Atomistic mechanisms of binary alloy surface segregation from

nanoseconds to seconds using accelerated dynamics

Richard B. Garza1,2 , Jiyoung Lee3,4, Mai H. Nguyen3, Andrew Garmon5,6, Danny

Perez5, Meng Li2, Judith C. Yang2, Graeme Henkelman3,4, Wissam A. Saidi1*

1
Department of Mechanical Engineering and Materials Science, University of
Pittsburgh, Pittsburgh, PA
2
Department of Chemical and Petroleum Engineering, University of Pittsburgh,
Pittsburgh, PA
3
Department of Chemistry, University of Texas at Austin, Austin, TX
4
Oden Institute for Computational Engineering & Sciences, University of Texas at
Austin, Austin, TX
5
Theoretical Division T-1, Los Alamos National Laboratory, Los Alamos, NM
6
Department of Physics & Astronomy, Clemson University, Clemson SC

*
Corresponding author. E-mail: [email protected]

Abstract

Although the equilibrium composition of many alloy surfaces is well understood, the rate of

transient surface segregation during annealing is not known, despite its crucial effect on alloy

corrosion and catalytic reactions occurring on overlapping timescales. In this work, CuNi

bimetallic alloys representing (100) surface facets are annealed in vacuum using atomistic

simulations to observe the effect of vacancy diffusion on surface separation. We employ multi-

timescale methods to sample the early transient, intermediate, and equilibrium states of slab

surfaces during the separation process, including standard MD as well as three methods to

perform atomistic, long-time dynamics: parallel trajectory splicing (ParSplice), adaptive kinetic

Monte Carlo (AKMC), and kinetic Monte Carlo (KMC). From nanosecond (ns) to second

timescales, our multiscale computational methodology can observe rare stochastic events not

typically seen with standard MD, closing the gap between computational and experimental

timescales for surface segregation. Rapid diffusion of a vacancy to the slab is resolved by all four

1
methods in tens of ns. Stochastic re-entry of vacancies into the subsurface, however, is only seen

on the microsecond timescale in the two KMC methods. Kinetic vacancy trapping on the surface

and its effect on the segregation rate are discussed. The equilibrium composition profile of CuNi

after segregation during annealing is estimated to occur on a timescale of seconds as determined

by KMC, a result directly comparable to nanoscale experiments.

1. Introduction

Alloy surfaces are typically enriched with one of their constituent elements, particularly in the

top layers, due to differences in the surface energies of the pure metals. This surface segregation

process leads to metallic de-mixing, which is of relevance to many different fields of research

such as catalysis and metallurgy, considering that in situ transformations can affect the chemical

activity or structural integrity gained from homogeneously alloying pure metals together. Alloy

surface segregation and ordering have been measured experimentally both in vacuum and gas

environments, with differences in the equilibrium alloy composition induced near the exposed

top layers for many bimetallic alloys[1-6]. Consequently, studies of alloy surface transformations

are influenced by prior phase separation in vacuum during pre-treatment. A decoupling of the

experimental environment from the pre-treatment environment is required to understand the

transient effect on non-equilibrium surface composition, since nanoscale elemental mapping is

not feasible on relatively short microsecond (µs) timescales even using electron and X-ray

diffraction.

Molecular dynamics (MD) simulations can generally resolve atomic transitions on the

picosecond (ps) to nanosecond (ns) timescales. While these short timescale dynamics can be

easily investigated using energetic models based on either classical force fields or first-principles

2
methods, longer timescales that are often more relevant experimentally are challenging to realize

with conventional resources and techniques[7-9]. As shown in the current work, alloy surface

segregation occurs over millisecond (ms) timescales that are impractical to obtain using

conventional MD simulations. This goal can be achieved with accelerated methods including

adaptive kinetic Monte Carlo (AKMC) and temperature-accelerated dynamics (TAD)[10-12].

These two methods were employed in a multiscale approach to study surface segregation in

ceramics (rocksalt oxides) during oxidation[13]. Also, more recently, surface segregation and

timescales in PdAu nanoparticles have been studied using AKMC[14]. However, alloy

segregation and timescales related to the metallic dopant dynamics are not well understood in

nanoscale thin films, as these require larger model systems to accurately simulate the region

between bulk and exposed surface. A recent KMC study investigated the dynamics in bulk NiFe

solid solution, reporting that the vacancy migration barrier is highly influenced by the local

composition, behavior that the authors predict will influence the rate of microscale phase

transformations[15]. Previously, only the dependence of activation energy on overall alloy

composition could be predicted [16]. The failure of these global models for concentrated solid

solutions is known, though they may be corrected by sampling transition energies in differing

chemical environments[17].

In this study, we model the non-equilibrium surface segregation process by employing

multiscale simulations methods that probe the dynamics from ps to second timescales. We focus

our investigations on cupronickel (CuNi), which exhibits surface segregation at elevated

temperatures in vacuum, enriching with Cu near the surface and Ni in the bulk[18, 19]. The

CuNi alloy is of interest for different applications with extreme environmental conditions

3
including marine settings due to its resistance to corrosion by seawater[20, 21], as well as a high-

temperature catalyst for thermal CO2-to-syngas-to-fuels conversion[22-25].

To date, the transient dynamics of segregation in CuNi have not been investigated. Previous

studies have employed Monte Carlo (MC) to show that the Ni solute concentration is

significantly decreased in the top three surface monolayers for all slab orientations, approaching

its value in the bulk at a depth of four or five atomic layers below the surface[26-28]. While MC

simulations describe the systems in equilibrium, they do not give a timescale for the segregation

process. In contrast, MD can provide a timescale for dynamics, but it has not been utilized before

to study transient dynamics of the segregation process. Limited investigations of dislocation slip

under applied stress (work hardening) or melting in CuNi [29, 30] have been carried out using

MD because these processes occur over ns timescales, yet the impact on larger scale reordering

on experimental timescales was not explored in those studies. Thus, the mechanism, as well as

the timescale, for transformation from a randomly mixed to ordered alloy on the nano- and

microscales remains unclear.

Accelerated methods can provide insight on longer timescales into atomistic mechanisms of

surface segregation phenomena. Previously, AKMC investigations of the segregation kinetics of

PdAu nanoparticles showed greater kinetic stability due to reduced strain in the mixed-

phase[14]. In the present work, we probe segregation dynamics in planar, (100) CuNi surfaces,

employing conventional MD and three accelerated dynamics methods: parallel trajectory splicing

(ParSplice), AKMC, and KMC with kinetic barriers derived from a cluster expansion. ParSplice

affords accurate system evolution up to ~10 µs, while AKMC and KMC simulate longer

timescales up to ms and seconds, respectively.

