Computational Methods in Heterogeneous Catalysis

Benjamin W. J. Chen†1, Lang Xu†, Manos Mavrikakis*

Department of Chemical and Biological Engineering, University of Wisconsin – Madison,

Madison, WI 53706

†
These authors contributed equally.

*corresponding author [email protected]

Abstract

The unprecedented ability of computations to probe atomic-level details of catalytic systems holds

immense promise for the fundamentals-based bottom-up design of novel heterogeneous catalysts,

which are at the heart of the chemical and energy sectors of industry. Here, we critically analyze

recent advances in computational heterogeneous catalysis. Firstly, we will survey the progress in

electronic structure methods and atomistic catalyst models employed, which have enabled the

catalysis community to build increasingly intricate, realistic, and accurate models of the active

sites of supported transition-metal catalysts. We then review developments in microkinetic

modelling, specifically mean-field microkinetic models and kinetic Monte Carlo simulations,

which bridge the gap between nanoscale computational insights and macroscale experimental

kinetics data with increasing fidelity. We finally review the advancements in theoretical methods

1
Current Address: Institute of High Performance Computing, Agency for Science, Technology and Research, 1
Fusionopolis Way, #16–16 Connexis, Singapore 138632, Singapore

1
for accelerating catalyst design and discovery. Throughout the review, we provide ample examples

of applications, discuss remaining challenges, and provide our outlook for the near future.

2
Biography

Benjamin W. J. Chen is a research scientist at the Institute of High Performance Computing

(IHPC), a member institute of the Agency for Science, Research, and Technology (A*STAR) in
Singapore. He received his Ph.D. from the University of Wisconsin–Madison under the
supervision of Professor Manos Mavrikakis in the Department of Chemical and Biological
Engineering, and his B.Sc. degree in Chemistry from Imperial College London in 2013. His
research focuses on identifying improved materials and catalysts for the storage and conversion of
energy via first-principles calculations and kinetic modeling.

Lang Xu is a postdoctoral research associate at the University of Wisconsin-Madison. He received

his Ph.D. in Chemical Engineering from the University of Wisconsin-Madison in 2018, under the
supervision of Professor Manos Mavrikakis, and his B.Eng. in Chemical Engineering from the
Hong Kong University of Science and Technology in 2012. His research focuses on applying
density functional theory calculations and microkinetic models to study the mechanism of
heterogeneously catalyzed reactions.

Manos Mavrikakis received his PhD in Chemical Engineering & Scientific Computing from the
University of Michigan, Ann Arbor. Following postdoctoral work at the University of Delaware
and the Technical University of Denmark, he joined the faculty at the University of Wisconsin-
Madison. He is currently the Paul A. Elfers Professor of Chemical Engineering. His research
interests include first-principles and microkinetic modeling of heterogeneous catalysis,
electrocatalysis, and chemoresponsive systems. He has received the 2009 Paul H. Emmett Award
in Fundamental Catalysis (North American Catalysis Society), the 2014 R.H. Wilhelm Award in
Chemical Reaction Engineering (AIChE), and the 2019 Gabor A. Somorjai Award for Creative
Research in Catalysis (ACS). He has been editor-in-chief of Surface Science since 2012.

3
Table of Contents

Abstract ........................................................................................................................................... 1 

1.  Introduction ............................................................................................................................. 7 

2.  Brief Theoretical Background for Electronic Structure Calculations ..................................... 9 

2.1.  Time-Independent Schrödinger Equation ........................................................................ 9 

  Hartree-Fock method .............................................................................................. 10 

2.2.  Density Functional Theory ............................................................................................. 11 

3.  Recent Advances towards Better Atomistic Models for Heterogeneous Catalysis .............. 13 

3.1.  Beyond DFT-GGA: Development of More Accurate Computational Methods ............ 13 

  Hybrid Functionals.................................................................................................. 14 

  vdW-Corrected and vdW-Inclusive Functionals .................................................... 18 

Non-Local Density Functionals ........................................................................................ 18

Empirical Corrections. ...................................................................................................... 22

  Random Phase Approximation. .............................................................................. 23 

3.2.  Development of Computational Methods for Large Systems ........................................ 25 

  Linear Scaling DFT Calculations............................................................................ 26 

  Machine-Learning Atomistic Interatomic Potentials. ............................................. 28 

3.3.  Catalyst Models for Large Nanoparticles and Small Nanoclusters................................ 32 

  Periodic Slab Models for Large Nanoparticles. ...................................................... 33 

Alloy Slab Models ............................................................................................................ 34

Strained Slab Models ........................................................................................................ 37

  Free and Supported Cluster Models for Small Nanocluster Catalysts .................... 37 

Modelling Supports ........................................................................................................... 40

4
 Capturing Dynamic Behavior ........................................................................................... 42

3.4.  Theoretical Treatment of Liquid-Phase Reactions ......................................................... 44 

  Implicit (Continuum) Solvent Models .................................................................... 45 

  Explicit Solvent Models .......................................................................................... 47 

  Microsolvation (Cluster-Continuum) Models......................................................... 50 

4.  First-Principles-Based Microkinetic Modeling for Heterogeneous Catalysis ...................... 51 

4.1.  First-Principles-Based Mean-Field Microkinetic Modeling .......................................... 51 

  Calculation of Rate and Equilibrium Constants...................................................... 53 

  Treating Lateral Interactions in Mean-field Microkinetic Models ......................... 55 

  Parameter Estimation for Evaluating Mean-field Microkinetic Models................. 57 

  Self-Consistency between Atomic-Level Energetics and Macroscopic Kinetics ... 60 

  Case Study: Mean-field Microkinetic Modeling of the Water Gas Shift Reaction 64 

  Case Study: Mean-field Microkinetic Modeling of Formic Acid Decomposition . 66 

4.2.  On-Lattice Kinetic Monte Carlo Simulations ................................................................ 69 

  Treating Lateral Interactions in KMC Simulations ................................................ 70 

  KMC Simulations for Clusters ............................................................................... 74 

  Treating Lattice Reconstructions in KMC Simulations .......................................... 76 

5.  Theoretical Methods for Rapid in-silico Catalyst Design and Discovery ............................ 80 

5.1.  Scaling and Brønsted-Evans-Polanyi Relations ............................................................. 80 

  Binding Energy Linear Scaling Relations............................................................... 80 

  Brønsted-Evans-Polanyi (BEP) Relations. ............................................................. 83 

5.2.  Sabatier Principle and Volcano Plots ............................................................................. 84 

  Initial Catalyst Screening using Volcano Plots ....................................................... 84 

5
   Breaking Scaling Relations ..................................................................................... 88 

3D Active Sites ................................................................................................................. 89

Alloying ............................................................................................................................ 89

Bifunctional Catalysts ....................................................................................................... 91

Strain ................................................................................................................................. 92

5.3.  Identification of Reactivity Descriptors ......................................................................... 93 

6.  Conclusions and Outlooks .................................................................................................... 98 

References ................................................................................................................................... 102 

6
1. Introduction

Catalysis—the science of accelerating chemical transformations—lies in the heart of modern

society. The production of fuels, commodity chemicals, pharmaceuticals, as well as consumer

products all rely heavily on catalytic processes. To facilitate the separation of the catalyst from the

product, heterogeneous catalysts, based mainly on metals and metal oxides, are widely used in

industrial processes. Examples of such processes include the large-scale production of commodity

chemicals such as ammonia1 and methanol.2

Due to the large economic and environmental footprint of the chemical industry, there are

enormous financial and societal incentives for improving the efficiency of catalysts. Catalytic

reactions designed to run at lower temperatures and pressures can lead to energy savings, which,

in turn, have positive effects on the environment. While catalyst development has typically been

based on trial-and-error experimental approaches, computational methods are playing an

increasingly important role by empowering guided approaches towards catalyst design and

development. This has been accelerated over the past two decades thanks to the rapidly increasing

computing power. As an example of the predictive capabilities of current computational methods,

novel Pt3Y and Pt3Sc catalysts more active than the state-of-the-art pure-Pt catalyst were designed

for the oxygen reduction reaction (ORR) using a purely computational descriptor-based approach,

which relied on understanding the important aspects of the reaction mechanism.3

Despite these sporadic successes, designing catalysts from computations remains challenging. The

complexity of real catalysts, from the role of the support to the dynamic nature of catalyst particles,

remain real and significant challenges that researchers contend with today. Additionally, the

process of designing improved catalysts from the atomic level up spans multiple length and time

7
scales. There is no single technique that can bridge the large materials and pressure gaps between

the nanoscale insights gained by computations and the macroscale reaction kinetics data collected

in experiments.

Figure 1. An illustration of a bottom-up multiscale computational heterogeneous catalysis study, from the
level of electrons to that of the catalytic reactor. (a) Electronic structure calculations provide predictive-
quality information about individual elementary steps in the catalytic cycle, such as adsorption energies and
activation energy barriers. (b) Kinetic models, such as kinetic Monte Carlo simulations, use this information
to evaluate the interplay between all elementary processes, establishing the intrinsic activity of the catalyst.
(c) Coarse-grained models over the catalyst microstructure simulate the inhomogeneous reaction conditions
at the length scale of a real reactor. (d) Interfacing these coarse-grained models with macroscopic heat and
mass transport models simulates the flow of reactants, products, and energy in the reactor. Adapted with
permission from reference 4. Copyright 2019 Springer Nature.

As a result, multiscale models for designing improved catalysts have been developed. These

models consist of several stages designed to bridge these gaps one at a time (Figure 1). The first

stage involves building an atomic-scale structural model for the catalyst and probing the energetics

of elementary steps in the reaction mechanism, usually with electronic-structure calculations. In

the second stage, the elementary steps are combined and fed into microkinetic models such as

mean-field microkinetic models (MF-MKMs) or kinetic Monte Carlo (KMC) simulations. These

models bridge the temperature and pressure gap between in silico models and the local catalyst

environment. Lastly, to account for inhomogeneities in temperatures and reactant concentrations

in the reactor, the models can be further integrated with heat and mass transport models.

8
Appropriate coarse graining must be applied so that the local environment at each point is

approximately constant.

While the multiscale approach remains the gold standard in modelling heterogeneous catalysis,

theoretical advances increased the fidelity and accuracy of each stage of the multiscale process.

Here, we review these recent advances, focusing on the first two stages of the multiscale modeling

framework—the first-principles characterization of intrinsic catalyst reactivity and their

subsequent use in microkinetic models—given that these factors most significantly impact the

intrinsic activity of catalysts.

The review is structured as follows. First, we briefly present the theoretical background for the

electronic structure methods commonly used in modern catalysis research. Second, we discuss

recent developments in electronic structure methods towards more accurate energetic descriptions

of reactions occurring on heterogeneous catalysts. Third, we discuss how improvements in

microkinetic models allow us to obtain a more in-depth understanding of heterogeneous catalysts

under reaction conditions. Lastly, we summarize the research efforts towards establishing suitable

descriptors for reactivity trends, which are key for enabling fundamentals-based catalyst design

and screening.

2. Brief Theoretical Background for Electronic Structure Calculations

2.1. Time-Independent Schrödinger Equation

Electronic structure methods aim to solve for the stationary states of a quantum mechanical system.

The equation that describes these states is the time-independent Schrödinger equation (TISE,

Equation 1):

9
 Ĥ  E (1)

where 𝐻 is the Hamiltonian operator, E is the total energy, and 𝜓 is the wavefunction which

contains all information about a system such as its ground-state energy and the probability

distributions of the electrons in the system. While analytically solvable for the simplest systems

such as the hydrogen atom, many-body electron-electron interactions hinder the solution of the

3N-dimensional TISE for practically relevant systems with N electrons.

Hartree-Fock method

The Hartree-Fock (HF) method is the simplest approach to tackle this problem. Under the HF

scheme, the solution to the TISE can be expressed as a Slater determinant of the one-electron

wavefunctions 𝜓 , which are the eigenvectors of the one-electron Fock operator (fi) (Equation 2):

f i i  Ei i
h2 2 ˆ ˆ (2)
ˆ ˆ ˆ
f i  T  V  VH ,i     V  VH ,i
2m

where 𝑇 is the operator for the kinetic energy, 𝑉 is the operator for the potential energy due to

electron-nucleus interactions, 𝑉 , is the operator for the Hartree-Fock potential arising from th

electron-electron electrostatic interactions of the ith electron with all other electrons. ℏ is the

reduced Planck’s constant, m is the electronic mass, and Ei is the energy of the one-electron

wavefunction 𝜓 .

The HF solution assumes that an electron moves independently of other electrons, except for the

effect of a mean-field Coulomb repulsion based on the average positions of all other electrons.

This neglect of the electron correlation leads to large errors in the total energy of the system. A

10
range of post-HF methods to better approximate such many-body electron interactions have been

developed. These include the configuration-interaction (CI)5,6 and coupled-cluster (CC)7,8

techniques, which express the wavefunction as combinations of Slater determinants of ground and

excited states, and Møller-Plesset (MP) perturbation theory,9,10 which attempts to recover the

correlation energy by perturbing the Hamiltonian.

HF and post-HF methods are widely used for small molecules. However, their computational cost

increases rapidly with system size. The HF method traditionally scales as O(N4) with respect to

the total number of electrons (N) in the system, whereas post-HF methods such as CI scale as

O(N7)11. For modelling heterogeneous catalysts, where at least tens to hundreds of atoms are

typically required, the computational cost of wavefunction-based methods remains prohibitive.

2.2. Density Functional Theory

An elegant alternative method for solving the TISE, Density Functional Theory (DFT) is based on

the premise that working with the electron density, a 3-dimensional variable, should be much more

efficient than working with the wavefunction, a 3N-dimensional variable. Developed by

Hohenberg, Kohn, and Sham in the 1960s,12,13 DFT is grounded on the theory that the ground-state

energy calculated from the TISE is a unique functional of the electron density. Kohn and Sham

produced a practical implementation of DFT, popularly known as Kohn-Sham DFT (KS-DFT), by

introducing a fictitious non-interacting reference system. In this system, the wavefunction can be

expressed as a single Slater determinant made up of the one-electron orbitals 𝜑 , whose densities

sum up to that of the real system. These orbitals have corresponding energies 𝜀 , which can be

obtained by solving the Kohn-Sham equation (Equation 3):

11
  h2 2 ˆ ˆ 
   V  VH  VˆXC  i   ii (3)
 2m 

The Hamiltonian of the Kohn-Sham equation is similar to the one-electron Hamiltonian in the HF

methods, except for an additional exchange-correlation operator, 𝑉 , which treats the many-body

electron interactions. No explicit analytic expression, however, exists for this term except for the

simple system of a uniform electron gas. An entire field of research is devoted to finding different

approximations for this exchange-correlation term; such approximations are known as exchange-

correlation (XC) functionals.14,15

Generalized gradient approximation (GGA) XC functionals, which use information about the

electron density and its local gradient, are one of the most successful families of functionals in

heterogeneous catalysis modeling. Members of this family include the Perdew-Wang (PW91)16

and the Perdew-Burke-Ernzerhof (PBE)17 functionals. The success of these functionals lie in their

low cost to accuracy ratio: the computational cost of DFT-GGA typically scales as O(N3),11

whereas the simplest post-HF method, second-order MP perturbation theory (MP2), scales as O(N5)

and has an accuracy similar to that of DFT-GGA in some cases.

Yet, the accuracy of DFT-GGA is far from what is needed to accurately predict chemical reactivity.

Chemical accuracy is defined as an error of less than 1 kcal / mol-1 (~0.04 eV). Depending on the

functional used, the estimated error of DFT-GGA in calculating binding energies is 0.2–0.3 eV for

strongly chemisorbed adsorbates and 0.4–0.7 eV for adsorbates with large contributions to

dispersion interactions. These errors were calculated by comparing calculated DFT values with

experimental single-crystal adsorption microcalorimetry (SCAC) measurements, one of the most

accurate methods to determine experimental adsorption energies, and temperature-programmed

12
desorption estimations.18,19 Improvements in electronic structure methods will thus be vital for

computational heterogeneous catalysis research in the future. Examples of some research efforts

to improve the accuracy of DFT beyond standard DFT-GGA is the focus of our next section

(Section 3.1).

3. Recent Advances towards Better Atomistic Models for Heterogeneous Catalysis

In this section, we focus on recent efforts to build better atomistic models for describing

heterogeneous catalysts. We first review how new DFT functionals have enabled the derivation of

increasingly accurate reaction energetics, and how emerging techniques such as machine learning

and linear-scaling DFT methods have also increased the speed of obtaining these energetics. We

then look at the atomistic models themselves, the complexities of which have been

underappreciated until recently. Lastly, we discuss how to treat solvation effects, which will be

vital for electrochemical and catalytic processes taking place in the liquid phase.

3.1. Beyond DFT-GGA: Development of More Accurate Computational Methods

The first step towards an accurate atomistic catalyst model is to select a suitable method for

describing reaction energetics. This is especially challenging since the method of choice must

accurately describe both the surface and the adsorbate, despite them being chemically very

different entities. Electronic structure methods, especially in the framework of DFT, have become

the method of choice for probing the reaction energetics since their development in the 1990s. Yet,

commonly used DFT-GGA functionals, such as PW9116 and PBE,17 suffer from the well-known

self-interaction error (SIE)—the interaction of an electron with its own density—a major source

of error for most KS-DFT functionals.20

13
One important consequence of the SIE is that it leads to overly delocalized electronic structures.

This is especially apparent for transition-metal oxides such as NiO21 and MnO2,22 where the d

electrons are not as localized to the transition metal atom as they should be. Additionally, the SIE

systematically underestimates band gaps for semiconductors and wide-gap insulators23,24 and over-

stabilizes systems with fractional charges such as the transition states of gas-phase reactions, thus

underpredicting the barriers of such reactions.25

Another deficiency of semi-local DFT-GGA functionals is their lack of long-range electron

correlation. As a result, van der Waals (vdW) interactions are poorly captured.26 These interactions

are critical for accurately describing the adsorption of large, aromatic molecules for example.27

In this section, we discuss recent advancements in electronic-structure methods that attempt to

rectify the shortcomings of DFT-GGA and examples of their applications in heterogeneous

catalysis. First, we discuss the use of hybrid XC functionals. Second, we review different methods

to account for vdW interactions. Lastly, we review the use of the random phase approximation

(RPA), which, albeit its computational cost, is one of the most accurate methods available for

modeling reactions for a variety of different catalysts including commonly used transition-metal

catalysts .

Hybrid Functionals

14
Figure 2. Jacob’s Ladder for density functional approximations. Adapted from the original Jacob’s Ladder
proposed by Perdew and Schmidt.28

To improve on the standard GGA functional, researchers proposed to add the non-interacting

kinetic energy density term as an input parameter. These meta-GGA functionals supposedly

improve the accuracy of the XC approximation.29 Yet, SIE still exists in the meta-GGA functionals.

Hybrid functionals, which incorporate a fraction of the exact exchange evaluated using HF/post-

HF methods to partially mitigate the SIE, were thus developed. The inclusion of the exact exchange

allows them to describe electronic exchange and correlation effects better than the GGA/meta-

GGA functionals. Hybrid functionals thus rank the second highest on Jacob’s Ladder, a well-

known hierarchy proposed by Perdew and Schmidt in 2003 to classify the accuracy of various DFT

methods (Figure 2).28 A variety of hybrid functionals capable of treating metallic systems have

been developed, most notably the PBE030 and Heyd–Scuseria–Ernzerhof (HSE0331,32 and

HSE0633) functionals. Note that the Becke, 3-parameter, Lee–Yang–Parr (B3LYP)34,35 functional,

although popular for modeling homogeneous catalysts, is rarely used to model transition-metal

heterogeneous catalysts. This is because it describes metallic systems poorly as a result of the LYP

functional underestimating correlation effects.36

15
Hybrid functionals are much more computationally expensive than GGA functionals, particularly

for systems with periodic boundary conditions: the evaluation of the exact exchange is much less

efficient using plane waves, where their delocalized nature means that all pairs of orbitals

contribute to the exact exchange,37 as compared with using a localized Gaussian basis set. As

demonstrated in a review by Xu and Carter, hybrid functionals are generally more accurate than

DFT+U, a commonly used semi-empirical method to treat the Coulomb interactions between

strongly correlated localized electrons with an additional Hubbard-like term.38 Yet, the cost of

hybrid functionals limits the system sizes that can be studied with them.38

A general strategy to lower the cost of hybrid DFT calculations is to first optimize structures

perform energy minimization using GGA functionals, and then perform single-point calculations

using hybrid functionals to obtain the energy. Gouveia and Coutinho showed that HSE06 total

energies for PBE-optimized structures have errors of < 1 meV / atom compared with fully self-

consistent HSE06 calculations for an oxygen vacancy in bulk MgO and an interstitial oxygen in

bulk Si.39 The errors in energy differences are even smaller due to error cancellation. This approach

was adopted by Zhao et al. to study the reaction mechanism for the water-gas shift reaction (WGSR)

on a Au nanowire supported by MgO(100).40 They observed that the use of HSE06 alleviates the

commonly-known issue of CO over-binding by GGA-PW91.41

Mazheika and Levchenko demonstrated that HSE accurately describes the adsorption of CO and

CO2 on Ni single atom catalysts (SACs) supported on MgO (NiSAC/MgO) upon proper

parameterization of the fraction of the exact exchange, α.42 The adsorption energies calculated by

HSE with an α of 0.2 – 0.3 compared well with those predicted by the wave-function-based

16
coupled cluster singles and doubles with perturbative triples (CCSD(T)) method, the gold standard

in quantum chemistry (Figure 3).42

Figure 3. Calculated adsorption energies of (a) CO and (b) CO2 on NiSAC/MgO using HSE, MP2, and
CCSD(T) methods. HSE values are plotted as a function of the fraction of exact exchange in the hybrid
functional, α. Reprinted with permission from reference 42. Copyright 2016 American Chemical Society.

