (2007) 265703. Python, Computing Sci. Eng. 11 (2009) 6–15.
[24] D. Wheeler, T. Keller, S.J. DeWitt, A.M. Jokisaari, D. Schwen, J.E. Guyer, L. [30] S. DeWitt, S. Rudraraju, D. Montiel, W.B. Andrews, K. Thornton, PRISMS-PF: A
K. Aagesen, O.G. Heinonen, M.R. Tonks, P.W. Voorhees, et al., PFhub: The phase- general framework for phase-field modeling with a matrix-free finite element
field community hub, J. Open Res. Softw. 7 (2019). method, npj Computational Materials 6 (2020) 1–12.
[25] S.M. Allen, J.W. Cahn, A microscopic theory for antiphase boundary motion and its
application to antiphase domain coarsening, Acta Metall. 27 (1979) 1085–1095.

11