4
 ParSplice [11] extends MD simulation times by leveraging parallel computers to carry

out parallelization in the time domain, in contrast to the usual domain decomposition approaches

that operate in the space domain [31, 32]. Thus, with ParSplice it is possible to simulate small

systems over very long timescales, again in contrast to conventional parallelization approaches

that are efficient at spatially decomposing large systems simulated over short timescales. This is

accomplished by concurrently generating a large number of independent, short trajectory

segments using a procedure that guarantees these segments can be assembled into a longer state-

to-state trajectory that is statistically correct[33]. Periodic quenching is used to identify

transitions between different metastable states and segment terminations. It can be shown that

ParSplice trajectories can become arbitrarily accurate by adjusting the estimate of the so-called

correlation time of the dynamics, at the expense of a computational overhead [34]. When the

dynamics follow from a sequence of rare events, ParSplice can provide a computational speedup

that scales with the number of processors used; it is therefore especially powerful when deployed

on massively parallel computers.

KMC-based methods do not have a fixed timestep; instead, they find the time elapsed for

the first escape from one state to another, allowing for large periods of vibrational motion in the

atomic system to be bypassed[14, 35]. These escape times correspond to reaction rates, which

are calculated adaptively or “on the fly” to construct an AKMC state model: nothing in the

output event table is predefined or assumed from prior knowledge[12, 14, 35]. AKMC uses

minimum-mode following searches or high-temperature MD to construct this event table: the

transition state energies (activation barrier heights) are found using single-ended saddle point

finding algorithms such as the dimer method[36-38]. To achieve further acceleration while

maintaining as much of this accuracy as possible, in the present study, we use off-lattice

5
dynamics from AKMC to fit a more approximate, lattice-based KMC model, which functionally

depends on the Ni-coordination via a cluster expansion.

Our multiscale approach produces a hierarchy of trajectory data for the segregation

process over a broad range of timescales. To represent trajectory data on varying scales and

fidelities, we track physical properties including the local defect chemical environment and its

influence on the correlated rates of segregation and vacancy migration. Further, we find that

segregation in the top layer is a function of the rare re-entry of a vacancy from the surface to the

subsurface, which only occurs with a frequency of 10 µs due to a high kinetic barrier, greatly

increasing the time required for the surface to be completely depleted of Ni. We also determine

that the number of dopant atoms in the 1st and 2nd coordination shells around the point defect

alters its migration energy, affecting the rate of composition change. The consistency of system

evolution during segregation by thermal annealing across all the accelerated methods is

examined on many timescales; this evidence supports each technique’s further use in multiscale

simulations in combination with the data processing methods used in this work.

2. Methodology

2.1 Validation of the Embedded-Atom Method Potential

A reliable interatomic potential is required to obtain accurate MD of surface segregation and

equilibrated structures. Here, we employed the embedded atom model (EAM) potential of

Fischer and collaborators, which is designed to model CuNi phase segregation across grain

boundaries and is parameterized with surface energies, lattice constants, vacancy migration

energies, and relevant quantities governing the rate of metallic phase separation [39]. We first

verified the applicability of this potential to study CuNi surface segregation by comparison of

6
calculated surface energies with those obtained from spin-polarized DFT calculations carried out

using VASP with the Perdew-Burke-Enzerhof (PBE) exchange-correlation functional[40-44],

see SI Appendix A, Table A1. Further, we verified that the segregation of Ni to the bulk—from

the 1st to the 2nd or 3rd layer—is always favorable i.e. ΔENi,surf→interior ≈ -0.4 to -0.3 eV from DFT

and EAM, also in agreement with previous ab initio calculations [45]. Further, we found good

agreement with the previously reported anisotropic, thermodynamic tendency for Ni to segregate

to exposed <100> and <110> facets in favor of <111> surfaces[19, 27].

2.2 Slab Models and Initial State for Annealing

We employed slab models of the (100) surface termination that are composed of 216 and 384

atoms by randomly substituting Cu with Ni. These slabs were 12 monolayers in thickness with

3x3 and 4x4 surface periodicity; nearly doubling the vacancy concentration (1:216 vs. 1:384

atoms) did not affect our observed mechanisms or energetics. Thus, systems of smaller surface

periodicity (3x3, 216 atoms) were simulated to realize the longest timescales for higher fidelity

time-averaging and activation energy histograms. MC simulations across the composition range

of 2.7 – 16at% Ni, and temperature range of 300 – 700 K showed no detectable variation in the

amount of Ni segregated. For this reason, we chose to exclusively simulate systems with 16at%

Ni at 500 K with the accelerated MD methods.

The equilibrium composition profile of the CuNi (100) surface alloy is determined using MC.

The method attempts 106 MC swaps of Cu and Ni atoms with subsequent minimization and an

acceptance probability as determined via the Metropolis algorithm[46]. MC composition profiles

are found to be consistent with similar reported profiles [18, 26-29, 47] across a range of

compositions and temperatures, see SI Appendix B. Such agreement with previous

7
computational and experimental results is a further validation of the reasonable accuracy of the

EAM potential.

2.3 Multiscale Simulation Hierarchy

The methods employed in our multiscale hierarchy are described as follows, in order of

accessible timescale: molecular dynamics in the NVT ensemble was generated using a Langevin

thermostat, as implemented in the LAMMPS code[48]. Many instances of LAMMPS were then

orchestrated by the EXAALT/ParSplice code [http://gitlab.com/exaalt] to carry out long-time

ParSplice simulations. Direct MD realized ns of simulated time using the velocity Verlet

algorithm and modest resources (4 processors). ParSplice accelerated these dynamics up to µs

timescales using 224 processors. Both standard MD and ParSplice were carried out in the same

way: after initializing particle velocities according to a Boltzmann distribution representing the

target temperature, the systems were equilibrated with a 2 fs timestep and a 1 ps temperature

relaxation time for 200 ps. After the equilibration phase of MD, a further 400 ns were simulated

to constitute the production phase, while ParSplice accelerated this simulation time up to 35 µs.

Minimization was done with the conjugate gradient algorithm using convergence criteria of 10-8

eV and 10-6 eV/Å for the energy and forces[49], respectively. Quenching is necessary for

ParSplice to identify state transitions and to properly terminate each parallel MD replica.

In contrast to ParSplice and MD, the AKMC and KMC models use a non-constant

timestep, which varies to match the timescale of the first escape time from a given state. Over

longer timescales, up to many seconds, KMC-based acceleration approaches the segregation

found with MC [12, 35]. KMC methods simulate the time evolution of the system, requiring a

predetermined event table in which the kinetic rate of each event is approximated through the

8
Arrhenius relation to the pre-calculated activation energy. At each step, a random event i is

selected from the table in the order from 1 to i with the condition
!"# $ !