Hybrid functionals have also found applications in photo- and electrochemical catalysis, since

accurate band structures are crucial for these systems.38 This is especially true for systems with

defects and impurities, which contribute electronic states within the band gap. Mino et al. studied

CO2 capture on anatase TiO2 using the PBE0 functional43 which offered a more accurate

description for the structural and electronic properties of the TiO2 surface. Gao et al. probed the

reduction of CO2 on Pd/Pt SACs supported on carbon nitride using HSE06; they predicted band

gap values for the carbon nitride support in good agreement with the experimental values.44

Another field where hybrid functionals have proven to be useful is in zeolite catalysis. This is

partially because the unit cells required to model simple zeolite structures are small with a minimal

number of atoms, and therefore tractable with expensive methods such as hybrid functionals. For

example, Paolucci et al. studied the selective catalytic reduction of NOx on Cu-exchanged SSZ-13

zeolites.45 They observed that HSE06 predicted weaker binding of NO on Cu(II) sites compared

with GGA, consistent with their spectroscopic observation that NO was absent on Cu(II) sites.

17
Additionally, Göltl and Hafner performed an extensive benchmarking study for CO and NO

adsorption on Cu- and Co-exchanged zeolites using different functionals; they demonstrated that

hybrid functionals (PBE0 and HSE) predicted more accurate binding structures and energies than

GGA.46–48

vdW-Corrected and vdW-Inclusive Functionals

The long-range electronic correlations that give rise to vdW interactions are not built into standard,

semi-local DFT-GGA functionals. This leads to underestimation of the binding strength of large,

closed-shell adsorbates that interact with catalytic surfaces mostly via vdW interactions, such as

aromatics and long-chain hydrocarbons.49 Two main approaches have been developed to

incorporate the effects of vdW interactions in DFT calculations: (1) DFT functionals with non-

local correlation effects, and (2) empirical vdW corrections. We now go through these two

approaches in more detail.

Non-Local Density Functionals

Among the many non-local density functionals, the Rutgers-Chalmers van der Waals Density

Functional (vdW-DF)50 is perhaps one of the most successful.50–52 Since its introduction in 2004,

the vdW-DF2,53 vdw-DF3,54 and the opt family of functionals: optPBE-vdW, optB88-vdW, and

optB86b-vdW,55 have been developed. These functionals incorporate a non-local correlation

component to the XC functional to treat long-to-medium ranged interactions. For additional

technical details on these methods, the readers may refer to a thorough review by Berland et al.26

A specific functional in the family of vdW-DF should be mentioned here, which is the Bayesian

error estimation functional with van der Waals (BEEF-vdW) XC functional developed by

18
Wellendorff et al.56 BEEF-vdW is a semiempirical functional parameterized to experimental data.

Regularization and cross-validation techniques were used to avoid overfitting. vdW interactions

were accounted for by the vdW-DF2 scheme. The BEEF-vdW functional has been used to study

various heterogeneous catalytic reactions such as CO2 electroreduction on MoS2 and MoSe2,57

methanol synthesis on Cu(211),58 and SCR of NOx on Cu-exchanged chabazites.59 Studt et al.

studied the methanol synthesis reaction on Cu(211) from the hydrogenation of a CO/CO2 mixture

using BEEF-vdW, showing that the BEEF-vdW functional is essential for correctly predicting

whether CO or CO2 is selectively hydrogenated.58

Figure 4. Potential energy curve for graphene adsorption on Ni(111) calculated using different functionals.
Random phase approximation (RPA) data are from references 60 (RPA1) and 61 (RPA2). The standard
deviation predicted by BEEF-vdW is indicated by the grey area. ΔE denotes the binding energy of graphene
on Ni(111) per C atom. d denotes the vertical distance between the graphene sheet and the Ni(111) surface.
Reprinted with permission from reference 56. Copyright 2012 American Physical Society.

To quantify the accuracy of BEEF-vdW, Wellendorff et al. mapped out and compared the potential

energy curves for graphene adsorption on Ni(111) using a variety of XC functionals. Then, they

19
compared them with highly accurate RPA- (discussed in detail in Section 3.1.3) predicted curves

(Figure 4).56 The RPA calculations (pink and purple lines; Figure 4) predict two local minima on

the potential energy curves. The vdW family of functionals, including BEEF-vdW (blue line in

Figure 4) were the only functionals able to qualitatively reproduce this behavior. However, BEEF-

vdW shows relatively large differences in absolute energy compared with RPA, indicating that it

is still unable to describe the graphene/Ni(111) system well.

In another benchmark of the accuracy of BEEF-vdW, Wellendorff et al.19 curated a database of

reaction energies obtained from highly accurate SCAC experiments and compared them with the

reaction energies predicted by the semi-local functionals LDA, PBEsol, PW91, PBE, RPBE and

BEEF-vdW.19 For reactions of strongly chemisorbed species where vdW interactions are minimal,

BEEF-vdW was the most accurate with a RMSE of 38.8 kJ mol-1. For reactions of weakly bound

adsorbates with significant vdW interactions, BEEF-vdW had the second-best performance with a

RMSE of 36.0 kJ mol-1 (PBEsol was most accurate with a RMSE of 29.9 kJ mol-1).19

A unique feature of the BEEF-vdW functional is its ability to estimate the uncertainty of its

predicted energies by calculating the energy of a system using an ensemble of different functionals

based around the optimal parameters. For example, these uncertainties are shown as error bars in

the potential energy curves of graphene adsorption on Ni(111) (grey area; Figure 4). The error bars

are small where BEEF-vdW reproduces the potential energy curve well (large distance between

graphene and the surface, d) and large where it does not (small d), showing that the error

estimations provided by BEEF-vdW may be useful for alerting researchers to how sensitive a

system is to the choice of XC functionals.

20
These error estimates can be further translated into uncertainties for reaction rates through error

propagation, as exemplified in a study by Medford et al., who generated a volcano plot (discussed

in detail in Section 5.2) together with the reliability ranges of the theoretical predictions for the

rates of ammonia synthesis on a series of transition-metal surfaces (Figure 5A-C).62 The DFT error

bars obtained from BEEF allow not only for the prediction of the volcano peak, but also the

probability of finding the optimal catalyst at a certain descriptor value (Figure 5D).

Figure 5. Ammonia synthesis rate trends. (A) Absolute turnover frequencies (TOFs) on metallic (211) step
sites of selected metals plotted as a function of N* adsorption energy. (B) TOFs divided by the TOF of Fe
as a function of N* adsorption energy relative to Fe. (C) Rate activity map obtained from linear scaling
relations between the adsorption energy of nitrogen and other intermediates; red lines and red areas
correspond to the uncertainty on the absolute (left y-axis) and relative rate (right y-axis), respectively. A
linear regression of the dissociative adsorption energy of N2 (2EN) versus the N–N transition-state energy
(ETS(N-N)) is shown in the inset, along with principal component ellipses of the metal ensembles. (D)
Probability that a rate volcano has an optimum at a given nitrogen adsorption energy. All plots were

21
generated at 673 K, 100 bar, and 10% equilibrium conversion. Reprinted with permission from reference
62. Copyright 2014 The American Association for the Advancement of Science.

Empirical Corrections

The second main approach for incorporating vdW interactions is to use empirical corrections. A

series of such methods have been developed, including the D2,63 D3,64,65 and D466 methods

developed by Grimme; the Tkatchenko-Scheffler (TS) method;67 and the density-dependent

energy correction (DDsC) method.68,69 These empirical corrections methods have the same aim of

capturing the local dynamic polarizability of individual atoms; this approach comes naturally from

perturbation theory.70

The D2 method calculates the dispersion term by summing up all the pairwise interaction terms

(calculated using empirical coefficients) inversely proportional to the sixth order of the interatomic

distance.63 The D3 method improves upon DFT-D2 by including an additional pairwise term

inversely proportional to the eighth order of the interatomic distance.64,65 The D4 method adds in

a dependence on the partial atomic charges when computing the dipole polarizabilities, increasing

its accuracy over the D3 method.66 The TS and DDsC methods both construct the dispersion

correction term in a similar fashion as the D2 method, except that the coefficients are charge-

density-dependent. Interested readers are directed to several comprehensive reviews for the

theoretical details of these empirical vdW correction methods.70–72

Compared with the vdW family of functionals, empirical vdW corrections are more commonly

used in heterogeneous catalysis to account for vdW interactions in DFT studies. Particularly, the

D2 and D3 methods are some of the most widely used methods for vdW corrections; the original

paper by Grimme63 in 2006 on the D2 method has received approximately 14,000 citations

22
according to the Web of Science. This is mainly because they are easy to implement—no change

to the XC functional itself is required—and also computationally inexpensive to use.72

Some notable applications of empirical vdW corrections are given below. Deng et al. applied

PBE+D2 in combination with experiments to study the electrocatalytic activity of Pt SACs

supported on MoS2 for the hydrogen evolution reaction (HER).73 Gao et al. studied the

photocatalytic reduction mechanism of CO2 on Pd/Pt SACs supported on graphitic carbon nitride

using PBE+D3.44 Kerber et al. calculated the binding energies of various hydrocarbons on silica

using PBE and PBE+D2 and compared them with experimental measurements.74 They observed

that PBE only accounts for a small fraction of the adsorbate-surface interactions, while PBE+D2

gives much better agreement with experiments.74

Random Phase Approximation.

At the top rung of Jacob’s ladder lies the random phase approximation (RPA). Combining the RPA

correlation energy with exact exchange (EX+cRPA) accurately treats both exchange and

correlation effects, especially when combined with single excitations.75 The RPA method has

several significant advantages over hybrid functionals. Firstly, the RPA for the correlation energy

seamlessly incorporates vdW interactions since it is fully non-local.76,77 Secondly, it accounts for

dynamic electron screening, so it can be applied to metallic systems. Thirdly, the SIE in the Hartree

energy is entirely canceled out by using exact exchange.

23
Figure 6. Predicted binding energies of CO on top sites of Pt(111) (triangles) and Rh(111) (squares) versus
the predicted surface energies for various semi-local density functionals, the hybrid HSE functional based
on PBE, and the random phase approximation (RPA). Experimental values are provided for comparison.
Surface areas are per unit area (u.a.). Adapted with permission from reference 78. Copyright 2010 Springer
Nature.

To accurately model heterogeneous catalysts, it is important to describe the electronic properties

of both the surface and the adsorbate correctly. This is, however, a difficult task for semi-local

functionals, which overly stabilize metal surfaces and overestimate binding strength at the same

time.79,80 While there is the expected inverse correlation between surface stability and adsorbate

binding strength within the family of semi-local functionals (Figure 6; open symbols), this

correlation lies on a line that has a constant offset from the experimental values (Figure 6; orange

filled symbols). This suggests a fundamental flaw with semi-local functionals. RPA values,

however, are distinctly offset from this line and are much closer to the experimental values (Figure

6; green filled symbols)—they accurately describe both the metal and the adsorbate. Indeed, RPA

has been used to study adsorption of a variety of molecules, such as CO78,81 and aromatic

molecules,82,83 on late transition metal surfaces with excellent agreement with experimental

geometries, site preferences, and binding energies.

24
Importantly, EX+cRPA solves the “CO puzzle”—the failure of GGA functionals to predict the

experimental top binding site for adsorption of CO on Pt(111).84 EX+cRPA predicts binding of

CO on the top sites of Pt(111) within the experimental binding energy error bars of ~0.2 eV,

whereas semi-local functionals predict hollow site adsorption and overbinding of CO by up to 1

eV.78 This is because semi-local functionals incorrectly describe the energies of the frontier

molecular orbitals of CO with respect to the Fermi energy of the metal surface.85,86

The unique ability of RPA to accurately describe vdW interactions and adsorption makes it a

promising technique for modeling heterogeneous catalysts, particularly transition-metal catalysts.

However, the computational costs of evaluating entire reaction pathways using RPA remain

prohibitive especially for periodic systems. Encouragingly, work on reducing the computational

costs of RPA calculations, such as cubic scaling algorithms for the RPA,87,88 has gradually enabled

calculations on larger systems. For example, calculating the energy of a unit cell with 64 Si atoms

sampled with 3 × 3 × 3 k points required just 5 hours on 128 cores.87 The active developments in

this area will be helpful for accelerating its adoption in the heterogeneous catalysis community.

3.2. Development of Computational Methods for Large Systems

Whereas some catalytic systems can be well-reproduced by simple cluster or slab models, others

such as supported nanoclusters can be extremely complex, requiring larger and more detailed

models to realistically simulate them. Dynamic phenomena may also play significant roles in

certain systems, such as in liquid-phase reactions, making it necessary to construct a time-

dependent description of the system beyond static energetics. In this section, we discuss some

promising methods to speed up DFT calculations and which enable researchers to address the

complexity of large systems.

25
 Linear Scaling DFT Calculations

Calculating the electronic eigenstates and eigenvalues of a system in KS-DFT involves the costly

diagonalization of a dense Hamiltonian matrix. This typically leads to an O(N3) scaling of the

computational cost, which limits the size of typical models of heterogeneous catalysis to 100-200

atoms before the computational cost quickly becomes intractable. However, the development of

linear scaling DFT methods potentially opens avenues for exploring much larger systems of up to

several thousands of atoms. While currently not widely employed for catalytic systems, this is an

opportune moment to briefly showcase such methods as their use will only increase with time. For

more technical and detailed discussions of such methods, the interested reader is referred to the

excellent reviews on this topic by Bowler and Miyazaki,89 and Goedecker.90

The key concept required for achieving linear scaling is the “nearsightedness of electronic matter”.

First introduced by Walter Kohn in 1996,91 this is the observation that electrons do not

communicate across large length scales. We illustrate one way to exploit this nearsightedness via

the density matrix reformulation of KS-DFT. From the occupied one-electron orbitals, 𝜑 , the

density matrix 𝜌 at zero temperature can be expressed as (Equation 4):

  r , r '   i  r  i*  r ' (4)

i

Due to the locality of quantum mechanical systems, the density matrix decays with distance. For

insulators and semiconductors, the decay is exponential allowing the elements of the density

matrix to be zeroed after a certain cutoff distance.92,93 The result is a sparse density matrix, whose

information increases linearly with system size after a certain point. Various codes such as

ONETEP,94 CONQUEST,95 BIGDFT,96 and SIESTA,97 have implemented linear scaling methods

26
for such insulating systems. A comprehensive overview of such codes is provided by Soler et al.
98
Such linear scaling methods have been successfully used to calculate the electronic structures of

AlGaAs quantum dots on a GaAs substrate, a system containing up to 12168 atoms, fully self-

consistently.99

For metallic systems, however, the decay rate of the density matrix is algebraic at 0 K. Although

exponential decay is recovered at finite temperature,100 it remains unclear if the exponent of decay

is large enough to ensure a sparse enough density matrix for practical applications.101 For Al, the

decay was found to be slow such that linear scaling algorithms only reach chemical accuracy at

truncation distances of 48-64 Bohr, which corresponds roughly to system sizes of thousands of

atoms. Therefore, methods that are pseudo-linear-scaling have been developed for metals. These

methods minimize the number of diagonalizations such that the computational cost is dominated

by linear-scaling operations at small system sizes. One example is the ensemble DFT method

implemented in ONETEP102, which has been used successfully to study free-standing Pt

nanoparticles up to 309 atoms103, as well as graphene-supported Pt nanoparticles with a total

system size of 1189 atoms104. The ability to model such large systems opens the door towards

addressing the effect of nanoparticle size on catalytic activity, an important question in

heterogeneous catalysis.

Considering the fundamental issues of the density matrix formalism, orbital-free density functional

theory (OF-DFT) may be a more promising linear scaling method for metals105. By eschewing the

one-electron wavefunctions and turning back purely to the electron density, OF-DFT theoretically

scales as O(NlogN). However, one major challenge is the lack of accurate kinetic energy

functionals. Although recently developed non-local kinetic energy density functionals based on

27
the properties of the free electron gas have begun to treat main group metals quite accurately

compared to KS-DFT106, they have not reached suitable levels of accuracy for treating the localized

electrons of transition metals107,108.

Finally, we note that although the scaling behavior of the cost of linear-scaling DFT is less steep

than that of KS-DFT, it possesses a larger multiplying prefactor (i.e., a larger coefficient).

Therefore, linear-scaling DFT currently remains slower than KS-DFT for small, typically studied,

system sizes of 100s of atoms. However, there is a crossover point at which linear-scaling DFT

becomes faster than KS-DFT, which is currently around 100 – 1000 atoms. We anticipate the rise

of the importance of linear-scaling algorithms as computational power increases, when it becomes

easier to perform calculations at and exceeding these crossover points90.

Machine-Learning Atomistic Interatomic Potentials.

Machine-learning interatomic potentials (ML-IAPs) are another avenue for pursuing faster-than-

DFT calculations. ML-IAPs are potentials fitted to a consistent set of first-principles data using

machine learning methods. In contrast to physics-based potentials, such as the embedded atom

method and Lennard-Jones potentials, these mathematical potentials do not have any knowledge

of the underlying physics of the systems they are trained on. However, due to their more flexible

form, they have an excellent speed-versus-accuracy ratio, possessing near-DFT accuracy while

being much faster than DFT calculations.109–111 In exchange for a small loss in accuracy, ML-IAPs

expand our ability to probe more reaction pathways, access longer timescales to explore new

catalytic phenomena, and increase the realism of our models by adding details such as supports

and co-adsorbates.

28
Figure 7. Construction of machine-learning interatomic potentials (ML-IAPs). A reference database that
samples the 3N-dimensional, Cartesian-coordinates based, potential energy surface (PES) of a material (N
= number of atoms) is created using first-principles methods such as DFT. The structure of the material is
then represented in terms of a machine-learnable form with the use of suitable descriptors, transforming the
PES from Cartesian-coordinate space to descriptor space. The ML-IAP is fitted by regression on the PES
in descriptor space. Reprinted with permission from reference 112. Copyright 2019 Wiley.

An illustration of the process of training an ML-IAP is shown in Figure 7. First, a database

containing the energies and atomic forces of reference structures is constructed. Then, the

structures are transformed from their original Cartesian coordinate representations into suitable

descriptors for training. Ideal descriptors should be invariant to transformations such as translation

and rotations to preserve as many symmetries as possible. Finally, the transformed potential energy

surface can be learned by applying regression techniques, such as neural networks, Gaussian

process regression, or even simple linear regression, to obtain the ML-IAP.

One of the earliest ML-IAPs was developed by Behler and Parrinello, who trained neural networks

(BPNNs) to represent high-dimensional potential energy surfaces for bulk Si.113 A crucial

breakthrough was their introduction of atom-centered symmetry functions, which efficiently

transform the structures into suitable descriptors for training. Their neural network was 5 orders

of magnitude faster than DFT, and achieved high accuracy with energetic root mean square errors

(RMSEs) of 5-6 meV per atom and force RMSEs of 0.2 eV/Å.113 Natarajan and Behler used these

29
neural networks to study the solvation of Cu surfaces by water.114 The good agreement of the

simulated water diffusion coefficient, equilibrium density, and hydrogen bond lifetime with

experimental values demonstrated the ability of the neural network to accurately describe solvation

environments.114

Various ML-IAPs have since been developed, such as moment tensor potentials,115 the Gaussian

Approximation Potential (GAPs),116 and the Spectral Neighbor Analysis Potential (SNAP).117

Although mainly used in materials science communities, there is increasing interest in such ML-

IAPs for heterogeneous catalysis.118,119 The Atomistic Machine-learning Potential (AMP),

developed by Khorshidi and Peterson,120 is one such potential developed for catalysis applications.

Ulissi et al. subsequently employed AMP to rapidly identify the most active Ni-Ga motifs for CO2

reduction from over 500 candidate structures of Ni-Ga bimetallic alloy catalysts.121

Curating a suitable training set is one of the most challenging parts of training a ML-IAPs. To

tackle this challenge, active learning techniques122,123 have been developed to facilitate the training

and use of ML-IAPs. In active learning, atomic configurations that are poorly described by the

potential are automatically identified on-the-fly while using the potential. These configurations are

added into the training database and the potential is retrained to improve its ability to describe

these configurations. By doing so, overfitting of the potential is reduced and the potential never

has to extrapolate beyond the available data, ensuring that it accurately describes the system at all

times.

VASP 6124,125, the latest version of the widely used Vienna ab initio Simulation Package, for

example, incorporates active learning of GAPs. Jinnouchi and coworkers showed that it can

accelerate MD simulations for bulk materials by more than two orders of magnitude with minimal

30
loss of accuracy.126,127 The accuracy of their simulations was proven by the ability to quantitatively

reproduce the first-principles melting points of Al, Si, Ge, Sn, and MgO.126 In a separate work,

they also showcased the use of ML-IAPs to access time scales in the nanosecond regime, which

allowed them to explore new phase transition phenomena in hybrid perovskites.127

Figure 8 Optimization of a single particle on the Müller-Brown potential using the BFGS and ML-Min
algorithms. (Top) The BFGS algorithm optimizes in the true potential, where each geometry step requires
an expensive functional evaluation. Each frame is a snapshot of the same optimization process. (Bottom)
The ML-Min algorithm optimizes in the predicted ML-potential, where the geometry steps are inexpensive.
Each frame represents one iteration of the ML-Min algorithm. Although the optimized structures from the
ML-Min algorithm in the first few iterations (first three frames) are not the true minima, they can be used
to refine the PES iteratively, so that the estimated PES eventually closely resembles the true PES (last
frame). Only a small number of evaluated structures are required to reproduce the true PES. Thus, this
active-learning procedure leads to savings in the number of total function evaluations required to converge
to the minimum. Reprinted from reference 119.

Active learning can also be used to accelerate geometry119 and transition-state optimizations.128

The ML-Min and ML-NEB algorithms found in the open-source CatLearn python package119 use

active-learning techniques in combination with Gaussian process regression to build up

increasingly accurate approximations to the potential energy surface at each iteration of the

optimization. An illustration of how the ML-Min algorithm for geometry optimization works is

shown in Figure 8. Since the optimization is always carried out in the approximate PES (DFT

calculations are only required to verify the final optimized structure) the overall procedure greatly

31
reduces the number of DFT function calls required. The ML-Min algorithm required three times

fewer function evaluations than the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm for

geometry optimizations of CO over Au(111).119 Similarly, the ML-NEB algorithm reduced the

number of function evaluations by one order of magnitude compared to the MDMin algorithm for

NEB optimizations of diffusion of metal adatoms on metal surfaces.128 Importantly, the energies

of the optimized intermediates and transition states obtained from ML-Min and ML-NEB are

identical to that of DFT since the final evaluation is always done on the true PES.

ML-IAPs also have many advantages over empirical force-fields. Most importantly, they are able

to accurately describe chemical reactions, unlike typical non-reactive force-fields. Even reactive

potentials such as ReaxFF,129,130 which were specially designed to accurately treat bond breaking

and formation, have limited transferability109,131 and need to be reparametrized for each new

system.132 Boes et al. compared ReaxFF and BPNNs for a variety of Au structures including

surfaces, bulk, and clusters.109 They showed that BPNNs were more accurate and could be trained

to describe all types of structures well, whereas ReaxFF described clusters poorly.109 The high

transferability and accuracy of ML-IAPs makes them promising alternatives to empirical force-

fields.

3.3. Catalyst Models for Large Nanoparticles and Small Nanoclusters

Besides a suitable electronic structure method, it is also important to determine how best to

approximate the catalyst particles and supports so that the active site can be properly captured. In

this section we discuss the commonly used atomistic models for large (> 5 nm) particles, as well

as recent advances in modeling small, sub-nanometer clusters are of increasing interest in

heterogeneous catalysis.

32
 Periodic Slab Models for Large Nanoparticles.

For large metal nanoparticles > 5 nm, periodic slab models are excellent approximations (Figure

9). They properly capture the delocalized electronic structure of metals and the flat local curvature

of large nanoparticles. Such slab models have thus become the mainstay for modeling

heterogeneous catalysts. Constructing the slab model, however, is an often-overlooked problem in

computational modeling. While slab models for metal surfaces can generally be easily generated,

inorganic crystalline materials such as metal oxides can be challenging since they may have

multiple terminations. To tackle this problem, Sun and Ceder developed an automated approach

based on covariant transformations,133 whereas Witman et al. used a graph theory approach.134

Another challenging area is the generation of moiré structures, relevant to the intense research into

2D materials.135 Examples of catalytic systems where such structures can be encountered include

the growth of graphene136 and ultrathin metal oxide films137,138 on transition metal surfaces.

Fortunately, the program CellMatch can facilitate the construction of moiré structures by

automatically determining possible common supercells given two unit cells.139

33
Figure 9. Atomistic illustration of an idealized 309-atom fcc octahedral nanoparticle. The particle exposes
the (111) (blue) and (100) (red) facets as well as the edge atoms which can be modeled by a (211) slab.
Grey spheres are metal atoms. Purple dashed lines denote the periodic boundary of a typical simulation cell
for a (211) slab.

One limitation of the periodic slab approach is the prohibitively high cost required to model

systems of low symmetry such as surfaces with defects. In such cases, it can be advantageous to

use embedded cluster approaches, such as quantum mechanics/molecular mechanics (QM/MM)

models.140,141 In these models, only the portion of interest of the system—the defect, for example—

is accurately modeled with first-principles, whereas the rest of the system is treated as point

charges or with classical atomistic potentials. This greatly reduces the computational cost while

preserving an accurate description of the site of interest.

Alloy Slab Models

Metal alloys catalysts deserve special mention due to their prevalence.142–144 Some alloy catalysts,

such as near-surface alloys (NSAs)145,146 and single atom alloys (SAAs),147 are straightforward to

model. Many times, however, the structure of alloy surface is often unclear, especially with the

34
lack of in situ experimental characterization. An initial model for an alloy surface is typically

generated by cleaving the most stable bulk alloy structure at the desired surface index. Yet, this

might not yield the surface with the lowest surface energy due to the varying tendency of metal

elements to segregate preferentially to the surface, even in vacuum.148,149 Cao, Niu, and Mueller

showed for PtNi surfaces that the compositions of first three surface layers differ from that of the

bulk due to these surface effects.150 Finding the most stable surface is, however, a difficult global

optimization problem, requiring methods such as simulated annealing. Since first-principles

calculations are too expensive for conducting such global searches, a common method used is to

parameterize a cluster expansion (CE) Hamiltonian from first-principles calculations instead.151–

153

Another challenge in alloy catalysis lies in modeling random alloys which have no fixed atomic

arrangements or unit cell structures. Although it is easy to create a slab with random atomic

configurations, a truly random distribution cannot be achieved with reasonable unit cell sizes

because of periodic boundary conditions. Sampling of the possible atomistic configurations of the

alloy surface with multiple surface models is one way to treat these systems.154,155

35
Figure 10. Predicting whether hydrogen-induced segregation is favorable for various near surface alloys
(NSAs). Alloys are denoted as solute/host pairs. On the x-axis, Eseg is the segregation energy—the energy
required for a solute atom to migrate from the bulk to the surface layer of the host metal. On the y-axis, we
have the difference between the hydrogen binding energies (at a coverage of θH = 1/4 monolayer) on the
close-packed surfaces of the pure solute (|BEHsol|) and the pure host (|BEHhost|). Hatched regions are where
hydrogen-induced segregation is expected. Alloys are subsurface alloys (see inset), unless ‘*’ is present,
which denotes overlayer alloys. Reprinted with permission from reference 145. Copyright 2004 Springer
Nature.

An important factor to consider with all alloy models is the possibility of adsorbate-induced

surface segregation.156–158 If an adsorbate binds stronger to certain elements, this may provide a

driving force for the more strongly binding elements to segregate to the surface.159 For example,

Greeley and Mavrikakis established a simple scheme for predicting hydrogen-induced segregation

behavior in bimetallic NSAs (Figure 10). If the difference in hydrogen binding energy on the two

component metals is larger than the intrinsic segregation energy, then hydrogen-induced

segregation is likely.145

36
To complicate matters, the coverage of adsorbates may also change during a reaction, resulting in

a dynamic situation whereby the degree of adsorbate-induced surface segregation changes with

time.160 Fluctuations in the reaction temperature and partial pressures of reactants may also induce

varying degrees of adsorbate-induced surface segregation.159 In these cases, the surface should be

treated more dynamically with kinetic Monte Carlo simulations,161 for example, which are

discussed in Section 4.2.

Because predicting the structure of an alloy surface is difficult, idealized surfaces, such as near

surface alloys (NSAs) with metal overlayers,145 are frequently used, especially in high-throughput

screening calculations162–164 that focus on studying reactivity trends. They may also serve as an

initial starting point for subsequent screening using other criteria such as the stability of the surface

with respect to segregation and dissolution.165

Strained Slab Models

The effects of tensile or compressive strain can be studied with slab models by artificially

increasing or decreasing the unit cell size in one or more directions. Such strained models are well-

suited for studying pseudomorphic overlayers166 and mechanically strained catalysts.167 The effect

of strain is well-known to alter surface reactivity of metal168–170 and metal-oxide170 surfaces

towards adsorbates, potentially making them more reactive171. Strain also could be a useful tool

for breaking scaling relations;166 we elaborate more about the effects of strain for breaking scaling

relations in Section 5.2.

Free and Supported Cluster Models for Small Nanocluster Catalysts

37
Figure 11. Differences in coordination environments between cluster and slab models. (a) An 18-atom
cluster with many different coordination environments. The coordination numbers (CN) of the atoms range
from CN=3 to CN=9. (b, c) Slab models, such as the (b) fcc(111) and (c) fcc(100) surfaces, have uniform
environments on their surface (CN=9 and CN=8, respectively). Spheres denote metal atoms. The dotted
lines mark out 2×2 unit cells.

While synthesized catalyst nanoparticles are traditionally large (> 5 nm), interest in small

nanoclusters (< 2 nm) has steadily increased since Haruta’s discovery of the extraordinary catalytic

properties of well-dispersed Au more than 30 years ago.172–174 Such clusters are suspected to play

a role even in homogeneous catalysis: Oliver-Meseguer et al. showed that 3- to 10-atom Au

clusters form spontaneously from solutions of gold salts, acting as the active sites for organic

reactions such as hydration of alkynes.175 The fervent research into heterogeneous cluster

catalysis—accelerated by the improved experimental capabilities for synthesizing and depositing

size-selected clusters with precise numbers of atoms—is evidenced by the multiple excellent

reviews on the topic in recent years.176–180

Slab models are unrealistic for modelling small sub-nanometer clusters as they do not reproduce

their discrete energy levels caused by quantum effects181,182 and their high proportion of

undercoordinated sites183–185 (Figure 11). At these small sizes, surface-tension induced strain186

38
and interactions with the support187,188 also significantly alter the electronic properties of the cluster.

Due to these unique properties, a cluster’s reactivity is highly sensitive to its size down to single

atoms—every atom makes a difference.189 Superatomic metal clusters are a classic example of

these reactivity differences.190,191 Because they have atom-like electronic structures,192 their

reactivity oscillates depending on whether they have an even number of atoms (closed-shell; less

reactive) or an odd number of atoms (open-shell; more reactive).193

Figure 12. CO oxidation over different-sized Pd clusters supported on rutile TiO2(110). The left axis
indicates the CO oxidation activity, obtained by temperature-programmed reaction. The right axis indicates
the deviation of the Pd 3d binding energy from the expected smooth bulk scaling (n-S) values, where n is
the number of cluster atoms and S is a fitted scaling factor. The CO oxidation activity correlates with the
deviation of the Pd 3d binding energy as determined by x-ray photoelectron spectroscopy (XPS). Positive
(negative) deviations lead to decreased (enhanced) CO oxidation activity. Isolated Pd atoms are not active;
CO oxidation activity of a single Pd atom is similar to the activity of the support (cluster size of 0).
Reprinted with permission from reference 194. Copyright 2009 The American Association for the
Advancement of Science.

Unsurprisingly, these reactivity differences translate into differences in catalytic activity: Kaden

et al. showed that the CO oxidation activity of mass-selected Pd2 clusters is approximately two

times greater than that of Pd1 clusters (Figure 12).194 The activity of the clusters correlates with the

deviation of the Pd 3d binding energy from the normal bulk scaling behavior, suggesting that the

activities of the clusters are controlled by their unique properties at the sub-nanometric regime.

39
Discrete cluster models, precise down to the exact number of atoms, are thus necessary for

accurately modeling cluster catalysts. Preparing a suitable model typically requires finding the

ground state structure, which is a complex global optimization problem. Multiple methods,

including simulated annealing,195,196 basin hopping,197,198 and genetic algorithms,199–201 have been

developed to tackle this challenge. Yet, locating the ground state structure of a cluster remains

computationally expensive.

Modelling Supports

The complexity of cluster reactivity studies is further compounded by the increasing experimental

and theoretical evidence that the catalyst support can substantially influence catalytic activity of

clusters. Small clusters have a larger proportion of their surface area in contact with the support,

leading to increased charge transfer between clusters and supports on a per atom basis. For example,

Lykhach et al. showed that the charge transfer per atom between the cluster and the support is

maximized for cluster sizes < 100 atoms.202 Additionally, the effects of supports on the catalyst

particle can persist to relatively large particle sizes: the heats of adsorption of Ag atoms onto Ag

nanoparticles supported on metal oxides converge to the bulk heat of Ag sublimation only at

particle sizes > 4 nm for CeO2-x(111) (x = 0.1 and 0.2) and Fe3O4(111), and > 6 nm on

MgO(100).203

Even supports traditionally considered inert may exert a large influence if they contain defects or

if the catalyst particles are sufficiently small; Gao et al. showed that the activities of Au1 and Au2

clusters for O2 dissociation were significantly affected by binding on pristine h-BN, and even more

so for h-BN with N and B vacancies.204 Schimmenti and Mavrikakis studied the decomposition of

formic acid over a Pd6 cluster supported on boron nitride (h-BN)205,206 and found that the reaction

40
proceeds via the formate (HCOO)-mediated pathway, whereas on an extended Pd(111) slab, more

representative of larger nanoparticle catalysts, the reaction proceeds through the carboxyl (COOH)

-mediated pathway instead.205 Interestingly, the type of vacancy on h-BN also influences the

selectivity of Pd6 clusters for the formic acid decomposition reaction,206 with N-vacancies favoring

the HCOO-mediated pathway and B-vacancies favoring the COOH-mediated pathway as a result

of the positive and negative charge induced on the cluster, respectively.

Besides altering the electronic structure of the cluster, supports may also impact reactivity by

forming metal-support interfaces. These interfaces are suspected active sites for many reactions

including CO oxidation,207 methanol synthesis,208 and the water gas shift reaction40,209,210. The high

activity of Au catalysts supported on TiO2 for room-temperature CO oxidation is an instructive

example of the bifunctional synergy of cluster catalyst particles with the catalyst-support interface.

In this example, the strongly-binding interface facilitates O2 dissociation, whereas the less

strongly-binding Au cluster facilitates recombination of CO* with the generated atomic O*.211,212

Due to the above factors, an explicit atomistic model for the catalyst-support interactions can be

crucial for cluster catalysts. The benefits of modeling supports, however, must be practically

weighed against its high computational cost. Modeling a support brings with it a whole new set of

challenges, such as determining the appropriate surface terminations,213 oxidation states (e.g. for

reducible oxides such as CeO2),214 and consideration of defects and vacancies215,216 in the support

material, all of which can affect the geometry and electronic structure of the supported catalyst,

and thus modulate its catalytic activity.217 For example, Yoon et al. showed that Au8 clusters on

MgO(001) are active for CO oxidation only when bound to F-center defects (oxygen vacancies),

whereas clusters bound to defect-free MgO(001) were inactive.188 DFT calculations suggest that

41
charge transfer from the F-center to the cluster strengthens the binding of O2 and thus increases

the catalytic activity of the cluster, whereas charge-neutral clusters are inert since they cannot

adsorb O2 well.188  

In another example, Sterrer et al. found using scanning tunneling microscopy (STM) that Au forms

2D clusters on thin 3-monolayer (ML) MgO(001) films supported on Ag(001), but forms 3D

clusters when the thickness of the MgO(001) films is increased to 8 ML.218 A similar effect was

observed when Mo-doped CaO(001) (2D) was used as opposed to a pristine CaO(001) surface

(3D).219 This was attributed to charge transfer from the support—specifically the Ag(001) substrate

and the Mo dopants—which drives the flattening of Au clusters to maximize electrostatic

interactions with the support.219

Capturing Dynamic Behavior

In the presence of adsorbates, clusters can also undergo substantial reconstruction. Since they have

low cohesive energies220 and stronger directional bonding due to their more localized orbitals,221

clusters strongly prefer to bind adsorbates in certain sites and geometries. Such adsorbate-induced

reconstructions, also known as cluster fluxionality, can dynamically generate active sites during

reaction.

One interesting example is illustrated by Wang et al., who used DFT calculations to model Au20

supported on CeO2.222 They showed that the adsorption of CO aided in the dissociation of single

Au atoms from the cluster, which were suspected to be the actual active sites for CO oxidation.222

In another example, Baxter et al. combined experiments with theory to study ethylene

dehydrogenation over Pt clusters supported on Al2O3 (Figure 13).223 The ground state of the

42
pristine cluster is a 3D prism. However, adsorption of ethylene molecules gradually shifts the

ground state towards a 2D layered structure. Upon adsorption of three ethylene molecules, the

cluster spontaneously undergoes a 3D to 2D reconstruction.223

Figure 13. Structures of Pt7 clusters adsorbed on Al2O3. (a) Ground state structure (I) and energetically
low-lying isomers (II-V) of Pt7/Al2O3. Panels I and IV are prismatic structures, while the rest are single-
layer structures. Adsorption energies (Eads), population prevalence at 700 K (P700 K), and the amount of
charge transfer from the support to the cluster (ΔQ) are given. ΔEads indicates the difference in energy from
the ground state structure. (b) Adsorption of one, two, and three ethylene molecules onto Pt7/Al2O3. The
differential binding energies of ethylene (Ei ethylene, i = 1-3), as well as the bonding character of the ethylene
molecules, are given. The prismatic structure is less stable than the single-layer structure when one ethylene
molecule is adsorbed. When three ethylene molecules are adsorbed, the prismatic structure reconstructs
spontaneously to the single-layer structure. Color code: grey – Pt, brown – Al, green – C, white – H, red –
O. Adapted with permission from reference 223. Copyright 2017 American Chemical Society.

This cluster fluxionality may be partially responsible for the high catalytic activity of clusters. In

the above study by Baxter et al., Pt7 clusters were found to be much more active than Pt8 clusters,

although their global minima had very similar 3D structures.223 However, a key difference between

Pt7 and Pt8 lies in the many energetically low-lying 2D isomers that the Pt7 possesses, which the

latter does not. Since the 2D isomers are thought to be more active than 3D isomers,223 the ability

43
of Pt7 to access these low-lying metastable states may explain its higher ethylene dehydrogenation

activity.

3.4. Theoretical Treatment of Liquid-Phase Reactions

Liquid-phase reactions are much more challenging to model than gas-phase reactions.224 Solvents

can interact with reactants, products, and intermediates in a variety of ways, significantly altering

reaction energetics and reaction mechanisms225,226. Furthermore, the facile diffusion of solvent

molecules leads to a breakdown in the harmonic approximation, which is commonly used to

estimate vibrational features in computational studies, making it extremely challenging to calculate

accurate free energies for solvated systems. In this section, we discuss three main methods to treat

solvent effects (Figure 14). In order of increasing computational cost, these are: 1) implicit solvent

(continuum) models, 2) microsolvation (cluster-continuum) models, and 3) explicit solvent models.

Recently, machine learning methods to study and treat solvation effects, such as the use of

convolutional neural networks to predict kinetic solvent parameters,227 have also been developed.

However, this is still a new and rapidly evolving field. We thus refer the interested reader to these

relevant papers for more details on this topic.228,229

Figure 14 Different approaches to treat solvation of a methoxide ion in water: Continuum, Cluster-
continuum, and Explicit-solvent. The light blue background represents a dielectric continuum that

44
approximates the solvent with a relative permeability of ε. The blue mesh represents the solute-solvent
interface. Color code for atoms: dark grey – C, red – O, white – H. Reprinted with permission from reference
230. Copyright 2020 Wiley.

Implicit (Continuum) Solvent Models

Computational-cost-wise the most efficient, implicit solvent models are widely used to simulate

the effects of solvation. Also known as continuum solvation models (CSMs),231,232 these methods

calculate a correction term arising from embedding the solute in a dielectric continuum. This

continuum represents the solvent in a mean-field formalism. Early models such as the Polarizable

Continuum Model (PCM)233 place the solute in a vacuum cavity created from spheres centered on

the solute atoms. However, the sharp discontinuity in dielectric constants at the solvent-solute

interface leads to discontinuities in atomic forces, making geometric optimizations and MD

simulations problematic.234 Density-based solvation models, such as the self-consistent continuum

solvation (SCCS) scheme235,236 and VASPsol,237,238 seek to solve this issue by factoring in the

electron density of the solute to engineer a smooth transition at the solute-solvent interface.