! 𝑟!"# < 𝑝# ! 𝑟$ ≤ ! 𝑟!
# # #

where ∑!# 𝑟! is the sum of the rate from event 1 to event i, ∑!"#
# 𝑟!"# is the sum of the rate from

event 1 to event i–1, 𝑝# is a random number between 0 and 1, and ∑$

# 𝑟$ is the total rate of the

event table. Time is then incremented by

−ln(𝑝% )
∑$
# 𝑟$

where a random number 𝑝% is drawn between 0 and 1. AKMC allows the system to find all

potential events without a need of a predetermined event table [45]. In order to search for events,

AKMC uses high-temperature MD and the climbing-image nudged elastic band (CI-NEB)

approach to calculate the saddles for the new states using the EON software[50].

To reach even longer timescales with KMC, an energy estimation based on the local

environment of the vacancy was generated by a cluster expansion. The energy was predicted to

be dependent on the concentration of Ni and Cu located in the first nearest neighbor shell of the

vacancy. In the FCC CuNi alloy, there are eight nearest neighbors for the vacancy on the surface

and twelve nearest neighbors for the vacancy in the subsurface. Sequentially, the energy-fitting

model was used to determine the barrier and rate of the events in the event table based upon the

trajectory data from our AKMC simulations. More information about the cluster expansion

method is provided in the SI, Appendix C. Further, we found satisfactory agreement between

AKMC and KMC by carrying out the dynamics up to 300 µs, further validating the KMC model

(see SI, Appendix D).

9
 Finally, the concept of an equilibrium rate was introduced to further accelerate the

timescale accessible by KMC. In this approximation, no rate was allowed to be larger than the

specified equilibrium rate with the assumption that all states connected by rates faster than the

equilibrium rate should already be in equilibrium. In the case of CuNi segregation, the planar

diffusion of a vacancy on the exposed surface is rapid, equilibrating on much shorter timescales

than for defect re-entry to the subsurface to occur even once. Ultimately, the effect of an artificial

equilibrium rate in KMC simulations is that these key transitions can be sampled more

effectively instead of the many horizontal transitions that do not alter the composition with

respect to surface depth.

In both AKMC and KMC simulations, the temperature was set to 500K and the prefactor

for the rates was fixed at 5×1012 s-1. The optimizer used in AKMC was L-BFGS, with a

convergence criterion of 0.01 eV/Å[49]. System evolution from AKMC reached ms timescales

running for a week on 24 cores. The following KMC timings are the average of five separate

runs in both cases: KMC realized seconds of simulation time running on a single processor for 2

hours with the added equilibration rate. KMC without the equilibrium rate was only simulated up

to 10 ms, since only one minute of wall clock time was necessary before the surface vacancy

trapping described above halted Ni segregation in the system.

3. Results and Discussion

The equilibrium composition profiles obtained with MC (Figure 1) exhibit Ni migration out of

the top three surface layers, while the concentration of the fourth and fifth layers approach the

bulk value. This trend did not quantifiably vary with changes in system temperature or Ni

concentration. The near-surface Ni concentration observed with the EAM potential agrees with

the profiles derived from previous MC simulations [18, 26-29] as well as experiments[1-3, 19].

10
Figure 1: Ni composition profiles of Cu-16at%Ni (100) surface as a function of layer depth after

MC annealing at 500 K. The slab has 12 layers with layers 1 and 12 exposed to the vacuum.

Composition is normalized by the number of atoms in a pristine FCC <100> layer (18) resulting

in a mismatch between the bulk composition of layer 6 (~22at% Ni) and the overall composition

(16at% Ni).

Next, we probed the segregation dynamics using MD, ParSplice, AKMC, and KMC

starting from the same initial configuration of the random alloy. Lattice vacancies are the

primary defect responsible for alloy segregation: the only other mechanism, self-interstitial

migration of metal atoms to octahedral or tetrahedral sites, is destabilized by greater formation

and migration energies than those for the point defect (vacancy)[51]. Hence, our simulations

included a single vacancy to facilitate surface segregation.

11
Figure 2: (left) Ni content in the top layer of Cu-16at%Ni(100) with a vacancy for (a) standard

MD, (b) ParSplice, (c) AKMC, (d) KMC, and (e) KMC with applied equilibrium rate.

Composition is normalized as in Figure 1. (right) Orthographic and side views of the final slab

model from each simulation are provided for reach simulation.

Figure 2 shows the composition over time within the top layer of our Cu-16at%Ni (100)

slab model for a single simulation of each type within our methodology. The concentration of Ni

12
atoms in the top surface layer decreases over time as it transitions from a uniform distribution to

that of equilibrium, as shown by the MC simulations in Figure 1. The accessible simulation time

increases sequentially for each method (MD, ParSplice, AKMC and KMC). In MD (Figure 2a),

the vacancy diffuses to the surface at 0.3 µs, displacing a Ni atom to the subsurface; up to the

total time of 0.4 µs the vacancy remains trapped on the surface so that the Ni concentration does

not change. With over thirty different MD simulations, we observed that the time for vacancy

percolation to the surface was consistently less than 1 µs, resulting in similar degrees of

segregation as shown in Figure 2a.

ParSplice (Figure 2b) shows similar dynamics as in standard MD, with vacancy

migration to the surface in a fraction of a µs. Throughout the simulation, the ParSplice trajectory

visited 8,278 topologically unique states while making 37,458 transitions, the vast majority of

which occurred after the vacancy had reached the surface. The parallel acceleration of ParSplice

increased the timescale from the standard MD simulation by a factor of ~100 up to 35 µs, but

even at this longer timescale the vacancy remained trapped on the surface and no additional Ni

segregation was observed. As expected, ParSplice is consistent with MD on timescales where

they overlap (0.4 µs) not only in terms of bulk vacancy diffusion but also the timescale at which

the vacancy diffuses to the surface. We have repeated the ParSplice simulations dozens of times

with different random seeds observing early vacancy diffusion to the surface (before 1 µs) in

nearly all trials, just as with MD.

With AKMC, we increased the simulation timescale by another order of magnitude

(Figure 2c): a total of 144,597 transitions evolved the system through 32,369 unique states (a

similar ratio of transitions to new states found as in ParSplice). Note that our AKMC approach

uses coarse-graining following the MC with absorbing Markov chains (MCACM) method,

13
allowing many more transitions to be considered via an analytic solution to the rate

equations[52]. From the AKMC dynamics, we can observe events in which the vacancy moves

from the surface to the subsurface at a timescale of roughly 50 µs. Over the simulation time of

300 µs, five such events were observed resulting in two Ni atoms and 3 Cu atoms migrating from

the surface to the subsurface.