In addition to the electrostatic interactions between the solute and the continuum, many codes,

including SCCS and VASPsol, also include free energy corrections (e.g., based on the formalism

by Sundararaman and Arias239) for non-electrostatic effects, such as the energy required for

creating the cavity and dispersion interactions between the solvent and the solute.

Implicit solvent models have found applications in many heterogeneous catalytic reactions. They

are particularly widely used in electrocatalytic systems, where the presence of electrolytes plays

an important role in modulating catalytic performance. For example, Xu et al. applied VASPsol to

account for solvent effects in the computational design of metal SACs as electrocatalysts for

oxygen reduction reaction (ORR), oxygen evolution reaction (OER), and HER.240 Additionally,

45
VASPsol has been used to study the C-C bond formation step for CO2 electroreduction (CO2RR)

on transition-metal single-crystal electrodes,241 as well as to probe the detailed CO2RR mechanism

on Cu(100) and Cu(111).242 Yet, a recent study by Sundararaman and Schwarz cautioned against

over-trusting implicit solvent models when studying electrochemical systems.243 In that study, the

authors tested how the surface charge of metal surfaces changes as a function of the applied electric

potential in the presence of several implicit solvent models. Although all models produced correct

trends in the charge-versus-potential profiles, the absolute value of the predicted charge at any

potential was still underestimated by more than two times compared with experiments.243 This

inadequacy of the DFT + implicit solvation model to accurately describe the amount of charge on

the surface could adversely affect the ability of such models to simulate electrochemical

elementary steps, especially those that involve significant charge transfer.

Another pertinent issue when using implicit solvent models is that since they treat the solvent in

an averaged manner, they typically cannot handle directional interactions between the solute and

solvent, such as hydrogen bonding and asymmetrically charged solvents. The charge-asymmetric

nonlocally determined local-electric (CANDLE) solvation model was specifically designed to

handle asymmetrically charged solvents.244 CANDLE is able to describe the solvation of both

anions and cations in water and acetonitrile more accurately than SCSS, which could only describe

either anions or cations, depending on the specific parameters used.244 CANDLE was employed

by Yu and coworkers to evaluate the performance of a Ni/Fe phosphide electrocatalyst for water

splitting.245 There is, however, a lack of sufficient benchmarking against other solvation methods

for us to critically assess its accuracy.

46
Another interesting implicit solvation method, the 3D-reference interaction site model (3D-

RISM),246,247 uses the statistical mechanics of liquids to calculate 3D site distribution functions for

each solvent interaction site around a solute. By reproducing the equilibrium solvent structure, 3D-

RISM incorporates directional bonding in an averaged way. Although applied in many problems,

including the interaction of water with a Cu(100) surface,248 3D-RISM is known to have issues

dealing with hydrophobic interactions.248,249 Further correction schemes have been developed to

deal with these interactions.248,250 However, its transferability is limited as a result.

Explicit Solvent Models

The most straightforward approach to rectifying the deficiencies of implicit solvent models is to

explicitly account for the solvent molecules at the atomistic level. Unfortunately, there are usually

many energetically low-lying configurations in which the solvent molecules may be arranged.

Therefore, static calculations are typically unable to capture the dynamic nature of the solvent at

finite temperatures and the resulting ensemble of possible interactions with the solvent and solute.

Molecular dynamic simulations are one method for more rigorously sampling the configurational

space of the solvated system. The long timescales needed to obtain converged statistical averages,

combined with the high computational cost due to the numerous explicit solvent molecules, make

this an extremely costly approach. However, the main advantage of combining explicit solvation

with MD simulations is that the free energies of the system, including entropic effects, can

theoretically be calculated from statistical mechanics. Unfortunately, the absolute free energies of

a system are notoriously difficult to compute since higher energy states contribute significantly to

the evaluation of the entropy,251,252 but MD simulations are preferentially biased towards low

energy states. Thankfully, free energy differences between two states, such as reaction free

47
energies, can be more easily obtained by using enhanced sampling techniques such as

thermodynamic integration253 or metadynamics.254,255

As an example, Gono et al. used an explicit solvent approach combined with thermodynamic

integration to calculate the free energies of reaction for the OER at the water-TiO2 interface.256

They found that the free energies of reaction of the rate-determining step (OH*→O*+H++e-) was

reduced by 0.5 eV upon solvation.256 Herron et al. incorporated the effects of electric potential

together with explicit water solvation to study methanol electro-oxidation over Pt(111),

specifically the selectivity of the first dehydrogenation step of methanol.257 They showed that

solvation lowers the barriers for C-H bond breaking towards *CH2OH more than O-H bond

breaking towards *OCH3 since *OCH3 repels water from the interface. The effects of electric

potential, however, favored the formation of *OCH3, lowering its barrier for formation by ~0.2 eV

when going from 0.0 VSHE to +0.4 VSHE (SHE = standard hydrogen electrode), although it did not

have any effect on the barrier of *CH2OH formation.257

When computational costs are a concern, static calculations using ice-like bilayer structures of

water are frequently used to mimic the effects of water solvation. They, however, do not include

the entropic effects due to the ensembles of multiple energetically low-lying solvation structures.

Often, the solvent structures used are also not the most stable, raising questions of their reliability

in describing the solvent environment. Ludwig et al studied if using the global minimum solvation

structure was important for accurately determining the solvation energies of various adsorbates

including CO*, COH*, and CHO*, on Cu(211).258 For systems of 5-10 water molecules, arbitrary

solvent structures reproduced the solvation energy obtained from the global minimum structure

reasonably well, within a standard deviation of 0.15 eV.258 However, they also find that the work

48
function depends strongly on the exact solvent structure, and thus rigorous sampling of the various

local minima is needed.258

Figure 15. Comparison of vacuum (no solvent), implicit solvent, and explicit solvent (ab initio molecular
dynamics (AIMD)) approaches for computing adsorption energies (ΔEad; top) and solvation energies (ΔEsolv;
bottom) of adsorbates on Cu(211) (left) and Cu(111), Au(111), and Pt(111) (right). For CO on Cu(211), the
symbols for vacuum and implicit solvent are overlapping. Error bars for the AIMD data are calculated from
four MD simulations. “×” symbols represent time-averaged data (for AIMD simulations) or site-averaged
data (for vacuum and implicit solvent); “+” symbols represent the most stable configuration as obtained
from geometry optimizations. Reprinted with permission from reference 259. Copyright 2020 AIP
Publishing.

Heenen et al. compared the performance of the implicit VASPsol method versus AIMD

simulations of explicit solvents for predicting the adsorption energy (without entropy) of

adsorbates including CO*, OH*, OOH*, COH* on various Au, Cu, and Pt surfaces (Figure 15).259

Compared with explicit solvent simulations, they found that the implicit solvent approach had

large errors in adsorption energies (mean average error (MAE) of 0.24 eV) similar to that of static

vacuum calculations with no solvent (MAE of 0.25 eV). Therefore, the accuracy of the adsorption

energies was not improved by using implicit solvents. This was revealed to be caused by the

neglect of directional hydrogen bonding and the competitive adsorption of water (i.e. the

49
displacement of water from the surface by the adsorbate), which is an endothermic process

commonly overlooked in studies of solvation on surfaces.260

Microsolvation (Cluster-Continuum) Models

Microsolvation models, or alternatively, hybrid-cluster-continuum models,224,230 arise from a

simple idea: incorporate short-range directional bonding by using a handful of explicit solvent

molecules and long-range electrostatic interactions by using continuum models. This combination

promises to achieve the best of the implicit and explicit solvation worlds. Microsolvation models

agree well with more complex bilayer models, giving very similar results (within 0.04 eV) for the

aqueous solvation energies of oxygenates on Pt nanoparticles,261 as well as qualitatively

reproducing the potential energy diagram (PED) obtained with bilayer water models for methanol

decomposition on the (111) facets of Pd, Pt, and Ru.262

Compared with more rigorous explicit solvation calculations, however, microsolvation models

possess multiple deficiencies. Sicinska et al. showed that a microsolvation model with 2 explicit

water molecules underestimated the free activation energy of decarboxylation of a 4-pyridylacetic

acid zwitterion by > 17 kcal mol-1 (0.74 eV).263 The good agreement with bilayer models but poor

agreement with full treatments of explicit models suggest that microsolvation models have similar

deficiencies as bilayer models: they typically only sample a handful of possible configurations and

do not account for dynamic solvent effects. Additionally, it is also unclear how many solvent

molecules are required to converge solvation effects.

Recently, Simm et al. developed a systematic method to increase conformational sampling and

automatically determine the number of water molecules needed to capture most of the solvent

50
effects.264 The development of systematic ways to converge the results from these microsolvation

approaches are the key for unlocking their full potential.

4. First-Principles-Based Microkinetic Modeling for Heterogeneous Catalysis

Electronic structure methods can offer valuable atomic-level insights into how reaction

intermediates interact and react over a catalytic surface. Yet, they are unable to predict

macroscopic reaction kinetics data such as reaction rates, apparent reaction barriers, and reaction

orders. To bridge this gap between atomic-scale phenomena and macroscopic properties,

microkinetic models are required. In this section, we discuss the key advances in the two

predominant types of microkinetic models used: MF-MKMs and KMC simulations. We place

special focus on how coverage effects can be self-consistently incorporated into these models to

rigorously describe the reaction environment and active site as a function of reaction conditions.

4.1. First-Principles-Based Mean-Field Microkinetic Modeling

Pioneered by Dumesic et al.265,266 as well as Stoltze and Nørskov267,268 in the 1980s-1990s, MF-

MKMs are a popular microkinetic modeling tool in the field of heterogeneous catalysis and are

frequently applied in combination with electronic structure calculation methods to study many

important industrial catalytic reactions, including ammonia synthesis,269,270 methanol synthesis271–

276
, and Fischer-Tropsch synthesis.277–279 

MF-MKMs neglect the spatial distribution of sites and the adsorbates on a catalytic surface.

Therefore, besides the reaction conditions (pressure, temperature, feed composition), it only needs

the thermodynamic and kinetic parameters for each elementary step in the reaction system as input.

Ideally, all possible elementary steps for a specific chemistry should be included without any bias

51
towards a specific reaction mechanism. A fully explicit MF-MKM requires no assumptions about

rate-limiting, quasi-equilibrated, or irreversible steps, which are commonly found in simplified

kinetic models.

The rate equations of all the elementary steps in the reaction network can be generally represented

as a system of differential algebraic equations. By solving these systems of equations, a MF-MKM

simulates the trajectory of adsorbate coverages and partial pressures of gas species through time.

This allows for the unbiased prediction of the overall reaction rate and the reaction fluxes through

each individual elementary step, from which one can deduce the preferred reaction mechanism

under any considered reaction conditions. MF-MKMs also naturally predict the surface coverage

of intermediates under reaction conditions, allowing us to identify potentially site-poisoning

species. Further, rate-limiting step(s) can be identified quantitatively by calculating how much the

overall rate depends on each elementary step, such as via the degree-of-rate-control analysis

proposed by Campbell.280 The ability to make these quantitative predictions makes MF-MKMs an

ideal tool for analyzing catalytic reaction mechanisms.

Here, we review the recent applications of MF-MKM in combination with DFT calculations. We

start by discussing methods for calculating rate and equilibrium constants from first principles. We

then discuss how to incorporate coverage effects due to lateral adsorbate-adsorbate interactions

into MF-MKMs. This leads to an overview of the development of parameter estimation methods,

which facilitate model-experiment comparison and quantification of the accuracy of the DFT

models. We combine these concepts to see how coverage self-consistent first-principles-based

MF-MKMs can be achieved. Lastly, we end with case studies of the WGSR and formic acid

decomposition, which showcase how first-principles-based MF-MKMs studies can help in

52
elucidating catalytic reaction mechanisms and the nature of the active site as a function of reaction

conditions.

Calculation of Rate and Equilibrium Constants

One of the major challenges in setting up a MF-MKM is obtaining reasonable thermodynamic and

kinetic parameters. Parametrizing a MF-MKM historically relied on semi-empirical methods, such

as bond-order conservation (BOC) and group additivity, to estimate reaction thermochemistry and

kinetics.281 Nowadays, electronic structure calculations can be used to calculate free energies from

first principles, from which rate and equilibrium constants can be determined (see e.g. Gokhale et

al.282).

When calculating free energies via statistical thermodynamics and ab initio vibrational frequencies,

it is common to approximate the vibrational modes of adsorbates as harmonic oscillators. This

works well at low temperatures and for strongly binding adsorbates, where the perturbations of the

metal-adsorbate bonds are kept to a minimum. Such vibrational modes are characterized by large

vibrational frequencies and are well-described by the harmonic oscillator approximation. However,

other factors such as adsorbate coverage may still affect the accuracy of these calculations. For

example, the vibrational frequencies of H* increase with increasing coverage due to repulsion

between adsorbates,283 resulting in a decrease in the entropy of H* with increasing coverage. This

decrease in entropy with increasing adsorbate coverage was experimentally observed by García-

Diéguez, Hibbitts, and Iglesia, who studied the adsorption of H* on Pt nanoparticles.284

For adsorbates that are weakly binding, solvated systems, or systems at high temperatures,

however, the harmonic oscillator approximation breaks down and anharmonic effects start to play

53
a large role.285,286 Campbell and Sellers recently developed a new method for measuring the

entropies of weakly adsorbed molecular species287 such as methanol and alkanes. They find that

the entropies of these molecules are much larger than that predicted from the harmonic oscillator

approximation, and that they can be readily estimated from the gas-phase entropy, an easily

obtainable quantity, of the molecules (Equation 5):

o
Sads T   0.70S gas
o
T   3.3R (5)

where 𝑆 𝑇 is the standard entropy of the adsorbate at temperature 𝑇, 𝑆 𝑇 is the standard

gas phase entropy, and R is the molar gas constant. The breakdown of the harmonic oscillator

approximation is characterized by the presence of low-lying frequencies in the calculated

vibrational modes. There are several methods for dealing with such modes. The simplest possible

approach involves replacing them with an arbitrary frequency cutoff e.g. 50 cm-1. Grimme also

proposed a quasi-rigid-rotor-harmonic-oscillator (RRHO) approximation whereby the entropy of

a vibrational mode is calculated as a weighted average of the entropies of a harmonic oscillator

and free rotor.288 As the frequency of the vibrational mode tends to zero, the entropy tends to that

of the free rotor, whereas as the vibrational mode tends to infinity, the entropy tends to that of the

harmonic oscillator. While these methods achieve the aim of avoiding the spuriously high

entropies286 due to artificially low frequencies, they have little physical motivation behind them.

Recently, several more physically-grounded approaches to deal with anharmonicity have been

developed. Most of these methods rely on incorporating knowledge of the PES for the translation

of the adsorbate over the surface to calculate free energies more accurately. For example, Sprowl,

Campbell, Árnadóttir noted that the PES corresponding to motion parallel to a surface is periodic

54
with low but nevertheless significant barriers. By identifying these adsorbate modes and treating

them with hindered translator and rotor models,289 they achieved good agreement with experiments

for small adsorbates such as methanol and methane. Taking this one step further, Bajpai et. al

mapped out entire PESs for monoatomic adsorbates and calculated their free energies via solving

for the quantum mechanical states on the PESs.290,291 In a similar vein, Jørgensen and Grönbeck

calculated the translational entropy via a semiclassical canonical partition function obtained by

integrating over the PES.

The most reliable methods to calculate free energies for systems where anharmonicity is prominent

rely on AIMD simulations with enhanced sampling techniques to rigorously sample all the relevant

modes of the system. Such techniques are widely used in the field of zeolite catalysis,292,293 for

example, as the confinement of adsorbate in zeolite pores leads to a breakdown in the harmonic

approximation.294 The use of advanced simulation techniques to ensure rigorous sampling of the

PES is well-covered in a recent Perspective by Collinge et al.285

Treating Lateral Interactions in Mean-field Microkinetic Models

While early DFT studies for heterogeneous catalytic systems were generally conducted at the low-

adsorbate coverage limit, it soon became appreciated that such models could be unrealistic,

especially if the MF-MKMs were predicting high coverages of intermediates. The lateral

interactions that adsorbates exert—typically destabilizing, although in rarer cases stabilizing

interactions exist (e.g. through hydrogen bonding)—could significantly alter the energetics of the

entire reaction network.295,296 If these interactions were neglected, the inputted DFT-calculated

binding and activation energies would become inconsistent with the MF-MKM predicted solution,

rendering the model unreliable. For a reliable model, we thus require a coverage self-consistent

55
solution where the inputted reaction energetics are evaluated at the coverages predicted by the MF-

MKM.297

Despite the importance of coverage self-consistency, there are limited examples in the literature

of such coverage self-consistent MF-MKMs. This is largely because the evaluation of energetics

on high-coverage surfaces is computationally demanding: it is nontrivial to identify an appropriate

surface coverage for the study, and even if this is known, determining the ground state

configuration of adsorbates given a certain coverage is a complex global optimization problem.

A common strategy for achieving coverage self-consistency involves developing coverage-

dependent energetics. Typically, one first identifies the most likely most abundant surface

intermediate (MASI), which is usually the most strongly bound surface species. Then, the reaction

energetics are evaluated in the presence of different coverages of the MASI. This allows one to fit

a function of the reaction energetics versus coverage, which can then be used by the microkinetic

model to update the energetics of the reaction network on-the-fly based on the predicted MASI

coverage at each time step.

This strategy was adopted Grabow et al. to study the WGSR on Pt. They derived correlations of

the binding energies of surface intermediates as a function of the CO* coverage.298 Then, they

applied Brønsted-Evans-Polanyi (BEP) relations to account for changes in activation barriers with

coverage. Gokhale et al. used an approach similar to that of Grabow et al. to model methanol

decomposition on Pt.282 In another example, Getman and Schneider studied the NO oxidation

reaction on a Pt catalyst.299 They performed DFT calculations on Pt(111) at three different

coverages of atomic O between 0 and 0.5 ML and established linear correlations for all the binding

and activation energies as a function of O* coverage.299 Unlike the work by Grabow et al., who

56
applied BEP relations, they explicitly fitted the activation energies as a function of coverage. They

observed that all adsorbed species and transition states were destabilized with increasing O*

coverage. Under typical NO oxidation conditions, the O* coverage was predicted to be 0.25 – 0.40

ML; these spectator O* species were crucial for weakening the strong binding of O* enough for it

to react with NO*.

One could argue, though, that the coverage obtained by this type of studies is from interpolation

and may not have been rigorously evaluated using DFT; hence, the validity of this approach is

limited by the accuracy of the coverage-dependent correlations. Additionally, for systems

involving multiple strongly bound species, the construction of multi-variable coverage

dependencies becomes much more challenging. The assumed MASI may also not be the actual

MASI once coverage effects are considered. These difficulties pose significant challenges towards

the construction of reliable MF-MKMs. One solution is to ensure the self-consistency of the MF-

MKM with respect to its predicted coverages and the coverages at which its DFT energetics are

evaluated at. This concept of a coverage self-consistent MF-MKM will be discussed further in

Section 4.1.4.

Parameter Estimation for Evaluating Mean-field Microkinetic Models

Catalyst models inevitably contain deficiencies such as incomplete reaction mechanisms,

unaccounted coverage effects, and errors intrinsic to DFT. These deficiencies may lead to

discrepancies between DFT-based microkinetic model predictions and experimentally measured

reaction rates that span multiple orders of magnitude. To put the problem into perspective a 0.06

eV error in the DFT calculated activation energy at room temperature results in an order of

magnitude error in the rate constant of an elementary step.

57
Parameter estimation—the adjustment of the DFT-derived energetics parameters to minimize

model-experiment discrepancies—is a powerful tool for evaluating the validity of postulated

models for the nature of the active site. For example, Singh and coworkers studied the

decomposition of formic acid decomposition over Au catalysts by a combination of MF-MKM

and parameter estimation.195,300 They observed large discrepancies between their DFT energetics

and the model solutions predicted by fitting of the experimental data. This served as a red flag,

informing them that their active site models, consisting of extended (111), (100), and (211)

surfaces, were implausible. More details of this case study will be discussed in Section 4.1.6.