With our KMC model (Figure 2d), we executed 4.7 million transitions generating 10 ms

of simulated time. Over these timescales, the system appears to approach equilibrium, with

fluctuations in the surface concentration between 0 and 5% Ni. However, what is not obvious

from these plots is that the vacancy spends all of its time in the first and second layer, so that Ni

segregation only occurs between the surface and subsurface layers. This behavior originates from

a disparity in barrier heights that embodies the “low-barrier problem”, more aptly referred to as

the “heterogeneous barrier” problem. Dynamics with a mix of low and high activation barriers

are inherently more difficult to accelerate since groups of states interconnected by low barriers

will always dominate the trajectory during naïve state space exploration[8, 11]. The barrier for

vacancy surface diffusion is 0.4-0.5 eV, whereas the barrier for the vacancy to go subsurface is

0.8-0.9 eV. Thus, we are simulating on the order of a million KMC steps with the vacancy

mostly diffusing on the surface for one subsurface diffusion event, offering a low chance for

segregation to occur. Even factoring in the small cost of each KMC step, this makes it impossible

to simulate an equilibrium distribution of Ni in the top three layers.

To further accelerate the dynamics and mitigate the “heterogeneous barrier” problem, in

our final simulation (Figure 2e), we perform KMC with the equilibrium rate approximation

described in the methods section, realizing 160 million transitions to reach a simulated time of 1

s. Here, the idea is that the vacancy will quickly reach local equilibrium diffusing in the top

14
layer, and no new states of interest are explored until subsurface diffusion occurs. Since vacancy

surface diffusion occurs on a timescale of ns, and diffusion of the vacancy to the subsurface

occurs on a timescale of µs, we chose an equilibrium rate of 40/µs to slow the surface diffusion

and accelerate the diffusion of the vacancy to the subsurface and below. Figure 2e shows that on

a timescale of seconds, the surface Ni concentration fluctuates around equilibrium after 0.4

seconds until the end of our simulation lasting 1 s.

In order to more accurately estimate the timescale required to obtain the equilibrium

profile, we performed five individual KMC simulations with the same equilibrium rate (40/µs),

extending the total time to 2.5 second corresponding to about 400 million transitions. Figure 3

shows the averaged Ni composition obtained from these KMC simulations in the bottom three

layers (referred as layer 1, 2 and 3) and the top three layers (referred as layer 12, 11, and 10) with

the equilibrium composition profiles obtained by MC together for comparison. From Figure 3a,

we can see that layer 1 reaches the equilibrium profile in less than 0.1 second and even

approaches zero Ni concentration at ~1.5 second. Layer 2 exhibits more dramatic changes. The

composition started from 30% Ni, dropped to the equilibrium composition of 10at% ~1.75

second, and remained near the equilibrium until the end of the simulation. While the MC

simulation left 16% Ni in layer 3, our KMC model shows more Ni segregations leaving 10% Ni

in the layer. This might explain the higher Ni composition profiles of the top two sublayers

(layer 11 and 10) in Figure 3b. While layer 12 shows similar behavior as layer 1, layer 11 did not

reach equilibrium until 2.5 second, and layer 10 did reach the equilibrium at ~0.6 s before the

composition increased again. The overall profile, nonetheless, clearly shows an increase in the

number of Ni atoms in bulk layers, indicating thermodynamic tendency for this dopant to remain

in the bulk rather than on the surface.

15
 (a) (b)
Figure 3: Ni composition averaged over five KMC simulations for the (a) bottom three layers

and (b) top three layers over 2.5 second with an equilibrium rate of 40/µs. The bold line

represents the 0.5-second-average calculated from the lighter single-frame datapoints, and the

dashed line corresponds to the average concentration of Ni by MC calculation with EAM

potential at each layer when the system reaches equilibrium. Composition is normalized as in

Figure 1.

Apart from surface effects, the local chemical environment around a vacancy is expected

to determine the system energetics and the rate of Ni composition change. To investigate this, we

tracked the number of Ni atoms within a 5 Å radius of the vacancy. The number of Ni atoms in

the local environment ranges from 0 to 12, with three distinct “bins” formed for low, mixed, and

high Ni-content environments. The spectrum of barriers calculated during AKMC for all vacancy

migration events within this trajectory is presented as a histogram in Figure 4a, which shows the

lowest transition state energies for vacancy migration in Ni-rich regions of the alloy.

Additionally, because the system has lower Ni content than Cu, the integrated peak area is

smaller for transitions into/within Ni-rich regions than for those with mixed and Cu-rich

compositions. Vacancy migration energies are shifted closer to 0.4 eV in the Ni-rich regions than

16
for migration in Cu-rich regions according to Figure 4a. We can deduce that vacancy migration

is favored in the Ni-rich regions, contributing to the lower dwell time near Ni as the vacancy

more rapidly diffuses towards and away from this dopant. Vacancies must slowly explore the

Cu-rich regions of the host lattice before returning to possibly segregate the Ni atoms away from

the surface. The order of integrated peak areas in Figure 4a also supports this conclusion.

(a) (b)

(c) (d)
Figure 4: (a) Histogram of vacancy migration energy barriers obtained from a 319 µs AKMC

simulation. (b-d) Total residence time for the vacancy in different chemical environments for (b)

MD, (c) ParSplice, and (d) AKMC. Cumulative times are normalized by the prevalence of each

composition type in the system rendering them unphysical.

17
 The overall time spent in low Ni, mixed, and high Ni content environments is also shown

with respect to time in Figure 4b-d using a cumulative measure of residence time for MD,

ParSplice, and AKMC and re-scaling this measure by the prevalence of each environment in the

system. The number of Ni atoms in the vicinity of the vacancy modulates these timescales

substantially but in similar fashion for all methods. Further, the logarithmic trendlines and their

orders of magnitude agree across overlapping timescales for MD/ParSplice (ns) and

ParSplice/AKMC (µs): this supports the hierarchy of methods utilized herein as a theoretical

foundation to connect correlated observations evolving across many timescales for the same

process. The cumulative dwell time and thermodynamic trends match for all of the simulation

methods in the order of integrated peak areas (Figure 4a) and residence times (Figure 4b-d):

t ,-,./ ,-,./ ,-,./

&'"()*+ > t 0)123 > t 5)"()*+ . Though this agreement does not hold at very early (ps) simulation

times with low cumulative sums, particularly for MD which is noisier as compared to accelerated

methods, the trend becomes evident in the ergodic limit. The residence times of vacancy

chemical environment presented in Figure 4b-d confirm that the local composition is an effective

determinant of where a vacancy spends most of its time during annealing and segregation.

Specifically, the position of Ni in the lattice is a minor determinant of the dynamics, since the

dopant slightly biases the vacancy’s random walk by ~0.1 eV. This effect has also been

documented in Ni-Fe surfaces annealed at 1100 K using KMC, where the local composition and

the identity of the atoms exchanging during segregation (solute vs. solvent) significantly

influenced the vacancy migration energy and the measured tracer diffusion coefficient[15].

18
4. Conclusions

In summary, we have applied three accelerated dynamics methods to examine the rate of

segregation in CuNi alloy from nanosecond to second timescales, reaching 1 second of

simulation time and the equilibrium composition with KMC. This composition profile shows no

Ni on the top surface monolayer, with less than 15at% Ni in the 2nd layer compared to 22at% Ni

in the bulk on a per layer basis, in agreement with MC predictions. Though most of our

accelerated methods were used to simulate up to the µs timescale, only modified KMC dynamics

could reach the equilibrium profile obtained from MC and previous experimental observations.

Our model estimates that the timescale for segregation in the top layer to reach equilibrium is on

the order of 0.1 ms, while equilibrium segregation does not penetrate to the 3rd layer until

timescales on the order of 100 ms. The equilibrium timescales for surface segregation of any

FCC bimetal can be determined with a combination of AKMC and KMC. However, KMC-based

methods require assumptions regarding transition state theory, and they do not resolve fast

atomic exchange processes as well as ParSplice and MD. Our study shows that a model for the

relationship between solvent distribution, activation energy, and large-scale phenomena like

segregation may be developed from further simulations of bimetallics on experimental

timescales.

5. Declaration of Interests

The authors declare that they have no known competing financial interests or personal

relationships that could have appeared or considered to influence the work reported in this paper.

6. Acknowledgments

This work was funded from NSF (DMR-1809085, CHE-210231, and CMMI-1905647) as well as

resources and collaboration from the University of Pittsburgh’s Computational Resource Cluster

19
(CRC), the Texas Advanced Computing Center, and the Theoretical Division T-1 of Los Alamos

National Laboratory.

7. References

1. Bardi, U., The atomic structure of alloy surfaces and surface alloys. Reports on Progress

in Physics, 1994. 57(10): p. 939-987.

2. Brongersma, H.H., M.J. Sparnaay, and T.M. Buck, Surface segregation in Cu-Ni and Cu-

Pt alloys; A comparison of low-energy ion-scattering results with theory. Surface

Science, 1978. 71(3): p. 657-678.

3. Erdélyi, Z., et al., Investigation of the interplay of nickel dissolution and copper

segregation in Ni/Cu(111) system. Surface Science, 2002. 496(1): p. 129-140.

4. Landa, A., et al., Development of Finnis–Sinclair type potentials for Pb, Pb–Bi, and Pb–

Ni systems: application to surface segregation. Acta Materialia, 1998. 46(9): p. 3027-

3032.

5. Jiang, C. and B. Gleeson, Surface segregation of Pt in γ′-Ni3Al: A first-principles study.

Acta Materialia, 2007. 55(5): p. 1641-1647.

6. Zhevnenko, S.N., et al., Surface and segregation energies of Ag based alloys with Ni, Co

and Fe: Direct experimental measurement and DFT study. Acta Materialia, 2021. 205: p.

116565.

7. Voter, A.F., F. Montalenti, and T.C. Germann, Extending the Time Scale in Atomistic

Simulation of Materials. Annual Review of Materials Research, 2002. 32(1): p. 321-346.

20
8. Miron, R.A. and K.A. Fichthorn, Multiple-Time Scale Accelerated Molecular Dynamics:

Addressing the Small-Barrier Problem. Physical Review Letters, 2004. 93(12): p.

128301.

9. Montalenti, F., M.R. Sørensen, and A.F. Voter, Closing the Gap between Experiment and

Theory: Crystal Growth by Temperature Accelerated Dynamics. Physical Review

Letters, 2001. 87(12): p. 126101.

10. Zamora, R.J., et al., The Modern Temperature-Accelerated Dynamics Approach. Annual

Review of Chemical and Biomolecular Engineering, 2016. 7(1): p. 87-110.

11. Perez, D., et al., Long-Time Dynamics through Parallel Trajectory Splicing. Journal of

Chemical Theory and Computation, 2016. 12(1): p. 18-28.

12. Henkelman, G. and H. Jónsson, Long time scale kinetic Monte Carlo simulations without

lattice approximation and predefined event table. The Journal of Chemical Physics, 2001.

115(21): p. 9657-9666.

13. Lavrentiev, M.Y., et al., Atomistic simulations of surface diffusion and segregation in

ceramics. Computational Materials Science, 2006. 36(1): p. 54-59.

14. Li, L., et al., Adaptive kinetic Monte Carlo simulations of surface segregation in PdAu

nanoparticles. Nanoscale, 2019. 11(21): p. 10524-10535.

15. Ferasat, K., et al., Accelerated kinetic Monte Carlo: A case study; vacancy and dumbbell

interstitial diffusion traps in concentrated solid solution alloys. The Journal of Chemical

Physics, 2020. 153(7): p. 074109.

16. Helander, T. and J. Ågren, A phenomenological treatment of diffusion in Al–Fe and Al–

Ni alloys having B2-b.c.c. ordered structure. Acta Materialia, 1999. 47(4): p. 1141-1152.

21
17. Osetsky, Y.N., L.K. Béland, and R.E. Stoller, Specific features of defect and mass

transport in concentrated fcc alloys. Acta Materialia, 2016. 115: p. 364-371.

18. Hennes, M., et al., Equilibrium segregation patterns and alloying in Cu/Ni nanoparticles:

Experiments versus modeling. Physical Review B, 2015. 91(24): p. 245401.

19. Webber, P.R., M.A. Morris, and Z.G. Zhang, Crystal-face specificity in surface

segregation of CuNi alloys. Journal of Physics F: Metal Physics, 1986. 16(4): p. 413-419.

20. Jin, T., et al., Surface Characterization and Corrosion Behavior of 90/10 Copper-Nickel

Alloy in Marine Environment. Materials, 2019. 12(11).

21. Tuck, C.D.S., C.A. Powell, and J. Nuttall, Corrosion of Copper and Its Alloys, in

Reference Module in Materials Science and Materials Engineering. 2016, Elsevier.

22. Zegkinoglou, I., et al., Surface Segregation in CuNi Nanoparticle Catalysts During CO2

Hydrogenation: The Role of CO in the Reactant Mixture. The Journal of Physical

Chemistry. C, Nanomaterials and Interfaces, 2019. 123(13): p. 8421-8428.