In another instructive example, Grabow and Mavrikakis performed MF-MKM study of methanol

synthesis on Cu(111).276 Their initial model using DFT-based energetics predicted reaction rates

very different from the experimental rates measured on industrial Cu/ZnO/Al2O3 catalysts. From

a parameter estimation analysis, they found that surface intermediates, particularly CxHyOz species,

needed to be stabilized by as much as 0.64 eV compared with the DFT values obtained on clean

Cu(111) to accurately describe the experimental data. From this, they concluded that either

surfaces with more undercoordinated atoms, such as Cu(100), Cu(110), or Cu(211) facets were the

active sites for methanol synthesis on Cu; or that the ZnO support plays an important role. Their

hypothesis was later confirmed by Behrens et al. who demonstrated through a combination of DFT

and in situ characterization methods that Cu steps decorated with Zn atoms were the actual active

site in industrial Cu/ZnO/Al2O3 catalysts for methanol synthesis.208

Due to the large of number of variables in a reaction network, the parameter estimation problem

is intrinsically complex with multiple possible solutions. Early examples of parameter estimation

studies, such as the above work on methanol synthesis by Grabow and Mavrikakis, generally

58
focused on finding solutions near the original DFT-derived initial guesses. This is, however, not

ideal since the located solutions may be far from the global minimum. Additionally, for systems

with large and complicated reaction mechanisms, the iterative process of guessing initial

parameters, solving the model, and updating the parameters based on model solution can be tedious

and time-consuming. A systematic procedure for tackling this global optimization problem is,

therefore, highly desired. Rubert-Nason et al. formulated the parameter estimation problem as a

single non-linear optimization problem (NLP) by combining it together with the MF-MKM.301 In

their formalism, they defined a unified objective function (obj) for their NLP using Equation 6:

obj  error    d   * p (6)

where error is the sum of squared errors between model-predicted reaction rates and experimental

measurements, d is the sum of the total influx minus the total outflux over all species (i.e.,

deviations from the reaction steady state), p is the sum of deviations from the original binding

energies and activation barriers calculated using DFT, and α and β are weighting parameters.

Minimizing obj allows for the spontaneous solution of the MF-MKM and the parameter estimation

problem up to 1000 times faster than the conventional method of treating the MF-MKM and

parameter estimation separately.

59
Figure 16. Schematic of a sequential method for solving the parameter estimation problem associated with
a microkinetic model. Adapted with permission from reference 302. Copyright 2017 American Chemical
Society.

Rangarajan et al. additionally developed a sequential-optimization-based framework to tackle this

optimization problem (Figure 16).302 Within this framework, the objective function to be

minimized is defined as a weighted summation of the model-experiment discrepancies and the

adjustments made to the DFT-derived energetics. The framework combines a differential-algebraic

equation (DAE) solver for solving the microkinetic model and a nonlinear numerical optimization

for minimizing the objective function. These are interconnected through a sensitivity analysis

scheme which identifies sensitive thermodynamic/kinetic parameters which influence the overall

reaction kinetics. Additionally, a Latin-hypercube-based multi-start approach, which has been

shown to be statistically more efficient for selecting values of input variables than random

sampling,303,304 was used to initialize initial guesses for the parameter adjustments needed,

enabling a rigorous search over the entire solution space. This sequential optimization framework

offers a promising tool for future parameter estimation analysis in first-principles-based MF-MKM

studies.

Self-Consistency between Atomic-Level Energetics and Macroscopic Kinetics

60
While MF-MKM has found major success for the mechanistic studies of many heterogeneous

catalytic reactions, there is still much room improving its synergy with electronic structure

calculations and reaction kinetics experiment. As described in the sections above, MF-MKMs by

themselves lack accurate means of describing the lateral interactions of adsorbates and the

coverage dependency of reaction energetics.

61
Figure 17. Proposed iterative scheme for arriving at a coverage self-consistent model of the catalytic active
site, using a combined approach of DFT calculations, reaction kinetics experiments, and microkinetic
modeling. Reprinted with permission from reference 297. Copyright 2020 American Chemical Society

62
To resolve these issues, our group proposed a synergistic framework combining first-principles

calculations, microkinetic modeling, and reaction kinetics experiments to construct self-consistent

MF-MKMs (Figure 17).297 Initially, DFT calculations are performed on a pre-chosen atomistic

catalyst model, such as a (111) surface for example, which serves as an initial guess for the catalyst

active site. The initial coverages of adsorbed species are typically chosen to be as close to the low

coverage limit as allowable by the choice of the unit cell, unless in situ experimental data for the

surface coverages are available. While performing the DFT calculations, reaction kinetics

experiments can be conducted in parallel. Once the calculations and experiments are completed,

the DFT results from the initial catalyst model are then used to construct and run a MF-MKM, the

predictions of which are then compared with experimental measurements (e.g.: reaction rates,

orders) after parameter estimation analysis.

The results are then assessed based on three criteria:

Criterion 1: Do the experimentally measured quantities (e.g. reaction rates, orders) agree with the

model-predicted ones after parameter estimation?

Criterion 2: Are the parameter adjustments within the typical DFT error bar of 0.1-0.2 eV?

Criterion 3: Are the predicted surface coverages consistent with the catalyst model used in the

DFT calculations?

If all three criteria are fulfilled, we have converged at a self-consistent solution. If not, the MF-

MKM needs to be revised by changing either the catalyst model (e.g.: by considering the (100)

facet instead of the (111) facet) or the coverage of spectator species. The DFT-energetics are then

reevaluated with the new model. This algorithm is repeated until all three criteria are satisfied.

63
Although this approach is computationally demanding, especially if multiple iterations are

required, it allows for rigorous generation of self-consistent computational models that accurately

reproduce the experimental results from reaction kinetics measurements, reflects the catalyst

structure of the active site, and captures the surface coverages of reaction intermediates decorating

the active site under reaction conditions.

Bhandari et al. used this approach to study formic acid decomposition on carbon-supported Pt

catalysts, by using Pt(111) and Pt(100) as the model facets for the active site.305 The first iteration

of their MF-MKMs, constructed using DFT-energetics evaluated on clean Pt surfaces, predicted

extremely high CO coverages of 0.92 and 0.93 ML on Pt(111) and Pt(100), respectively. Therefore,

they chose a higher CO coverage of 0.44 ML, near the saturation coverage of CO* experimentally

measured by nuclear microanalysis (NMA) on Pt(111) single crystals,306 for their second iteration.

Strikingly, the MF-MKMs constructed using the DFT energetics evaluated on the 0.44 ML CO-

covered surfaces predicted CO coverages of 0.33 and 0.50 ML on Pt(111) and Pt(100), respectively.

Since these coverages are close to the initial assumption of a 0.44 ML CO-covered surface,

coverage self-consistency was deemed to be reached. Based on these self-consistent MF-MKMs,

the authors observed that the experimentally measured reaction rates could be described by model

predictions on the 0.44 ML CO-covered Pt(111) surface only, while Pt(100) was likely poisoned

by CO* and did not contributed to the catalyst activity. Therefore, they concluded that the 0.44

ML CO-covered Pt(111) surface is a suitable model for the active sites on Pt/C catalysts during

formic acid decomposition.

Case Study: Mean-field Microkinetic Modeling of the Water Gas Shift Reaction

64
The WGSR (CO + H2O → CO2 + H2) is an industrially important reaction for hydrogen production

from steam reforming of methane and its subsequent purification. Cu is well-known to catalyze

low-temperature WGS at around 473 K.267,307 Despite its seemingly simple chemistry, the reaction

mechanism for the WGSR was once controversial. It was initially believed that the WGSR

proceeded via the redox mechanism on Cu, which involves the activation of water to form OH*

and H*, followed by O* formation via either direct OH* dissociation308 or disproportionation

(2OH* → H2O* + O*).267,307,309 Formate (HCOO*), presumably formed from CO* and OH*, was

also proposed as a possible key reaction intermediate,310–312 since HCOO* was experimentally

detected on Cu.313,314 Early modeling work for the WGSR primarily focused on either the redox

mechanism or the HCOO-mediated mechanism,267,307 while it was unsettled which one is clearly

preferred.

Figure 18 (a) Reaction network for the WGSR. A reaction scheme including both the surface redox
mechanism and the carboxyl mechanism is outlined. The DFT-calculated thermochemistry (E) and the
activation energy barriers (Ea) for all the elementary steps are given in eV. The minimum energy pathway
for the WGSR is highlighted with green. (b) Microkinetic model predictions for surface coverage of formate
(HCOO**), hydrogen (H*), carbon monoxide (CO*), and vacant sites as a function of total pressure for a
Cu catalyst operating at 463 K and 118 mL/cc flow of 7% CO, 21% H2O, 8.5% CO2, and 38% H2 with
balance inerts. Coverage of other species is < 10-4. Reprinted with permission from reference 315. Copyright
2008 American Chemical Society.

Departing from this line of thinking about WGSR, Gokhale et al.315 proposed a third mechanism

for WGS on Cu(111) involving the formation of a carboxyl (COOH*) intermediate from CO* and

OH*. From DFT-based calculations, the COOH* mechanism had lower activation energy barriers

65
compared to the other two WGSR mechanisms (Figure 18a). Formulating and solving a MF-MKM

that included all three possible WGSR mechanisms, they showed that COOH*mechanism was

clearly the preferred reaction pathway for WGSR catalyzed by Cu nanoparticles supported on

Al2O3. Importantly, under typical experimental conditions for WGSR, they predicted HCOO* to

be the MASI (Figure 18b), consistent with the spectroscopic observation of HCOO* on Cu

surfaces. Yet, this MASI was not responsible for turning reactants to products, but rather a

spectator species. Instead, *COOH is the intermediate responsible for turn-overs and remains

spectroscopically elusive to this date. These results were subsequently confirmed by Madon et al.

in another first-principles-based MF-MKM study combined with reaction kinetics experiments for

WGSR on Cu/Al2O3.316 The carboxyl mechanism has also been proposed as the dominant

mechanism for the WGSR on Pt(111),298 Pd(111)317 and Pt3Re(111).318

This case study exemplifies the key role of first-principles-based MF-MKMs for identifying key

reaction steps and intermediates, which can be difficult to determine from experimental studies

alone.

Case Study: Mean-field Microkinetic Modeling of Formic Acid Decomposition

Formic acid (HCOOH or FA) is a promising hydrogen (H2) storage compound due to its high H2

content (4.4 wt. % H2, 53 g H2 L-1 at 298 K), ease of transportation and storage, low toxicity, low

flammability, and wide availability from renewable sources such as biomass.319 Decomposing FA

to unlock its stored hydrogen requires a catalyst. It is, however, challenging to find suitable

catalysts for this reaction since undesired side products CO and H2O may be formed together with

the desired products, CO2 and H2.

66
Interestingly, several experimental studies have demonstrated that finely dispersed supported Au

catalysts exhibit excellent activity and almost complete selectivity towards H2.320–325 Mavrikakis

and coworkers combined DFT-based microkinetic modeling with reaction kinetics experiments to

determine the reaction mechanism and the nature of the active site for FA decomposition over SiC-

supported Au catalysts.195,300 Using parameter estimation, they first obtained the “true” reaction

energetics required for matching the experimentally measured reaction rates. From this, they ruled

out the possibility that Au(111), Au(100), or Au(211) were the active sites,300 since these surfaces

bind species significantly weaker (Figure 19a; red, blue, and green lines) than that required to

match the experiments (Figure 19a; purple line). This suggests that the active sites are much more

undercoordinated than the step-edge atoms on a Au(211) slab.

Figure 19. (a) Potential energy diagrams (PEDs) for FA decomposition via the formate-mediated pathway
on Au(111) (red), Au(100) (blue), and Au(211) (green) facets from DFT calculations. Also shown is the
PED of the best fitted solution (purple) to the experimental reaction rates measured over Au/SiC catalysts,
along with the confidence intervals (dashed lines) for the transition states. Reprinted with permission from

67
reference 300. Copyright 2014 Wiley. (b) PED for FA decomposition via the formate-mediated pathway on
Au18 (red line). PED for Au(211) is shown by a black line for comparison. Reprinted from reference 195.
(c) Schematic illustration of the HCOOH decomposition process on Au18. Color code for structures: Au –
yellow, C – grey, O – red, H – white. A pair of triangular ensembles, each comprised of three CN = 5 atoms,
are the active sites for FA decomposition on Au18 and are indicated in pink.

Motivated by this finding, Li et al. examined sub-nanometeric Aun clusters (n = 4–25) as potential

Au active site models for FA decomposition.195 Among which, Au18 was proposed to be the active

species since it yielded reasonable agreement with experiments: its PED for FA decomposition

was closest to the experimentally fitted solution (Figure 19b). The atoms on Au18 with a

coordination number (CN) of five are likely the active sites since on these atoms, the lowest

barriers for all elementary steps can be found. (Figure 19c).

This microkinetic modeling study of experimental reaction kinetics experiments for FA

decomposition on Au/SiC demonstrates how first-principles-based microkinetic modeling can be

used to reverse engineer experimental results to gain insights into the true energetics of a reaction,

facilitating comparison with first-principles calculations. This is a powerful approach for

elucidating the nature of the active sites in heterogeneous catalysis.297

Despite these recent advances, MF-MKMs still possess some intrinsic limitations. Due to the

mean-field approximation, MF-MKMs are well-suited for catalytic surfaces with uniform site

distributions such as those found in large (> ~5 nm) metal nanoparticles which mainly expose

extended metal facets. However, the mean-field approximation starts to break down for systems

that have strong spatial correlations, due either to the intrinsic geometry of the system, such as in

the case of the inhomogeneous surfaces of small clusters, or to adsorbate-adsorbate interactions.

KMC simulations are generally more suitable for modeling these catalytic systems, as discussed

in the section below.

68
4.2. On-Lattice Kinetic Monte Carlo Simulations

Figure 20. Processes in on-lattice KMC simulations. (a) Mapping of O2 dissociative adsorption on a
fcc(111) surface as a reaction pattern on the corresponding KMC lattice. (b) All possible reaction patterns
(outlined in dotted lines) are first searched for given an initial configuration. One reaction pattern is chosen,
and the lattice state is updated to give the final configuration. Adapted with permission from reference 326.
Copyright 2012 American Chemical Society.

On-lattice KMC simulations reproduce the spatial distribution of catalyst sites and adsorbates with

higher fidelity than MF-MKMs. MF-MKMs, as discussed in the previous section, typically adopt

a mean-field approach; since there is no defined lattice, spatial information about the adsorbates

and lattice sites cannot be included. On the other hand, on-lattice KMC simulations track the

positions of each adsorbate on a user-defined lattice. This on-lattice approach works well for

approximating atomistic models where adsorbates bind onto distinct, well-defined high-symmetry

sites. As an example, the reaction pattern for adsorbed molecular O2 dissociating on a fcc(111)

surface is shown in Figure 20a. The adsorption geometry change, from top sites (O2*) to fcc sites

(2O*), is readily captured. Besides dissociating, adsorbates can also diffuse, react, and desorb. In

a KMC simulation, each of these events is a state transition. These transitions are simulated as

stochastic Markov processes, where the probability of transitions pij occurring between states i and

j is independent of the history of the state and follows an exponential distribution depending on

69
the transition rate constant, kij. The probability of a transition between i and j occurring after time

t is given by Equation 7:

pij  t   kij exp   kij t  (7)

In any given lattice configuration, there may be multiple possible reactions. In each step of a KMC

simulation, only one of these patterns is chosen to be executed (Figure 20b). A simple algorithm

is to draw a random time tevent from the exponential distribution in equation 4 for all possible

reaction events. The event with the lowest tevent is chosen to be executed and the time of the system

is incremented by the chosen tevent. The next step can then begin by using the same procedure.

More efficient algorithms are possible; for more details on such algorithms, we refer the interested

reader to other reviews.327–329

Facilitated by the widespread sharing of efficient and user-friendly KMC codes, KMC simulations

have recently gained much attention from the heterogeneous catalysis modeling community. These

freely available codes include the graph-theoretical lattice simulation package Zacros330,331 (2011),

as well as kmos332 (2015) and MonteCoffee333 (2018). The latter two packages are written in

Python and can interface with the atomic simulation environment (ASE) which is commonly used

by the atomistic simulation community.334 These packages have lowered the barriers to entry for

building sophisticated KMC simulations to model a wide variety of complex catalytic chemistries,

including the WGSR,335 methanol oxidation,336 and the deacon reaction.337

Treating Lateral Interactions in KMC Simulations

Although it is possible in MF-MKMs to account for adsorbate-adsorbate interactions in a mean-

field way by assuming changes in the energy of the reactants as a function of coverage,299,298 MF-

70
MKMs have intrinsic difficulties in accounting for interactions that arise from spatial correlations

between adsorbates, such as adsorbate islanding,338 intrinsic inhomogeneities in site distributions

such as in small clusters,339 and confinement of adsorbates, for example, along the rows of (110)

surfaces.340 Such spatial correlations lead to a breakdown in the mean-field approximation,

rendering a mean-field treatment of lateral interactions highly inaccurate.

In on-lattice KMC models, adsorbate interactions are commonly treated by the use of cluster

expansion (CE) models, which expand lateral interactions in terms of point, pair-wise, and

multibody interactions between adsorbates.341 This captures how the adsorbates are spatially

distributed, allowing a more accurate description of the local adsorbate interactions, which are

critical for accurate catalytic modeling. For example, Hansen and Neurock simulated the

hydrogenation of ethylene on Pd(100) using DFT-based energetics and an interaction model

derived from bond-order conservation methods.342 They showed that interactions between

adsorbed H* alter the probability of two H* being located beside an adsorbed ethylene molecule,

crucially affecting the predicted H2 reaction order.

71
Figure 21 Mean-field kinetic models versus kinetic Monte Carlo (KMC) simulations for CO oxidation over
RuO2(110). Simulations were carried out at T= 510 K, p(O2) = 5.5 mbar, with varying p(CO). (a) Turnover
frequencies (TOFs) (left) and adsorbate coverages (right) predicted by KMC simulations. (b) TOFs and
adsorbate coverages predicted by mean-field kinetic models. Subscripts “ot” and “br” represent on-top and
bridge sites respectively. (c) Spatial distribution of Oot and vacancies after CO2 recombinative desorption
for three kinetic models. (I) KMC with no repulsive interactions. Vacancy pairs are formed, which lead to
poisoning of the surface as O2 dissociative adsorption occurs easily. (II) Mean-field approach. Random
mixing of the vacancies prevents O poisoning, allowing higher activity. (III) KMC with repulsive
interactions. Repulsive interactions disperse vacancies, preventing O poisoning and allowing high activity.
Reprinted with permission from reference 343. Copyright 2014 The Royal Society of Chemistry. 

CE models are usually trained with first-principles data beforehand and then used on-the-fly in

KMC simulations to rapidly evaluate the energies—including the effects of lateral interactions—

of the lattice configurations encountered. In this way, any effects of neighboring co-adsorbates on

the energetics of possible reaction events can be rigorously treated. Using BEP relations, these

72
changes in reaction energies can also be propagated to the activation energies of the reaction events,

thus affecting their probabilities of occuring.331

Schneider and coworkers used such an approach, a lattice reaction framework together with

GCMC and statistical averages over the system, to achieve most of the advantages of KMC

simulations without having to construct a full-blown KMC model. They studied NO oxidation on

a variety of Pt surfaces using a DFT-based cluster expansion scheme to describe how O binding

varies with coverage.344,345 They found good agreement with experimental apparent activation

energies and rate orders. The activity of the catalysts was found to be controlled by how strongly

O* binds to the catalyst, which was strongly dependent on O* coverage. At low O* coverages, the

surface binds O* too strongly, thus hindering its reaction with NO* to form NO2. However, at

higher coverages, the surface is passivated by increasing amounts of O*, weakening O* binding

and thus allowing the reaction to proceed.344

Yet, it is important to be cautious when using KMC simulations with CE models. Because of the

high fidelity with which KMC models reproduce spatial distributions, an accurate description of

the lateral interactions in KMC models is even more important than for MF-MKMs. Hess and Over

compared the CO oxidation rates predicted by KMC simulations (Figure 21a) and MF-MKMs

(Figure 21b) varied as a function of the interaction strength between O* adatoms.343 At low O-O

interaction strength, there is a large difference between the TOFs predicted by the two models: the

KMC-predicted TOF is 4 orders of magnitude lower than that from the mean-field microkinetic

model. It turns out that the KMC simulations overestimate the number of pair vacancies and

underestimate the number of point vacancies (Figure 21c; panel I). Since O2 easily dissociates in

73
these pair vacancies, the catalyst is prone to O* poisoning. MF-MKMs assume perfect mixing and

thus do not have this problem even if the O-O interaction strength is low (Figure 21c; panel II).

With increasing interaction strength, however, the number of pair vacancies in the KMC model is

reduced (Figure 21c; panel III). This reduces the differences in the TOFs predicted by the KMC

simulations and MF-MKMs. The high fidelity with which KMC simulations reproduce spatial

distributions is thus a double-edged sword—accurate parameters are needed so that we do not end

up simulating the wrong physics with high precision.

Despite the ability of CE models to describe adsorbate adlayers with high fidelity, the adoption of

CE models has been hindered by the cost and tediousness of building and training them.

Fortunately, recently several freely available python packages, such as ICET346 and CELL347, have

been developed to automate the construction of CE models. Machine learning methods have also

been applied to facilitate the construction of more complex CE models involving bidentate

species.348 Another problem with using CE models is that they may greatly decrease the speed of

KMC simulations, by up to 4 orders of magnitude.349 This has prompted research into techniques

to speed these calculations up by constructing superclusters, which merge multiple smaller clusters

to reduce the number of CE parameters in the simulation.349

KMC Simulations for Clusters

Small nanoclusters are typical examples where MF-MKMs fail. Fundamentally, clusters exhibit

discrete adsorbate coverages due to the small number of adsorbates that can be adsorbed on a

cluster at any time. They also present a spatially inhomogeneous surface due to the many unique

coordination environments present. These factors lead to a breakdown of the mean-field

74
approximation, necessitating the use of KMC simulations for simulating catalytic reactions on

clusters with higher fidelity.