23. Tan, Q., Z. Shi, and D. Wu, CO2 Hydrogenation to Methanol over a Highly Active Cu–

Ni/CeO2–Nanotube Catalyst. Industrial & Engineering Chemistry Research, 2018.

57(31): p. 10148-10158.

24. Austin, N., B. Butina, and G. Mpourmpakis, CO2 activation on bimetallic CuNi

nanoparticles. Progress in Natural Science: Materials International, 2016. 26(5): p. 487-

492.

25. Lortie, M., et al., Synthesis of CuNi/C and CuNi/-Al2O3 Catalysts for the Reverse Water

Gas Shift Reaction. International Journal of Chemical Engineering, 2015.

26. Good, B., G. Bozzolo, and J. Ferrante, Surface segregation in Cu-Ni alloys. Physical

Review B, 1993. 48(24): p. 18284-18287.

22
27. Foiles, S.M., Calculation of the surface segregation of Ni-Cu alloys with the use of the

embedded-atom method. Physical Review B, 1985. 32(12): p. 7685-7693.

28. Tréglia, G., B. Legrand, and P. Maugain, Surface segregation in CuNi and AgNi alloys

formulated as an area-preserving map. Surface Science, 1990. 225(3): p. 319-330.

29. Panizon, E., et al., Study of structures and thermodynamics of CuNi nanoalloys using a

new DFT-fitted atomistic potential. Physical Chemistry Chemical Physics, 2015. 17(42):

p. 28068-28075.

30. Xiang, M., et al., Shock-induced plasticity in semi-coherent {111} Cu-Ni multilayers.

International Journal of Plasticity, 2018. 103: p. 23-38.

31. Perez, D., B.P. Uberuaga, and A.F. Voter, The parallel replica dynamics method -

Coming of age. Computational Materials Science, 2015. 100: p. 90-103.

32. Voter, A.F., Parallel replica method for dynamics of infrequent events. Physical Review

B, 1998. 57(22): p. 13985-13988.

33. Perez, D., R. Huang, and A.F. Voter, Long-time molecular dynamics simulations on

massively parallel platforms: A comparison of parallel replica dynamics and parallel

trajectory splicing. Journal of Materials Research, 2018. 33(7): p. 813-822.

34. Le Bris, C., et al., A mathematical formalization of the parallel replica dynamics. Monte

Carlo Methods and Applications, 2012. 18(2): p. 119-146.

35. Xu, L. and G. Henkelman, Adaptive kinetic Monte Carlo for first-principles accelerated

dynamics. The Journal of Chemical Physics, 2008. 129(11): p. 114104.

36. Henkelman, G. and H. Jónsson, A dimer method for finding saddle points on high

dimensional potential surfaces using only first derivatives. The Journal of Chemical

Physics, 1999. 111(15): p. 7010-7022.

23
37. Kästner, J. and P. Sherwood, Superlinearly converging dimer method for transition state

search. The Journal of Chemical Physics, 2008. 128(1): p. 014106.

38. Heyden, A., A.T. Bell, and F.J. Keil, Efficient methods for finding transition states in

chemical reactions: Comparison of improved dimer method and partitioned rational

function optimization method. The Journal of Chemical Physics, 2005. 123(22): p.

224101.

39. Fischer, F., G. Schmitz, and S.M. Eich, A systematic study of grain boundary segregation

and grain boundary formation energy using a new copper–nickel embedded-atom

potential. Acta Materialia, 2019. 176: p. 220-231.

40. Kresse, G. and J. Furthmüller, Efficient iterative schemes for ab initio total-energy

calculations using a plane-wave basis set. Physical Review B, 1996. 54(16): p. 11169-

11186.

41. Kresse, G. and J. Furthmüller, Efficiency of ab-initio total energy calculations for metals

and semiconductors using a plane-wave basis set. Computational Materials Science,

1996. 6(1): p. 15-50.

42. Kresse, G. and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-

wave method. Physical Review B, 1999. 59(3): p. 1758-1775.

43. Perdew, J.P., K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made

Simple. Physical Review Letters, 1996. 77(18): p. 3865-3868.

44. Blöchl, P.E., Projector augmented-wave method. Physical Review B, 1994. 50(24): p.

17953-17979.

45. Ruban, A.V., H.L. Skriver, and J.K. Nørskov, Surface segregation energies in transition-

metal alloys. Physical Review B, 1999. 59(24): p. 15990-16000.

24
46. Beichl, I. and F. Sullivan, The Metropolis Algorithm. Computing in Science and

Engineering, 2000. 2(1): p. 65-69.

47. Sakurai, T., et al., New result in surface segregation of Ni-Cu binary alloys. Physical

Review Letters, 1985. 55(5): p. 514-517.

48. Plimpton, S., Fast Parallel Algorithms for Short-Range Molecular Dynamics. Journal of

Computational Physics, 1995. 117(1): p. 1-19.

49. Sheppard, D., R. Terrell, and G. Henkelman, Optimization methods for finding minimum

energy paths. The Journal of Chemical Physics, 2008. 128(13): p. 134106.

50. Henkelman, G. and H. Jónsson. EON: Long timescale dynamics. 2021 July 23, 2021];

Available from: https://theory.cm.utexas.edu/eon/.

51. Lam, N.Q., L. Dagens, and N.V. Doan, Calculations of the properties of self-interstitials

and vacancies in the face-centred cubic metals Cu, Ag and Au. Journal of Physics F:

Metal Physics, 1983. 13(12): p. 2503-2516.

52. Novotny, M.A., Monte Carlo Algorithms with Absorbing Markov Chains: Fast Local

Algorithms for Slow Dynamics. Physical Review Letters, 1995. 74(1): p. 1-5.

53. Donnelly, R.G. and T.S. King, Surface composition and surface cluster size distribution

of Cu-Ni alloys via a monte carlo method. Surface Science, 1978. 74(1): p. 89-108.

54. Eymery, J. and J.C. Joud, Surface segregation in binary Cu-Ni and Pt-Ni alloys using

Monte Carlo simulation. Surface Science, 1990. 231(3): p. 419-426.

55. Shumway, R.H. and D.S. Stoffer, Time Series Analysis and Its Applications (Springer

Texts in Statistics). 2005: Springer-Verlag.

56. Xiao, P. and G. Henkelman, Kinetic Monte Carlo Study of Li Intercalation in LiFePO4.

ACS Nano, 2018. 12(1): p. 844-851.

25
 Supplementary Information

Atomistic mechanisms of binary alloy surface segregation from nanoseconds

to seconds using accelerated dynamics

Richard B. Garza1,2 , Jiyoung Lee3,4, Mai H. Nguyen3, Andrew Garmon5,6, Danny Perez5, Meng
Li2, Judith C. Yang2, Graeme Henkelman3,4, Wissam A. Saidi1*
1
Department of Mechanical Engineering and Materials Science, University of Pittsburgh,
Pittsburgh, PA
2
Department of Chemical and Petroleum Engineering, University of Pittsburgh, Pittsburgh, PA
3
Department of Chemistry, University of Texas at Austin, Austin, TX
4
Oden Institute for Computational Engineering & Sciences, University of Texas at Austin,
Austin, TX
5
Theoretical Division T-1, Los Alamos National Laboratory, Los Alamos, NM
6
Department of Physics & Astronomy, Clemson University, Clemson SC
*
Corresponding author. E-mail: [email protected]

Appendix A: EAM Validation Using DFT

The surface energies of doped Cu surfaces up to 2.7at%Ni were computed using DFT and the

EAM potential of focus in this work: this was done not only to validate the forcefield’s

26
application in the study of surface segregation, but also to examine the process with ab initio

accuracy[39]. We used the Vienna Ab Initio Package (VASP) [40-42, 44] in conjunction with

Perdew-Burke-Ernzerhof (PBE) [43] exchange-correlation functional and an energy cutoff of

500 eV to expand the wavefunction. We used a 3 x 3 x 1 k-point mesh generated with a

Monkhorst-Pack routine to sample the Brillouin zone for <100>, <110>, and <111> slab models

composed of ~100 atoms. All calculations are spin polarized. We used a slab approach to model

the surfaces and included 10 Å of vacuum between exposed surfaces to mitigate spurious

interactions in the non-periodic direction. One of the host lattice atoms in the system (Cu) were

replaced with Ni at both surface (1st layer) and subsurface (2nd layer) positions. The surface

energy of a pure Cu slab model with (hkl) orientation including Natoms is calculated relative to the

bulk energy per atom ECu and the total exposed area 2A:
"#$%
7'(8 9! "5&'()* 9,#-.
+#
E+6/ = %:
.

The equation for surface energy of models containing one dopant atom is very similar to the

above, and requires the bulk energy per atom of the dopant (Ni in this study):

3-;23 <'/6 <'/6

7'(8
E8 − (N.,-07 − 1)E&' − E5)
E+6/ =
2A

The pure Cu surface energies for orientations with low Miller index (h2+k2+l2 < 3) are

shown in Table A1, which were calculated with DFT and EAM. As is known from previous

calculations and measurements of surface energy for pure Cu, <111> surfaces are most

thermodynamically stable, followed by <100> and <110> surfaces. The validation with pure Cu

surfaces finds 2.6% average relative error between the EAM potential and first-principles. We

then replaced one atom in these structures with Ni to study the effect of doping at different

surface depths.

27
 DFT (J/m2) EAM (J/m2)

Cu (100) 1.412 1.357

Cu (110) 1.477 1.488

Cu (111) 1.208 1.246

Table A1: Surface energies of pure Cu from DFT and the EAM potential of Fischer et al

Oriented slabs along (100), (110), and (111) planes exhibit segregation trends as a

function of crystallographic direction in the FCC system for Ni, evident in Table A2. EAM

calculations of Cu surface energy with an included Ni dopant at various locations have only

4.7% average relative error compared to DFT results. Additionally, the difference in minimized

energies between structures with Ni on the exposed surface or beneath it (subsurface) is 0.38-

0.39 eV for <100> and <111> Cu surfaces, yet this gap is only 0.17 eV for doped <111> Cu

surface. The same trend was found to cause the equilibrium segregation profiles from previous

MC studies of CuNi surface separation, agreeing with the calculated energetics in our own

validation step[18, 26, 27, 53, 54]. Overall, the trend indicates that Ni will segregate towards

subsurface sites to relax the local surface tension.

CuNi(100) CuNi(110) CuNi(111)

DFT EAM DFT EAM DFT EAM

Layer #
(J/m2) (J/m2) (J/m2) (J/m2) (J/m2) (J/m2)

1st Layer 1.387 1.371 1.457 1.512 1.165 1.264

2nd Layer 1.360 1.365 1.430 1.505 1.150 1.256

3rd Layer 1.364 1.362 1.429 1.501 1.152 1.257

28
Table A2: Surface energies of Cu with 1x Ni dopant atom included on the surface, subsurface,

or interior (1st, 2nd, or 3rd layer) to illustrate segregation trends from DFT and EAM

The monovacancy formation energies for pure Cu and pure Ni are 1.29 eV and 1.57 eV,

respectively.

Appendix B: Composition Profile Estimation MC Data

Each step of the simulated annealing to equilibrium is dependent on the previous configuration

(serial correlation), so the observed samples for calculating the estimated mean composition have

a biased standard error that is overestimated by a naïve error calculation. To correct this,

autocorrelation of subsequent frames in the MC process was estimated by assuming samples are

generated by a first order autoregressive process (AR(1)):

X, = c + ρX,"# + ε'=<).723

where Xt is the estimated value at time t by the AR(1) model, c is an optional shifting constant

(typically zero), and ε the corrected error we wish to estimate. The process has an analytically

known correction to find the unbiased, standard error of the estimated mean from n

measurements and depends on ρ to calculate the measurement error for each layer[55]:

1 + 2δ⁄n
σ'=<).723 = σ<).723 ;
2δ
1−
n(n − 1)

(n − 1)ρ − nρ% + ρ=>#

δ=
(1 − ρ)%

A custom Perl script was used to analyze the output data from MC, finding the

composition in each layer and tracking it over all the annealed samples. These measurements

were exported to a Python script which estimated values of ρ for the data using the ‘statsmodels’

29
module, and the above equations estimated accurate errors for plots of the equilibrium

composition profiles.

Appendix C: Cluster Expansion Model for Kinetic Monte Carlo (KMC) simulation

E
a0
E
a
Energy, E

∆E

0 Reaction coordinate, r a

Figure C1: Parabolic construction showing energy diagram of a system. The blue curve is the

initial state, the green curve is the final state with the same energy as the blue curve while the red

curve is the final state with lower energy than the blue curve (Recreation)[56].

The reaction coordinate between two states is identified and analyzed using high-temperature

sampling and climbing image nudged elastic band (CI-NEB) in AKMC. Contrastingly, our KMC

model approximates these transition energies using an equation parametrized from barriers

already known from AKMC: in Figure C1, the energy of the red curve is lower than the green

30
curve due to a change in local environment of the vacancy. As a result, the barrier is lower and

the rate can be calculated analytically as [56]

(4E.? + ∆E)%
E. =
16E.?
"9&
rate = Ae 6@

where E. is the barrier from one state to a lower-energy state (blue and red), E.? is the barrier

between two isoenergetic states (blue and green) also known as the intrinsic barrier, ∆E is the

energy difference between the initial and the final state, A is the pre-factor in s-1, k is the

Boltzmann constant 8.617 eV.K-1, and T is temperature in K. The intrinsic barrier E.? is obtained

from AKMC data in Table C1.