Figure 22. CO oxidation activity of Pt nanoparticles of different shapes as a function of temperature, as

predicted from KMC simulations. Pt atoms are colored by their generalized coordination number.
Simulations were conducted under pressures of 20 mbar CO and 10 mbar O2. Adapted with permission
from reference 4, which was adapted from reference 350. Copyright 2019 Springer Nature.

KMC simulations of cluster reactivity have revealed some extraordinary properties of clusters that

could not have been achieved by the use of MF-MKMs. Recently, we studied the decomposition

of FA over Au18 using DFT-based KMC simulations together with a CE Hamiltonian to account

for lateral interactions.339 We showed that although Au18 has two active sites, the strong attractive

interactions between adsorbates when they occupy both active sites lead to poisoning the active

sites by pairs of HCOO* and H* adsorbates produced in the course of FA decomposition. As a

result, each Au18 cluster can only catalyze the decomposition of one HCOOH molecule at a time.

This is a unique example of kinetic isolation between turnovers that was able to explain the peculiar

75
formation of solely HD from monodeutrated FA;339 such specific mechanistic insights would not

have been discovered by using MF-MKMs.

In other examples, Jørgensen and Grönbeck used KMC simulations to study how the structural

complexities of nanoparticles of different shapes would affect their activity for CO oxidation.350

They showed that kinetic coupling between sites with widely varying reactivities was key for

enhancing the overall catalytic activity of the nanoparticle.350 Additionally, Nikbin and coworkers

showed, using DFT-based KMC simulations of CO oxidation over Aun– (n=6, 8, 10) clusters, that

the different sites on the cluster can have significantly different reactivities.351,352 They attributed

the high activity of Au8 clusters for CO oxidation to a multitude of effects including charge transfer,

cluster reconstruction, and different reaction mechanisms.352

Treating Lattice Reconstructions in KMC Simulations

In many KMC simulations, reconstructions of the catalyst model are not fully modelled. Although

the energetics of the reconstruction can be partially included via CE models, the underlying lattice

is usually assumed constant. However, as discussed in Section 3.3, catalyst particles are highly

dynamic, especially in the presence of adsorbates, which provide a large driving force for

reconstruction. Even large nanoparticles can undergo significant reconstruction: Cu nanoparticles

with diameters 3–6 nm, for example, were observed by in situ transmission electron microscopy

(TEM) to reconstruct under methanol synthesis conditions as a result of the adsorption of water.353

KMC simulations are a promising method to study such processes, which occur on length scales

larger and time scales longer than attainable by first-principles calculations.

76
Several interesting strategies to study reconstructions of 2D surfaces in KMC studies have been

adopted. Noussiou and Provata employed a lattice switching scheme to simulate the hex

reconstruction over Pt(100) for CO oxidation, which occurs above a certain coverage of CO.354

Hoffman, Scheffler, and Reuter studied the reduction of a pre-oxidized Pd(100) surface by CO by

using a multilattice KMC simulation. They defined two overlapping lattices in their simulations

corresponding to pristine Pd(100) and oxidized Pd(100), respectively. In the oxidized regions, the

pristine Pd(100) lattice is inactivated by populating it with a null species. This species cannot

undergo any reaction except for diffusing to the oxidized Pd(100) lattice to signify that the oxidized

region is now reduced.355 Recently, Papanikolaou et al. used KMC simulations to study how

changes in CO coverage may affect the distribution of surface metal alloy ensembles of different

size and shapes.356 They were able to achieve this by defining reaction steps for the swapping of

surface atoms, which allows the surface to evolve simultaneously while reactions are occurring on

the surface.

Despite the advances in methods to describe reconstructions of 2D lattices, to our knowledge there

are no methods that describe reconstructions of 3D particles. However, KMC studies of particle

growth and sintering may provide some clues on how this can be achieved. Particle growth and

sintering processes are crucial to the activity and stability of catalysts and have become an

increasingly important fundamental aspect of catalysis modeling. These processes underlie our

ability to synthesize metastable clusters of any shape and morphology,357 which serve as excellent

model systems to systematically study the effects of structure and chemical composition on

catalytic reactions.358–360 It is therefore exciting that DFT calculations361 and KMC simulations362

have seen increasingly applications in particle growth models.

77
Figure 23. KMC simulations of metal nanoparticle growth. (a) High resolution transition electron
microscopy (HRTEM) images of the sintering of Au nanoparticles over 128 seconds at 300 K. The scale
bar represents 1 nm. (b) Snapshots of KMC simulations of the sintering of∼4 nm Au nanoparticles (4812
atoms) at 600 K show similar geometries as the images in (a). Reprinted with permission from reference
363
. Copy 2019 American Physical Society.

In one example, Li et al. studied the growth and migration of Pt nanoparticles on SiO2(001) via

KMC simulations.364 Their models were able to reproduce the equilibrium shapes of the metal

nanoparticles as predicted by Wulff constructions. Lai and Evans used KMC simulations and

managed to qualitatively reproduce the experimentally observed sintering of two ~4 nm Au

nanoparticles using a simple energetics model where the energies of the nanoparticle was

decomposed into the sum of coordination-dependent energies of single atoms (Figure 23). Barriers

for the diffusion of the Au atoms were obtained by using BEP relations with different intercepts

depending on the type of facet on which the atom is diffusing on.363 These examples validate the

use of KMC approaches to model particle diffusion and growth.

The predictive power of KMC approaches may be increased by combining them with first-

principles studies for elucidating the atomistic-scale mechanisms of nanoparticles synthesis and

78
growth. For example, Fichthorn and coworkers used first-principles phase diagrams to shed light

on the interactions between adsorbed halides and capping agents on Cu(111) and Cu(100) surfaces,

which helped to explain the anisotropic growth of Cu nanowires.365 Zhang et al. also calculated

the barriers for Pt substitution versus hopping on Pd(100) surfaces to understand how Pd etches

out of Pt@Pd nanoparticles to give hollow Pt nanocages.360 It would be exciting to see researchers

adopting these first-principles insights towards more accurate descriptions of the evolution of

catalyst structures during KMC simulations.

The research on KMC simulations for catalysis and particle growth have mostly been developing

on separate tracks. However, we envision that KMC simulations that treat the dynamics of the

catalyst surface as kinetic events on the same level as the chemical transformations of reaction

intermediates will help yield significant insights into the interplay between the dynamic catalyst

surface and adsorbates. While adapting on-lattice methods to treat lattice reconstructions may be

initially cost-effective, it will be necessary to gradually move towards more computationally-

intensive off-lattice methods to accommodate the wide range of possible particle transformations

and reconstructions in real catalytic processes.366,367

One of the key challenges that will inhibit the widespread adoption of KMC is the high complexity

and mental burden of manually probing reaction mechanisms. Adaptive KMC methods, which

automatically construct reaction mechanisms on-the-fly during the simulation can help alleviate

this problem.368,369 The adaptive KMC method leverages on automated saddle point searches using

for example, the dimer method,370 metadynamics,254,371 and high temperature MD simulations.

Adaptive KMC methods are also suitable for enabling higher fidelity off-lattice simulations,372

which are crucial for studying lattice reconstructions.366,367

79
5. Theoretical Methods for Rapid in-silico Catalyst Design and Discovery

Obtaining first-principles data and building first-principles kinetic models for catalysts is often

tedious and not computationally feasible for large scale catalyst screening and discovery. One of

the key research thrusts in computational heterogeneous catalysis is to find ways to predict the

binding energies of species without having to perform costly first-principles calculations. We will

now introduce these methods and explore how they can help us understand the activity trends of

different catalysts and rapidly identify promising catalysts for subsequent more rigorous screening.

5.1. Scaling and Brønsted-Evans-Polanyi Relations

Binding Energy Linear Scaling Relations

Figure 24. Binding energies of CHx intermediates (crosses: x = 1; circles: x = 2; triangles: x = 3), NHx
intermediates (circles: x = 1; triangles: x = 2), OHx, and SHx intermediates plotted against binding energies
of C, N, O, and S, respectively. The binding energy of species A is defined as the total energy of A adsorbed
in the lowest energy position on the surface minus the sum of the total energies of A in vacuum and the

80
clean surface. The data points represent results for close-packed (black), open (blue), and stepped (red)
surfaces on various transition-metal surfaces. From reference 373, which was originally adapted from
reference 374. Copyright 2007 American Physical Society.

One of the most influential concepts for enabling large scale catalyst screening is the linear scaling

relation. The early work by Abild-Pedersen et al.374 demonstrated that the binding properties of

small molecular species are mostly governed by those of the atom via which the species is adsorbed

on the surface. For example, there exists linear correlations between the calculated binding

energies of CHx*, NHx*, OHx*, and SHx* with those of C*, N*, O*, and S*, respectively (Figure

24). Interestingly, the slopes of the linear scaling relations depend on the number of valence

electrons of the adsorbate. For example, atomic C* has four valence electrons, while CH*, CH2*,

and CH3* have valencies of 3, 2, and 1, respectively; therefore, the slopes of the scaling relations

of the binding energies of CH*, CH2*, and CH3* versus that of atomic C* are approximately 0.75

(3/4), 0.50 (2/4), and 0.25 (1/4), respectively.

Albeit its simplicity, similar types of linear scaling relations have since been observed for many

surface adsorbates, such as for C2 hydrocarbon species,375 whose binding energies also scale

linearly with that of atomic C*. Interestingly, the rule can also be generalized to bidentate C2HX

species that bind to the surface through both C atoms. For example, the binding energy of CHCH*,

which has a valency of 2 for each C atom, scales linearly with the binding energy of C with a slope

of approximately 1 (2/4 + 2/4).375

Scaling relations are especially crucial for simplifying more complex reaction systems since they

can predict the binding energies of a group of surface intermediates given the binding energy of

one or more structurally similar intermediates. This can be used to rapidly construct the

thermochemical profiles of reaction mechanisms. For example, Studt et al. applied C2 scaling

81
relations to estimate the thermodynamics for the selective hydrogenation of acetylene to ethylene

on various transition metal alloy surfaces using just a single binding energy: that of CH*.376 This

allowed them to quickly identify catalysts with favorable reaction thermochemistry.376 Medford

and coworkers present another instructive example of the use of scaling relations by probing the

conversion of synthesis gas to higher alcohols, which has a complex reaction mechanism involving

numerous CxHyOz* reaction intermediates.377,378 Specifically, they were able to scale the binding

energies of the CxHyOz* species with the binding energies of just C*, O*, or a linear combination

of the two, effectively reducing the number of parameters needed to describe the overall reaction

thermochemistry to two. Finally, we note that strategies involving scaling relations are particularly

valuable for electrocatalytic reaction systems as the onset potentials of electrocatalytic reactions,

such as ORR,379 electrochemical synthesis of ammonia,380 and formic acid electrooxidation,381 are

largely governed by thermochemistry alone.

While scaling relations have been established for many catalytic systems and are useful for

reducing system complexity, it should be cautioned that these relations are strongly dependent on

the adsorbates’ binding geometries. In the previously discussed study by Abild-Pedersen et al.,374

the offsets (but not the slopes) of the scaling relations for the binding energies of CH3* against that

of C* differed depending on whether they were constructed using 1) binding energies obtained on

the optimal binding sites or 2) binding energies obtained on top sites only. This indicates that

scaling relations are site-specific.374 Hong et al. arrived at a similar conclusion by studying the

scaling relation between COOH* and CO* on a series of (111) and (211) transition-metal surfaces,

as well as metal-doped MoS2 edges. They found distinct correlations for the (111) and (211)

surfaces, respectively. Yet, no linear scaling relation was found for metal-doped MoS2.382 Scaling

relations are thus only transferrable between materials of the same class and structure—no

82
universal correlation exists. However, this also offers opportunity for catalyst design through

“breaking” the scaling relations, which will be discussed in detail in Section 5.2.

Brønsted-Evans-Polanyi (BEP) Relations.

In addition to scaling relations between binding energies, there also exist linear correlations

between the thermochemistry of elementary reaction steps and their respective transition state

energetics. These are the BEP relations, which correlate the activation energies of a reaction with

its reaction energies. Early examples of BEP relations were established for non-catalytic or

homogenous catalytic systems: the original postulate by Brønsted was established for acid- and

base-catalyzed reactions;383 whereas Evans and Polanyi discovered similar correlations for radical

reactions.384 However, computational methods have since discovered many examples of BEP

relations in heterogeneous catalysis as well. For example, Nørskov and coworkers demonstrated

the existence of a BEP relation for simple diatomic dissociation reactions of N2, NO, CO, O2 on a

variety of transition metal surfaces.385 Other examples include the BEP relations developed for

methanol decomposition,386 CO2 reduction,387 glycerol decomposition388 and WGSR.389,390

The combination of scaling and BEP relations greatly facilitates the simplification of the

description of complex reaction systems, effectively reducing the parameters required for a full

description of their reaction kinetics and thermochemistry to just a handful. These relations can be

further incorporated into generating reactivity volcano plots which allow for efficient initial

catalyst design and screening, as further discussed in Section 5.2. We should caution readers

though, about the deficiency of scaling and BEP relations: these correlation methods come

naturally with fitting errors, which, in combination with the intrinsic error of DFT binding energies,

can propagate rapidly throughout the multiscale modeling pipeline and result in large errors in the

83
predicted reaction observables such as reaction rates and selectivities.391 Therefore, while scaling

and BEP relations serve as powerful tools for initial catalyst design and screening, detailed DFT-

based reaction mechanism studies on specific catalyst surface models may still be needed to verify

the suitability of proposed catalysts before beginning any in-depth experimental verification.

5.2. Sabatier Principle and Volcano Plots

Figure 25. (a) Volcano plot of the predicted oxygen reduction reaction (ORR) activity of close-packed
transition metal surfaces as a function of the calculated binding energy of O*. Binding energies are
referenced to the Ru(0001) values. Reprinted with permission from reference 392. Copyright 2004 American
Chemical Society. (b) Two-dimensional volcano plot of the predicted CO methanation turnover frequencies
(TOF) as a function of the calculated binding energies of O* and C* on the (211) stepped surfaces of
transition metals. Reprinted with permission from reference 393.

Paul Sabatier, awarded the 1912 Nobel Prize for Chemistry, theorized that the ideal catalyst binds

reaction intermediates with an optimal strength that is neither too strong nor too weak.394 This is

termed the Sabatier principle. Thanks to recent advances in computational catalysis, this concept

can be illustrated quantitatively through volcano plots correlating the activity of a catalyst with

binding energies of key intermediates.

Initial Catalyst Screening using Volcano Plots

84
The work by Nørskov et al. for the oxygen reduction reaction (ORR) is a classic example of

volcano plots in heterogeneous catalysis.392 They found that the computed ORR activity of various

transition metal surfaces against how strongly they bind atomic O* takes the shape of a volcano

with a peak at intermediate O* binding energies (Figure 25a). This feature can be rationalized as

follows: On the ascending arm of the volcano (e.g. W, Mo, Fe), hydrogenation of oxygenated

intermediates is the rate determining step (RDS) due to the strong binding of O*. Weaker O*

binding in this regime increases the overall rate of reaction. On the descending arm (e.g. Ag, Au),

however, dissociative adsorption of O2 is the RDS since O* binds too weakly. Stronger O* binding

in this regime would increase the overall rate of reaction. Near the peak of the volcano (e.g. Pd,

Pt), O* has an ideal binding strength that balances the kinetics of both hydrogenation of

oxygenated species and dissociative adsorption of O2, which leads to a maximum in the overall

ORR activity.

Fundamental surface properties that correlate with their catalytic activity are termed descriptors.

While the ORR activity can be correlated to a single descriptor, the O* binding energy, there can

be multiple descriptors for a single reaction. For example, in another study by Nørskov and co-

workers, the catalytic activity of the methanation reaction was expressed as a function of both the

binding energies of atomic O* and C* (Figure 25b).393 Assuming that the detailed volcano plot for

a reaction is already in place, rapid in-silico initial screening of new catalysts is possible as these

descriptors, typically binding energies, are inexpensive to compute from first principles.

When the form of volcano plot is unknown, volcano plots may be constructed from first principles,

by first developing scaling and BEP relations. Consider the example of the ORR in Figure 25a:

the O* binding energy serves as a suitable descriptor since most reaction intermediates are

85
adsorbed via their O atom to the catalyst surface. Given a specific O* binding energy, scaling

relations can be applied to approximate the binding energies of other reaction intermediates in the

reaction mechanism such as OOH* and OH*. BEP relations then allow for estimation of reaction

barriers from the predicted reaction energies of the elementary steps. These energetics are then fed

to a microkinetic model to obtain macroscopic kinetic data such as reaction rates. By plotting the

predicted reaction rate versus a range of O* binding energies, a volcano plot is ultimately

constructed.

Volcano plots have been very successful for rationalizing catalytic activity and designing new

catalysts.395,396 For example, Jensen et al. developed a volcano for the ORR in alkaline media,

which led to their discovery of Cu/Pt(111) NSAs showing 60 times higher catalytic activity in 0.1

M KOH than the commonly used Pt(111) ORR catalyst in 0.1 M HClO4,397 although Pt catalysts

are typically less active in base than in acid.398 Andersson et al. studied the methanation reaction

using volcano plots, and discovered that Fe–Ni alloys could be active, cost-efficient catalysts, a

result that was subsequently verified by experiment.399

86
Figure 26. Volcano plot relating the turnover frequency of O2 dissociation versus the O* binding energy,
with and without O-O interactions, at 600 K and 0.1 bar of O2(g). The vertical red line represents the
chemical potential of O* at which there is an equimolar ratio of NO(g) and NO2(g). Reprinted with
permission from reference 400. Copyright 2014 Creative Commons CC BY.

However, there are several rarely discussed assumptions regarding the use of volcano plots to

screen for new catalysts. Firstly, it is usually assumed that the mechanism does not change across

the descriptor space. New pathways may, however, show up especially in regions far away from

the sampled descriptor space. Secondly, descriptor values are typically computed in the limit of

low coverage when screening for new candidate catalysts. While this assumption is valid for

weakly binding surfaces, stronger binding surfaces may be highly populated with adsorbates,

tainted by impurities,401 or even undergo state changes such as oxidation and carburization.402,403

In other words, depending on the nature of the catalyst surface, the environment of the active site

under similar reaction conditions can be dramatically different. Taking the WGSR for example,

CO* is the MASI for Pt surfaces,298 whereas HCOO* is the MASI for Cu surfaces.315 These effects

are not accounted for in the calculated descriptor values, potentially leading to the incorrect

characterization of where a candidate catalyst belongs in the descriptor space. Interestingly, Frey

87
et al. showed that although the surfaces span a wide range of O* binding energies and activities at

the zero-coverage limit, the volcano for O2 dissociation activity against O* binding energy is

flattened after incorporating lateral interactions—most surfaces cluster together with similar

binding energies and activities (Figure 26).400 This shows that care must be taken to recreate

realistic surface conditions when calculating descriptors on different surfaces.

Breaking Scaling Relations

Figure 27 Strategies for preferentially stabilizing COH* relative to CO* on metal surfaces, thus breaking
the scaling relations between these adsorbates. Such strategies may also be generally applicable to other
adsorbates. Adapted with permission from reference 373; originally adapted from 404. Copyright 2012
American Chemical Society.

Volcano plots and descriptors provide a methodological and systematic way to screen for

promising catalysts. Yet, scaling relations, based on which volcano plots are constructed,

intrinsically place a constraint on the maximum activity of the ideal catalyst. To develop catalysts

more active than that those predicted by the volcano, it is necessary to “break” the scaling relations

by tuning the energies of intermediates independently of one another.396,405 Although commonly

referred to as “breaking” the scaling relations, this strategy is more accurately described as

88
circumventing the scaling relations, since the objective is to move from one set of scaling relations

onto another. This can be achieved in variety of ways as summarized in Figure 27.

3D Active Sites

As discussed in Section 5.1.1, scaling relations depend strongly on the binding geometry. Scaling

relations can thus be disrupted by designing catalysts which bind reaction intermediates differently.

For example, one could design 3D active sites to take advantage of the additional binding

dimension that traditional 2D catalyst surfaces lack. This can be achieved by using ligands to

selectively tether adsorbates on their non-binding end. Alternatively, a three-dimensional catalyst

can be used. For example, DFT studies showed that RuO2 nanopores ~6 nm in diameter selectively

stabilize OOH* relative to OH*, thus reducing the overpotential for the oxygen evolution

reaction.406 This is because OOH*, being longer than OH*, can form hydrogen bonds with the

opposite channel wall, while OH* cannot.

Alloying

Alloying is another method for altering binding geometries. By eliminating specific ensembles on

catalyst surfaces, we can bias the binding site preferences of the adsorbate.407,408. Liu et al. showed

that on Pt/Cu SAAs, CO* is forced to bind on the top sites of isolated Pt atoms, although it prefers

to bind on bridge sites at higher Pt loadings where Pt ensembles can form.409 Forcing CO* to bind

to top sites weakens its binding, freeing up more sites for H2 activation. The catalyst’s activity for

hydrocarbon hydrogenation was therefore increased.409 By analyzing a variety of reactions using

DFT calculations, Darby et al. found that scaling relations for SAAs which do not alter the

preferred adsorption site of an adsorbate are similar to those for pure metal surfaces.410,411 They

89
thus attributed the ability of SAAs to break scaling relations to their ability to change the preferred

binding sites of adsorbates. However, the electronic (ligand) effects of alloying cannot be totally

ruled out: for example, the unique activity of SAAs is thought to be partially due to the free-atom-

like state of the solute atom.412,413

Alloying can not only affect the scaling relations for adsorbates, it can also help break the scaling

relations for transition states, i.e.: BEP correlations. To the best of our knowledge, the first example

reported in the literature for such an effect was the discovery of ideal bimetallic NSAs, which

activated H2 much easier that monometallic noble metal surfaces although they bind H*, the

product of H2 dissociation, as weakly as these metal surfaces.145 This new class of active sites can

operate at lower temperatures than those on monometallic surfaces, while maintaining enhanced

reaction activity and selectivity. The unique reactivity of NSA was suggested to be due to an

electronic (ligand) effect, specifically the difference in the interactions of the antibonding orbitals

of the transition state for H2 dissociation with the d-metal states.145

BEP correlations may also be broken by geometric effects. For example, Pt/Cu SAAs inhibit

reactions that require multiple reaction sites, one example being propane dehydrogenation.414 A

key descriptor for propylene selectivity in this reaction is the difference in the propylene desorption

energy and the barrier for further propylene dehydrogenation. For monometallic and bimetallic

alloy catalysts, both these quantities are linearly correlated to the first propane dehydrogenation

barrier (a general descriptor of the reactivity of the surface). For Pt/Cu SAAs (blue symbols; Figure

28a), however, these correlations are broken as the SAAs give rise to high barriers for further

dehydrogenation although they have propylene desorption energies similar to that of other alloys

(grey, green, and orange symbols; Figure 28a). Because propylene binds on top sites, SAAs do not

90
affect the desorption energy of propylene. Propylene dehydrogenation, however, is inhibited since

it requires ensembles of Pt atoms, which are not present on Pt/Cu SSAs (Figure 28b).414

Figure 28 Breaking scaling relations for propylene dehydrogenation by using a Pt/Cu single atom alloy
(SAA). (a) Scaling relations of the free energy barriers for the first dehydrogenation step versus further
deep dehydrogenation (left axis; grey symbols and stars) and versus the propylene desorption energy (right
axis; green symbols and circles). The deep dehydrogenation barrier is defined as the highest barrier for
dehydrogenation of propylene to C3H4. (b) Binding energy differences (ΔE) of various dehydrogenated
intermediates between various Pt and Pt/Cu SAA surfaces. Structures of the intermediates on Pt(111) and
Pt/Cu(111)-SAA are provided. A positive ΔE denotes weaker binding on Pt/Cu SAA surfaces. Color-code
for structures: Pt—blue; Cu—orange; C—gray; H—white. Adapted with permission from reference 414.
Copyright 2018 Creative Commons CC BY.

Bifunctional Catalysts

Bifunctional catalysts represent another class of catalysts which can circumvent the scaling

relations by combining two classes of materials, such as in a metal alloy,415 a metal catalyst with

91
an oxide support,416 or even a metal support with a oxide catalyst.417–419 For example, for metal-

oxide-supported Au catalysts the Au/metal oxide interface excels at oxygen activation, whereas

Au itself facilitates oxidation of CO using the oxygen spilled over from the interface.211,212 Using

MF-MKMs, Andersen et al. showed that as long as the two materials do not obey the same scaling

relation, bifunctional materials can indeed bring a large increase in activity compared to the

coupling of two materials following the same scaling relation.420,421

Metal/metal-oxide interfaces indeed have vastly different scaling relations.418,422,423 Interestingly,

Choksi, Majumdar, and Greeley also found that the linear scaling relations for adsorbates at

Au/MgO interfaces are strangely not constrained by bond order conservation, unlike for other

classes of materials such as metals and zeolites.422 For example, the slope of binding energies for

NH2* vs. N* would be 0.33 if using bond order reasoning; however, they found that at the Au/MgO

interface, the slope was 0.64 instead. They postulate that this deviation is due to the electrostatic

interactions between the dipole induced by the charge transfer between the metal and the oxide,

and the dipole of the adsorbate.422

Another interesting example of potential bifunctionality is nanoporous gold, which is generated

by dealloying of AuAg alloys.424 The process of dealloying leaves residual amounts of silver of

several at. %. The residual silver is proposed to increase supply of reactive O* either by O2

dissociative adsorption on the silver and subsequent spillover, or by perturbing the local electronic

structure of Au to facilitate O2 dissociative adsorption. This enabled the highly selective oxidative

coupling of methanol to methyl formate at low temperatures of less than 80°C,425 as well as the

high activity for CO oxidation at ambient pressures and temperatures of -20°C.426

Strain

92
Strain has also been shown to be a suitable way to break scaling relations. Khorshidi et al. recently

developed an eigenstress model to explain why strain strengthens binding at certain sites but

weakens binding at others.166 They showed that this principle could be exploited to break scaling

relations by applying anisotropic strain which selectively destabilizes the IS while stabilizing the

transition state.166

Indeed, calculations by Xie et al. showed that strain breaks the scaling relations between ORR

intermediates on N-doped graphene.427 By applying tensile strain, they found that the adsorption

strength of O* could be altered without changing the adsorption strength of OH* and OOH*.

Interestingly, compressive strain, which leads to a rippling deformation, does not break the scaling

relations. This was attributed to the similar tendency of adsorption of O* and tensile strain to

facilitate breaking of the C-N bonds in the N-doped graphene substrate.427 Additionally, Kropp

and Mavrikakis studied how strain affected the reactivity of graphene film. They determined that

strain affected the binding energies of O2 differently depending on its adsorption modes (e.g.

bidentate versus monodentate adsorption). This may serve as a starting point towards leveraging

upon such adsorption mode differences to break scaling relations.428

5.3. Identification of Reactivity Descriptors

Although scaling relations and volcano plots have been used to considerable success in

heterogeneous catalysis research, there is a lack of underlying physical principles behind the use

of the binding strength of one adsorbate as an energetic descriptor for another. This has motivated

the search for more fundamental quantities that can be used to describe the reactivity of catalysts

with higher accuracy, particularly for transition metal surfaces.

93
Early experimental work by Rodriguez and Goodman hinted at the use of the electronic-structure

properties of the surface atoms as a potential descriptor.429 They probed the adsorption of CO on

Pd, Ni, and Cu thin films pseudomorphically deposited on various metal single-crystal substrates

and observed a strong correlation between the CO desorption temperature and the core-level shift

of the metal surface measured via X-ray photoelectron spectroscopy (XPS).

Figure 29. Schematic illustration of the orbital interactions between an adsorbate’s valence level and the s
and d states of a transition-metal surface during the formation of an adsorbate-surface bond. DOS = density
of states. Reprinted with permission from reference 393; originally adapted from reference 430. Copyright
2005 Springer Nature.

With the rise of electronic structure calculations, theorists sought descriptors for surface reactivity

that were easier to compute than core-level shifts for example. In 1995, Hammer and Nørskov

proposed the d-band model, which is now one of the key electronic-structure reactivity

descriptors.431 The basic concepts behind the d-band model can be understood in terms of the

interactions of the adsorbate valence level and the metal s and d states, as illustrated in Figure

29.While the interactions with the s states give rise to an approximately constant contribution to

the binding energy on all transition metals, the interactions with the d states differ depending on

the metal. This can be understood by examining the anti-bonding states formed by coupling with

the d states, since the strength of the adsorbate-surface bond is determined by the filling of these

94
anti-bonding states. It turns out that the filling of the anti-bonding states correlates well with the

position of the d-band center relative to the Fermi level. This is because higher-energy metal d

states give rise to a higher energy anti-bonding states, which are less filled, thereby leading to

stronger adsorbate-surface interactions.

Figure 30. (a) Electrochemically measured changes in the hydrogen adsorption potential (ESCE) on various
Pd overlayer alloys versus the calculated shift of the d-band center (δεd) relative to that of Pd(111).
Reprinted with permission from reference 393 with data from reference 432. Copyright 2005 John Wiley and
Sons. The red circle represents the entry for Pd(111). (b) Binding energy of CO* plotted as a function of
the d-band center of various metal surfaces. Adapted with permission from reference 168. Copyright 1998
American Physical Society.

The d-band center is widely celebrated as a highly accurate yet easy-to-compute reactivity

descriptor in heterogeneous catalysis by transition metals and their alloys. The utility of this

descriptor has been verified by both theory and experiments. Kibler et al. studied the

electrochemical properties of Pd monolayers deposited on various close-packed transition metal

substrates and showed that the experimentally measured hydrogen adsorption potentials (ESCE)

correlated linearly with the calculated d-band center of the surface Pd atoms (Figure 30a).432

Additionally, Mavrikakis, Hammer, and Nørskov calculated the binding energies of CO* on

transition-metal surfaces of different local structure (e.g.: terraces, steps, kinks), on bimetallic

alloys and on surfaces experiencing different strain levels.168 As shown in Figure 30b, there are

95
excellent linear correlations between the CO* binding energy and the surface metal d-band center

for all the metallic systems studied.

Figure 31. (a) Solid lines: Projected density of states (PDOS) onto Pd 4d orbitals of clean Pd and Pd/Ag
alloys; dashed lines: PDOS onto the C 2p orbital of C adsorbed on Pd and Pd/Ag alloys. Vertical solid lines
denote the d-band center of a surface Pd atom in the (111) surface of the Pd and Pd alloys. (b) Illustration
of the variations in metal d-band shape induced by energy misalignment and interatomic coupling upon
alloy formation. (c) Adsorption energies (ΔE) of CH3* on pure transition metals and Pd alloys versus the
upper d-band edge (ɛu) of surface atoms. Reprinted with permission from reference 433. Copyright 2014
American Physical Society.

While the original d-band model is simple and widely applicable for many adsorbate/transition-

metal systems, there still exist outliers and exceptions where its accuracy is not satisfactory. For

example, Xin and Linic demonstrated434 that for adsorbates with nearly filled valence shells (e.g.,

OH*, F*, and Cl*) and metallic surfaces with nearly fully occupied d-bands (d9 and d10 metals),

the adsorbate-metal interactions do not follow the trend predicted by the d-band model, due to the

strong repulsion between the adsorbate states and the metallic d states.

In an attempt to develop a more accurate descriptor, Xin et al. proposed an improved d-band model

by considering the shape of the d-band in addition to the d-band center. The concept is illustrated

in Figure 31a-b:433 the adsorbate-metal interactions are governed by not only the center and filling

96
of the metal d-band, but also the band width and potentially higher moments of the d-band,

including skewness and kurtosis. They thus defined a new descriptor: the upper d-band edge (εu),

which is the highest peak position of the Hilbert transformation of the projected d-DOS, through

which the information about the higher moments of the metal DOS is included. They evaluated

the ability of their improved d-band model to predict the CH3* binding strength on a range of

monometallic transition metal surfaces as well as Pd alloys. Promisingly, they observed a linear

correlation between the DFT-calculated CH3* binding energies and εu for all the metals and alloys

studied (Figure 31c).433 This modified d-band model was able to predict the reaction energies for

the elementary steps involved in acetylene hydrogenation on various Pd3M alloys, where M is a

transition metal, with a mean absolute error (MAE) of just 0.07 eV compared to DFT calculations,

which is a significant improvement over the original d-band center model which gave an MAE of

0.15 eV.433

While the d-band model has found much success for describing adsorbate-metal interactions

extended metal surfaces, it is unclear if it works for metal clusters. Depending on the metal and

the size of the cluster, quantum effects, such as the presence of discrete energy levels, may strongly

affect cluster reactivity. The reactivity of small Au clusters, for example, may be largely affected

by quantum effects . Consider the binding of O* on a Au13 cluster, which has an average electron

level spacing near the d-band center of 0.7 eV since the d-band center lies far away from the Fermi

level. Because this is a significant proportion of the adsorbate-metal coupling matrix element of

3.4 eV, the quantum modulations of the O* binding energy on Au13 are therefore predicted to be

large.181 However, quantum effects are not expected to significantly affect O* binding on Pt13. For

Pt13, the d-band center lies close to the Fermi level where the density of states is high and the

average electron level spacing is near the d-band center is 0.03 eV. Since this is small compared

97
to the adsorbate-metal coupling matrix element of 4 eV,182 the O* binding energy remains

unaffected by quantum effects.

Additionally, how the d-band shifts with particle size for these small clusters still remains unclear.

d-band theory predicts a lower d-band center with particle size, as confirmed for experimentally

for Pt NPs down to 5 nm in diameter435 and also for Pd NPs down to 9 nm.436 For very small Au

clusters though, there is conflicting experimental and theoretical evidence. While photoelectron

spectroscopy (PES) experiments show that the Au d-band center indeed becomes lower with

decreasing particle size (for diameters ~0.5–3 nm),437 calculations instead predict that d-band

center for Au NPs becomes higher with decreasing particle size.438 More research is therefore need

to confirm whether the d-band center model can be applied to small clusters. 

6. Conclusions and Outlooks

Covering orders of magnitude in length and time scales, the modeling of heterogeneous catalysis

is truly a multidisciplinary field. While first-principles computations can give remarkably detailed

insights into the energetics of catalytic reactions occurring at the nanoscale, there is no simple way

to use these insights to rationalize the macroscopic reaction kinetics data that experiments provide.

Instead, a collection of different computational methods each specializing in the simulation of a

specific length and time scale, provides the key for connecting the experimental and theoretical

worlds. This multiscale approach is currently the gold standard for modeling heterogeneous

catalysts and forms the central basis of our review.

In Section 3, we reviewed how to construct physically and chemically accurate atomistic models,

which lies at the foundation of the multiscale approach. Specifically, we explored how improved

98
first-principles methods have enabled us to more accurately treat the interactions between the

different components of the catalytic system, including the surface, adsorbates, as well as possible

solvents. We also saw how realistic nanoparticle and cluster catalysts models can be built for a

variety of different catalyst types. These improvements will enable the catalysis community to

steadily march towards the holy grail of chemical accuracy for modelling the elementary steps of

catalytic reactions.

In Section 4, we analyzed the advances in microkinetic modelling, emphasizing the importance of

coverage self-consistency for realistic modeling of surface conditions. We also saw how kinetic

Monte Carlo simulations can enable the modeling of coverage effects and of catalysis on clusters

with high fidelity. These improvements in microkinetic modeling methods will enhance the

essential feedback loop between experiment and theory for identifying the nature of the active site

while the reaction is taking place and for developing new catalysts.

In Section 5, we finally looked at the advances in methods for identifying improved catalysts.

Approximations such as theoretical scaling relations, which capitalized on the similarity of binding

sites across different metals, have played a pivotal role for efficiently designing new catalysts in

the last decade. Yet, the next wave of research efforts will likely be marked by a push towards

breaking these scaling relations. This involves more rigorously reproducing realistic reaction

environments, individually assessing the unique active sites of various types of catalysts, and

studying how different active sites can synergistically couple with each other. This will be the key

for climbing above the volcano peaks, especially for well-studied reactions where creative

solutions are required to improve on existing, well-established catalysts.

99
A prominent recurring theme throughout this review was the increasing use of automation and

open-source software for improving catalytic modeling workflows. Although computational

resources will increase unboundedly, allowing unprecedented modelling of larger systems such as

supported nanoparticles, the ability of researchers to make sense of the complexity of reaction

networks and dynamic changes in the catalyst with time will become the limiting factor. Improved

automation, machine-learning tools, and the ability to integrate the plethora of new computational

methods and techniques into workflows will be key for improving our efficiency in designing new

heterogeneous catalysts.

We wrap up our review with a brief description of some of the key challenges facing researchers

in the field of computational heterogeneous catalysis. In the near term, the community will need

to continue building capabilities for probing atomic-scale surface phenomena and reaction

mechanisms in the presence of complex reaction environments. For example, much work remains

to be done in the areas of solvent modeling, especially with more complicated solvents such as

ionic liquids. Another example is the treatment of metal oxide catalysts, many of which require

highly accurate and costly hybrid-DFT calculations. Such limitations hinder our atomic-level

understanding of these catalysts despite their promise for many industrially relevant reactions. For

developments of this type to take place, the computational catalysis community will need to

proceed hand-in-hand with the experimental community studying model systems, for instance, to

benchmark the energetics of adsorption of species on active sites, the electronic structure of which

needs to be described more accurately than currently possible.

In the longer term, we envision challenges in translating the significant knowledge and insights

gained from computational studies into real-world catalysts and applications. As an example,

100
although we can now routinely predict highly active catalysts, we cannot yet predict how to

synthesize them or tell whether they would be stable under reaction conditions. To tackle this, new

methods will be necessary to further bridge the gap between theory and the entire experimental

chain from synthesis to deployment in industrial settings. Additionally, we envision the field

gradually venturing out of the realm of metal and metal oxide catalysts into new classes of

heterogeneous catalysts, for example, two-dimensional materials, including hexagonal boron-

nitride catalysts.439 Together with this shift, many of the concepts discussed in the present review,

such as catalytic descriptors and scaling relations, would need to be redeveloped as they are

applicable mainly to conventional metallic and metal-oxide catalytic systems. The development of

“universal” descriptors and scaling relations, which hold across a wide range of catalytic materials,

would be a significant step advancing the field in this new direction.

Competing Interests

The authors declare no competing financial interests.

Acknowledgments

This work was supported by the Department of Energy, Office of Basic Energy Sciences, Catalysis

Science Program (Grant DE-FG02-05ER15731). B.W.J.C thanks the Agency for Science,

Technology and Research (A*STAR) Singapore for partial support via a graduate fellowship. DFT

studies by our group discussed in this article were performed mostly at the National Energy

Research Scientific Computing Center, supported by DOE contract DE-AC02-05CH11231. We

thank Dr. Tibor Szilvási for his insightful comments on the manuscript.

101
References

(1) Nitrogen Statistics and Information, National Minerals Information Center

https://pubs.usgs.gov/periodicals/mcs2020/mcs2020-nitrogen.pdf.
(2) Dalena, F.; Senatore, A.; Basile, M.; Knani, S.; Basile, A.; Iulianelli, A. Advances in
Methanol Production and Utilization, with Particular Emphasis toward Hydrogen
Generation via Membrane Reactor Technology. Membranes 2018, 8, 98.
(3) Greeley, J.; Stephens, I. E. L.; Bondarenko, A. S.; Johansson, T. P.; Hansen, H. A.; Jaramillo,
T. F.; Rossmeisl, J.; Chorkendorff, I.; Nørskov, J. K. Alloys of Platinum and Early
Transition Metals as Oxygen Reduction Electrocatalysts. Nat. Chem. 2009, 1, 552–556.
(4) Bruix, A.; Margraf, J. T.; Andersen, M.; Reuter, K. First-Principles-Based Multiscale
Modelling of Heterogeneous Catalysis. Nat. Catal. 2019, 2, 659–670.
(5) Szalay, P. G.; Müller, T.; Gidofalvi, G.; Lischka, H.; Shepard, R. Multiconfiguration Self-
Consistent Field and Multireference Configuration Interaction Methods and Applications.
Chem. Rev. 2012, 112, 108–181.
(6) Shavitt, I. The Method of Configuration Interaction. In Methods of Electronic Structure
Theory; Schaefer, H. F., Ed.; Springer, Boston, MA: Boston, MA, 1977; pp 189–275.
(7) Bartlett, R. J.; Musiał, M. Coupled-Cluster Theory in Quantum Chemistry. Rev. Mod. Phys.
2007, 79, 291–352.
(8) Bartlett, R. J. Many-Body Perturbation Theory and Coupled Cluster Theory for Electron
Correlation in Molecules. Annu. Rev. Phys. Chem. 1981, 32, 359–401.
(9) Grimme, S. Improved Second-Order Møller-Plesset Perturbation Theory by Separate
Scaling of Parallel- and Antiparallel-Spin Pair Correlation Energies. J. Chem. Phys. 2003,
118, 9095–9102.
(10) Cremer, D. Møller-Plesset Perturbation Theory: From Small Molecule Methods to Methods
for Thousands of Atoms. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2011, 1, 509–530.
(11) Van Santen, R. A.; Neurock, M. Concepts in Theoretical Heterogeneous Catalytic
Reactivity. Catal. Rev. 1995, 37, 557–698.
(12) Hohenberg, P.; Kohn, W. Inhomogeneous Electron Gas. Phys. Rev. 1964, 136, B864–B871.
(13) Kohn, W.; Sham, L. J. Self-Consistent Equations Including Exchange and Correlation
Effects. Phys. Rev. 1965, 140, A1133–A1138.
(14) Mardirossian, N.; Head-Gordon, M. Thirty Years of Density Functional Theory in
Computational Chemistry: An Overview and Extensive Assessment of 200 Density
Functionals. Mol. Phys. 2017, 115, 2315–2372.
(15) Becke, A. D. Perspective: Fifty Years of Density-Functional Theory in Chemical Physics.
J. Chem. Phys. 2014, 140, 18A301.
(16) Perdew, J. P.; Wang, Y. Accurate and Simple Analytic Representation of the Electron-Gas
Correlation Energy. Phys. Rev. B 1992, 45, 244–249.
(17) Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized Gradient Approximation Made Simple.
Phys. Rev. Lett. 1996, 77, 3865–3868.
(18) Brown, W. A.; Kose, R.; King, D. A. Femtomole Adsorption Calorimetry on Single-Crystal
Surfaces. Chem. Rev. 1998, 98, 797–832.
(19) Wellendorff, J.; Silbaugh, T. L.; Garcia-Pintos, D.; Nørskov, J. K.; Bligaard, T.; Studt, F.;
Campbell, C. T. A Benchmark Database for Adsorption Bond Energies to Transition Metal
Surfaces and Comparison to Selected DFT Functionals. Surf. Sci. 2015, 640, 36–44.