Table C1: Intrinsic barrier of a vacancy migration event taken from AKMC data

Location

of the Surface Subsurface

vacancy

Moving
Cu Ni Cu Ni
atom

Intrinsic

barrier 0.453 0.487 0.635 0.681

𝐄𝐚𝟎 (eV)

To acquire ∆E, the energy of each state is needed for the equation

∆E = EC,) − EC,E

31
where EC,) is the total energy of the configuration with the vacancy at location i (final) and EC,) is

the total energy of the configuration with the vacancy at location j (initial). Vacancy formation

energy is calculated by

E8 = EC − E? + µ.

where E8 is the vacancy formation energy, EC is the total energy of the configuration with the

vacancy, E? is the total energy of the configuration without the vacancy (initial configuration)

and µ. is the chemical potential of the atom removed from the initial configuration to make the

vacancy.

The vacancy formation energy can then be used to determine ∆E:

∆E = (E8,) + E? − µ. ) − (E8,E + E? − µ. ) = E8,) − E8,E

Vacancy formation energy is predicted by the cluster expansion method, which samples multiple

alloy structures uniformly at random vacancies with different local environments. The energies

of these structures were calculated and used to parametrize the model shown in Figure C2.

a) b)

32
c) d)

Figure C2: Average and standard deviation values of vacancy formation energy at (a) surface

and (b) subsurface. Cluster expansion fitting model for the vacancy formation energy at (c)

surface and (d) subsurface.

While the cluster expansion model for KMC works for surface-surface and subsurface-

subsurface diffusion, it fails to illustrate correct surface-subsurface migrations (Figure C2b) since

the identity of the migrating atom (counter to the vacancy movement) is not included in this

model: Ni requires higher barriers (~0.95-1.0 eV) than Cu (~0.8-0.9 eV) to move from surface to

subsurface according to AKMC data. To fix this significant qualitative limitation of the cluster

expansion model, a new definition for the barriers of a vacancy migration from surface to

subsurface is proposed based on the AKMC data:

E.,7'(8"7'< = ∆E + E.,7'<"7'(8

The AKMC data shows that the barriers of a vacancy migration from surface to subsurface are

very similar to the energy difference of final and initial states (∆E in the equation above), and

values of , E.,7'<"7'(8 are very small (0.049 eV for moving Cu atom and 0.067 eV for moving Ni

atom) regardless of local environment. Figure C3 depicts an interpretation to calculate the

33
barriers of a vacancy migration from surface to subsurface by adding the difference in energy,

∆E, between two states to E.,7'<"7'(8 .

vacancy at subsurface vacancy at subsurface

moving atom at surface moving atom at surface

Eₐ sub-surf

Eₐ sub-surf

Eₐ surf-sub
∆𝐸

vacancy at surface vacancy at surface

moving atom at subsurface moving atom at subsurface

Figure C3. An illustration to determine the barrier of vacancy migration from surface to

subsurface using the existed barrier of vacancy migration from subsurface to surface.

By the new definition, the barrier difference between Ni and Cu originates from the energy

difference of two states ∆E, and its correction is significant to obtain correct dynamics. To

include this information, ∆E was calculated from EON, where AKMC runs, then compared with

the numerically calculated ∆E from cluster expansion. This discrepancy is illustrated in Figure

C4[50]: the errors are approximately 0.09 eV for Cu and 0.24 eV for Ni, so these correction

values were chosen to improve the cluster expansion model.

34
Figure C4: An illustration of discrepancy between ∆E from EON and from cluster expansion

Appendix D: Comparison of AKMC and KMC results

Figure D1 presents the energy profiles obtained from AKMC and KMC simulations that reach

timescale of 300 µs, realizing 144,597 and 124,650 transition states respectively. Both methods

show two Ni segregation events away from the surface at 100 ~ 125 µs and 170 ~ 200 µs, which

corresponds to the energy drops in Figure. This comparison supports that our KMC model can

simulate the similar behaviors that were observed in AKMC.

Figure D1: Energetic profiles of AKMC (blue) and KMC (red) simulations with respect to the

total time of 300 µs.

35
References:

[1] F. Fischer, G. Schmitz, S.M. Eich, A systematic study of grain boundary segregation and

grain boundary formation energy using a new copper–nickel embedded-atom potential, Acta

Materialia 176 (2019) 220-231.

[2] G. Kresse, J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and

semiconductors using a plane-wave basis set, Computational Materials Science 6(1) (1996) 15-

50.

[3] G. Kresse, J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations

using a plane-wave basis set, Physical Review B 54(16) (1996) 11169-11186.

[4] G. Kresse, D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave

method, Physical Review B 59(3) (1999) 1758-1775.

[5] P.E. Blöchl, Projector augmented-wave method, Physical Review B 50(24) (1994) 17953-

17979.

[6] J.P. Perdew, K. Burke, M. Ernzerhof, Generalized Gradient Approximation Made Simple,

Phys. Rev. Lett. 77(18) (1996) 3865-3868.

[7] M. Hennes, J. Buchwald, U. Ross, A. Lotnyk, S.G. Mayr, Equilibrium segregation patterns

and alloying in Cu/Ni nanoparticles: Experiments versus modeling, Physical Review B 91(24)

(2015) 245401.

[8] B. Good, G. Bozzolo, J. Ferrante, Surface segregation in Cu-Ni alloys, Physical Review B

48(24) (1993) 18284-18287.

[9] S.M. Foiles, Calculation of the surface segregation of Ni-Cu alloys with the use of the

embedded-atom method, Physical Review B 32(12) (1985) 7685-7693.

36
[10] R.G. Donnelly, T.S. King, Surface composition and surface cluster size distribution of Cu-

Ni alloys via a monte carlo method, Surface Science 74(1) (1978) 89-108.

[11] J. Eymery, J.C. Joud, Surface segregation in binary Cu-Ni and Pt-Ni alloys using Monte

Carlo simulation, Surface Science 231(3) (1990) 419-426.

[12] R.H. Shumway, D.S. Stoffer, Time Series Analysis and Its Applications (Springer Texts in

Statistics), Springer-Verlag2005.

[13] P. Xiao, G. Henkelman, Kinetic Monte Carlo Study of Li Intercalation in LiFePO4, ACS

Nano 12(1) (2018) 844-851.

[14] G. Henkelman, H. Jónsson, EON: Long timescale dynamics.

https://theory.cm.utexas.edu/eon/, 2021 (accessed July 23, 2021.).

37