102
(20) Bao, J. L.; Gagliardi, L.; Truhlar, D. G. Self-Interaction Error in Density Functional Theory:
An Appraisal. J. Phys. Chem. Lett. 2018, 9, 2353–2358.
(21) Rohrbach, G.; Hafner, J.; Kresse, G. Molecular Adsorption on the Surface of Strongly
Correlated Transition-Metal Oxides: A Case Study for CO/NiO(100). Phys. Rev. B 2004,
69, 075413.
(22) Kitchaev, D. A.; Peng, H.; Liu, Y.; Sun, J.; Perdew, J. P.; Ceder, G. Energetics of MnO2
Polymorphs in Density Functional Theory. Phys. Rev. B 2016, 93, 045132.
(23) Mori-Sánchez, P.; Cohen, A. J.; Yang, W. Localization and Delocalization Errors in Density
Functional Theory and Implications for Band-Gap Prediction. Phys. Rev. Lett. 2008, 100,
146401.
(24) Sham, L. J.; Schl̈ter, M. Density-Functional Theory of the Energy Gap. Phys. Rev. Lett.
1983, 51, 1888–1891.
(25) Mori-Sánchez, P.; Cohen, A. J.; Yang, W. Many-Electron Self-Interaction Error in
Approximate Density Functionals. J. Chem. Phys. 2006, 125, 201102.
(26) Berland, K.; Cooper, V. R.; Lee, K.; Schröder, E.; Thonhauser, T.; Hyldgaard, P.; Lundqvist,
B. I. Van Der Waals Forces in Density Functional Theory: A Review of the VdW-DF
Method. Reports Prog. Phys. 2015, 78, 066501.
(27) Carrasco, J.; Liu, W.; Michaelides, A.; Tkatchenko, A. Insight into the Description of van
Der Waals Forces for Benzene Adsorption on Transition Metal (111) Surfaces. J. Chem.
Phys. 2014, 140, 084704.
(28) Perdew, J. P.; Schmidt, K. Jacob’s Ladder of Density Functional Approximations for the
Exchange-Correlation Energy. In AIP Conference Proceedings; AIP, 2003; Vol. 577, pp 1–
20.
(29) Tao, J.; Perdew, J. P.; Staroverov, V. N.; Scuseria, G. E. Climbing the Density Functional
Ladder: Nonempirical Meta–Generalized Gradient Approximation Designed for Molecules
and Solids. Phys. Rev. Lett. 2003, 91, 146401.
(30) Perdew, J. P.; Ernzerhof, M.; Burke, K. Rationale for Mixing Exact Exchange with Density
Functional Approximations. J. Chem. Phys. 1996, 105, 9982–9985.
(31) Heyd, J.; Scuseria, G. E.; Ernzerhof, M. Hybrid Functionals Based on a Screened Coulomb
Potential. J. Chem. Phys. 2003, 118, 8207–8215.
(32) Heyd, J.; Scuseria, G. E. Efficient Hybrid Density Functional Calculations in Solids:
Assessment of the Heyd–Scuseria–Ernzerhof Screened Coulomb Hybrid Functional. J.
Chem. Phys. 2004, 121, 1187–1192.
(33) Krukau, A. V.; Vydrov, O. A.; Izmaylov, A. F.; Scuseria, G. E. Influence of the Exchange
Screening Parameter on the Performance of Screened Hybrid Functionals. J. Chem. Phys.
2006, 125, 224106.
(34) Stephens, P. J.; Devlin, F. J.; Chabalowski, C. F.; Frisch, M. J. Ab Initio Calculation of
Vibrational Absorption and Circular Dichroism Spectra Using Density Functional Force
Fields. J. Phys. Chem. 1994, 98, 11623–11627.
(35) Becke, A. D. Density‐functional Thermochemistry. III. The Role of Exact Exchange. J.
Chem. Phys. 1993, 98, 5648–5652.
(36) Paier, J.; Marsman, M.; Kresse, G. Why Does the B3LYP Hybrid Functional Fail for
Metals? J. Chem. Phys. 2007, 127, 024103.
(37) Carnimeo, I.; Baroni, S.; Giannozzi, P. Fast Hybrid Density-Functional Computations
Using Plane-Wave Basis Sets. Electron. Struct. 2019, 1, 015009.
(38) Xu, S.; Carter, E. A. Theoretical Insights into Heterogeneous (Photo)Electrochemical CO2

103
 Reduction. Chem. Rev. 2019, 119, 6631–6669.
(39) Gouveia, J. D.; Coutinho, J. Can We Rely on Hybrid-DFT Energies of Solid-State Problems
with Local-DFT Geometries? Electron. Struct. 2019, 1, 015008.
(40) Zhao, Z.-J.; Li, Z.; Cui, Y.; Zhu, H.; Schneider, W. F.; Delgass, W. N.; Ribeiro, F.; Greeley,
J. Importance of Metal-Oxide Interfaces in Heterogeneous Catalysis: A Combined DFT,
Microkinetic, and Experimental Study of Water-Gas Shift on Au/MgO. J. Catal. 2017, 345,
157–169.
(41) Zeng, Z.; Greeley, J. Theoretical Study of CO Adsorption on Au Catalysts under
Environmental Catalytic Conditions. Catal. Commun. 2014, 52, 78–83.
(42) Mazheika, A.; Levchenko, S. V. Ni Substitutional Defects in Bulk and at the (001) Surface
of MgO from First-Principles Calculations. J. Phys. Chem. C 2016, 120, 26934–26944.
(43) Mino, L.; Spoto, G.; Ferrari, A. M. CO2 Capture by TiO2 Anatase Surfaces: A Combined
DFT and FTIR Study. J. Phys. Chem. C 2014, 118, 25016–25026.
(44) Gao, G.; Jiao, Y.; Waclawik, E. R.; Du, A. Single Atom (Pd/Pt) Supported on Graphitic
Carbon Nitride as an Efficient Photocatalyst for Visible-Light Reduction of Carbon Dioxide.
J. Am. Chem. Soc. 2016, 138, 6292–6297.
(45) Paolucci, C.; Parekh, A. A.; Khurana, I.; Di Iorio, J. R.; Li, H.; Albarracin Caballero, J. D.;
Shih, A. J.; Anggara, T.; Delgass, W. N.; Miller, J. T. et al. Catalysis in a Cage: Condition-
Dependent Speciation and Dynamics of Exchanged Cu Cations in SSZ-13 Zeolites. J. Am.
Chem. Soc. 2016, 138, 6028–6048.
(46) Göltl, F.; Hafner, J. Structure and Properties of Metal-Exchanged Zeolites Studied Using
Gradient-Corrected and Hybrid Functionals. I. Structure and Energetics. J. Chem. Phys.
2012, 136, 064501.
(47) Göltl, F.; Hafner, J. Structure and Properties of Metal-Exchanged Zeolites Studied Using
Gradient-Corrected and Hybrid Functionals. II. Electronic Structure and
Photoluminescence Spectra. J. Chem. Phys. 2012, 136, 064502.
(48) Göltl, F.; Hafner, J. Structure and Properties of Metal-Exchanged Zeolites Studied Using
Gradient-Corrected and Hybrid Functionals. III. Energetics and Vibrational Spectroscopy
of Adsorbates. J. Chem. Phys. 2012, 136, 064503.
(49) Liu, W.; Carrasco, J.; Santra, B.; Michaelides, A.; Scheffler, M.; Tkatchenko, A. Benzene
Adsorbed on Metals: Concerted Effect of Covalency and van Der Waals Bonding. Phys.
Rev. B 2012, 86, 245405.
(50) Dion, M.; Rydberg, H.; Schröder, E.; Langreth, D. C.; Lundqvist, B. I. Van Der Waals
Density Functional for General Geometries. Phys. Rev. Lett. 2004, 92, 246401.
(51) Román-Pérez, G.; Soler, J. M. Efficient Implementation of a van Der Waals Density
Functional: Application to Double-Wall Carbon Nanotubes. Phys. Rev. Lett. 2009, 103,
096102.
(52) Klimeš, J.; Bowler, D. R.; Michaelides, A. Van Der Waals Density Functionals Applied to
Solids. Phys. Rev. B 2011, 83, 195131.
(53) Lee, K.; Murray, É. D.; Kong, L.; Lundqvist, B. I.; Langreth, D. C. Higher-Accuracy van
Der Waals Density Functional. Phys. Rev. B 2010, 82, 081101.
(54) Chakraborty, D.; Berland, K.; Thonhauser, T. Next-Generation Nonlocal van Der Waals
Density Functional. J. Chem. Theory Comput. 2020, 16, 5893–5911.
(55) Klimeš, J.; Bowler, D. R.; Michaelides, A. Chemical Accuracy for the van Der Waals
Density Functional. J. Phys. Condens. Matter 2010, 22, 022201.
(56) Wellendorff, J.; Lundgaard, K. T.; Møgelhøj, A.; Petzold, V.; Landis, D. D.; Nørskov, J.

104
 K.; Bligaard, T.; Jacobsen, K. W. Density Functionals for Surface Science: Exchange-
Correlation Model Development with Bayesian Error Estimation. Phys. Rev. B 2012, 85,
235149.
(57) Chan, K.; Tsai, C.; Hansen, H. A.; Nørskov, J. K. Molybdenum Sulfides and Selenides as
Possible Electrocatalysts for CO2 Reduction. ChemCatChem 2014, 6, 1899–1905.
(58) Studt, F.; Abild-Pedersen, F.; Varley, J. B.; Nørskov, J. K. CO and CO2 Hydrogenation to
Methanol Calculated Using the BEEF-VdW Functional. Catal. Letters 2013, 143, 71–73.
(59) Janssens, T. V. W.; Falsig, H.; Lundegaard, L. F.; Vennestrøm, P. N. R.; Rasmussen, S. B.;
Moses, P. G.; Giordanino, F.; Borfecchia, E.; Lomachenko, K. A.; Lamberti, C. et al.
Consistent Reaction Scheme for the Selective Catalytic Reduction of Nitrogen Oxides with
Ammonia. ACS Catal. 2015, 5, 2832–2845.
(60) Varykhalov, A.; Sánchez-Barriga, J.; Shikin, A. M.; Biswas, C.; Vescovo, E.; Rybkin, A.;
Marchenko, D.; Rader, O. Electronic and Magnetic Properties of Quasifreestanding
Graphene on Ni. Phys. Rev. Lett. 2008, 101, 157601.
(61) Mittendorfer, F.; Garhofer, A.; Redinger, J.; Klimeš, J.; Harl, J.; Kresse, G. Graphene on
Ni(111): Strong Interaction and Weak Adsorption. Phys. Rev. B 2011, 84, 201401.
(62) Medford, A. J.; Wellendorff, J.; Vojvodic, A.; Studt, F.; Abild-Pedersen, F.; Jacobsen, K.
W.; Bligaard, T.; Norskov, J. K. Assessing the Reliability of Calculated Catalytic Ammonia
Synthesis Rates. Science 2014, 345, 197–200.
(63) Grimme, S. Semiempirical GGA-Type Density Functional Constructed with a Long-Range
Dispersion Correction. J. Comput. Chem. 2006, 27, 1787-1799.
(64) Grimme, S.; Antony, J.; Ehrlich, S.; Krieg, H. A Consistent and Accurate Ab Initio
Parametrization of Density Functional Dispersion Correction (DFT-D) for the 94 Elements
H-Pu. J. Chem. Phys. 2010, 132, 154104.
(65) Grimme, S.; Ehrlich, S.; Goerigk, L. Effect of the Damping Function in Dispersion
Corrected Density Functional Theory. J. Comput. Chem. 2011, 32, 1456–1465.
(66) Caldeweyher, E.; Ehlert, S.; Hansen, A.; Neugebauer, H.; Spicher, S.; Bannwarth, C.;
Grimme, S. A Generally Applicable Atomic-Charge Dependent London Dispersion
Correction. J. Chem. Phys. 2019, 150, 154122.
(67) Tkatchenko, A.; Scheffler, M. Accurate Molecular Van Der Waals Interactions from
Ground-State Electron Density and Free-Atom Reference Data. Phys. Rev. Lett. 2009, 102,
073005.
(68) Steinmann, S. N.; Corminboeuf, C. A Generalized-Gradient Approximation Exchange Hole
Model for Dispersion Coefficients. J. Chem. Phys. 2011, 134, 044117.
(69) Steinmann, S. N.; Corminboeuf, C. Comprehensive Benchmarking of a Density-Dependent
Dispersion Correction. J. Chem. Theory Comput. 2011, 7, 3567–3577.
(70) Grimme, S.; Hansen, A.; Brandenburg, J. G.; Bannwarth, C. Dispersion-Corrected Mean-
Field Electronic Structure Methods. Chem. Rev. 2016, 116, 5105–5154.
(71) Grimme, S. Density Functional Theory with London Dispersion Corrections. WIREs
Comput. Mol. Sci. 2011, 1, 211–228.
(72) Stöhr, M.; Van Voorhis, T.; Tkatchenko, A. Theory and Practice of Modeling van Der
Waals Interactions in Electronic-Structure Calculations. Chem. Soc. Rev. 2019, 48, 4118–
4154.
(73) Deng, J.; Li, H.; Xiao, J.; Tu, Y.; Deng, D.; Yang, H.; Tian, H.; Li, J.; Ren, P.; Bao, X.
Triggering the Electrocatalytic Hydrogen Evolution Activity of the Inert Two-Dimensional
MoS2 Surface via Single-Atom Metal Doping. Energy Environ. Sci. 2015, 8, 1594–1601.

105
(74) Kerber, T.; Sierka, M.; Sauer, J. Application of Semiempirical Long-Range Dispersion
Corrections to Periodic Systems in Density Functional Theory. J. Comput. Chem. 2008, 29,
2088–2097.
(75) Ren, X.; Tkatchenko, A.; Rinke, P.; Scheffler, M. Beyond the Random-Phase
Approximation for the Electron Correlation Energy: The Importance of Single Excitations.
Phys. Rev. Lett. 2011, 106, 153003.
(76) Ren, X.; Rinke, P.; Joas, C.; Scheffler, M. Random-Phase Approximation and Its
Applications in Computational Chemistry and Materials Science. J. Mater. Sci. 2012, 47,
7447–7471.
(77) Lebègue, S.; Harl, J.; Gould, T.; Ángyán, J. G.; Kresse, G.; Dobson, J. F. Cohesive
Properties and Asymptotics of the Dispersion Interaction in Graphite by the Random Phase
Approximation. Phys. Rev. Lett. 2010, 105, 196401.
(78) Schimka, L.; Harl, J.; Stroppa, A.; Grüneis, A.; Marsman, M.; Mittendorfer, F.; Kresse, G.
Accurate Surface and Adsorption Energies from Many-Body Perturbation Theory. Nat.
Mater. 2010, 9, 741–744.
(79) Hammer, B.; Hansen, L. B.; Nørskov, J. K. Improved Adsorption Energetics within
Density-Functional Theory Using Revised Perdew-Burke-Ernzerhof Functionals. Phys. Rev.
B 1999, 59, 7413–7421.
(80) Zhao, Q.; Kulik, H. J. Stable Surfaces That Bind Too Tightly: Can Range-Separated
Hybrids or DFT+U Improve Paradoxical Descriptions of Surface Chemistry? J. Phys. Chem.
Lett. 2019, 10, 5090–5098.
(81) Ren, X.; Rinke, P.; Scheffler, M. Exploring the Random Phase Approximation: Application
to CO Adsorbed on Cu(111). Phys. Rev. B 2009, 80, 045402.
(82) Garrido Torres, J. A.; Ramberger, B.; Früchtl, H. A.; Schaub, R.; Kresse, G. Adsorption
Energies of Benzene on Close Packed Transition Metal Surfaces Using the Random Phase
Approximation. Phys. Rev. Mater. 2017, 1, 060803(R).
(83) Rohlfing, M.; Bredow, T. Binding Energy of Adsorbates on a Noble-Metal Surface:
Exchange and Correlation Effects. Phys. Rev. Lett. 2008, 101, 266106.
(84) Feibelman, P. J.; Hammer, B.; Nørskov, J. K.; Wagner, F.; Scheffler, M.; Stumpf, R.;
Watwe, R.; Dumesic, J. The CO/Pt(111) Puzzle. J. Phys. Chem. B 2001, 105, 4018–4025.
(85) Gil, A.; Clotet, A.; Ricart, J. M.; Kresse, G.; García-Hernández, M.; Rösch, N.; Sautet, P.
Site Preference of CO Chemisorbed on Pt(1 1 1) from Density Functional Calculations. Surf.
Sci. 2003, 530, 71–87.
(86) Stroppa, A.; Termentzidis, K.; Paier, J.; Kresse, G.; Hafner, J. CO Adsorption on Metal
Surfaces: A Hybrid Functional Study with Plane-Wave Basis Set. Phys. Rev. B 2007, 76,
195440.
(87) Kaltak, M.; Klimeš, J.; Kresse, G. Cubic Scaling Algorithm for the Random Phase
Approximation: Self-Interstitials and Vacancies in Si. Phys. Rev. B 2014, 90, 054115.
(88) Wilhelm, J.; Seewald, P.; Del Ben, M.; Hutter, J. Large-Scale Cubic-Scaling Random Phase
Approximation Correlation Energy Calculations Using a Gaussian Basis. J. Chem. Theory
Comput. 2016, 12, 5851–5859.
(89) Bowler, D. R.; Miyazaki, T. O(N) Methods in Electronic Structure Calculations. Rep. Prog.
Phys. 2012, 75, 036503.
(90) Goedecker, S. Linear Scaling Electronic Structure Methods. Rev. Mod. Phys. 1999, 71,
1085–1123.
(91) Kohn, W. Density Functional and Density Matrix Method Scaling Linearly with the

106
 Number of Atoms. Phys. Rev. Lett. 1996, 76, 3168–3171.
(92) He, L.; Vanderbilt, D. Exponential Decay Properties of Wannier Functions and Related
Quantities. Phys. Rev. Lett. 2001, 86, 5341–5344.
(93) Ismail-Beigi, S.; Arias, T. A. Locality of the Density Matrix in Metals, Semiconductors,
and Insulators. Phys. Rev. Lett. 1999, 82, 2127–2130.
(94) Prentice, J. C. A.; Aarons, J.; Womack, J. C.; Allen, A. E. A.; Andrinopoulos, L.; Anton,
L.; Bell, R. A.; Bhandari, A.; Bramley, G. A.; Charlton, R. J. et al. The ONETEP Linear-
Scaling Density Functional Theory Program. J. Chem. Phys. 2020, 152, 174111.
(95) Bowler, D. R.; Miyazaki, T. Calculations for Millions of Atoms with Density Functional
Theory: Linear Scaling Shows Its Potential. J. Phys. Condens. Matter 2010, 22, 074207.
(96) Genovese, L.; Neelov, A.; Goedecker, S.; Deutsch, T.; Ghasemi, S. A.; Willand, A.; Caliste,
D.; Zilberberg, O.; Rayson, M.; Bergman, A.; Schneider, R. Daubechies Wavelets as a Basis
Set for Density Functional Pseudopotential Calculations. J. Chem. Phys. 2008, 129, 014109.
(97) Soler, J. M.; Artacho, E.; Gale, J. D.; García, A.; Junquera, J.; Ordejón, P.; Sánchez-Portal,
D. The SIESTA Method for Ab Initio Order- N Materials Simulation. J. Phys. Condens.
Matter 2002, 14, 2745–2779.
(98) Ratcliff, L. E.; Mohr, S.; Huhs, G.; Deutsch, T.; Masella, M.; Genovese, L. Challenges in
Large Scale Quantum Mechanical Calculations. Wiley Interdiscip. Rev. Comput. Mol. Sci.
2017, 7, e1290.
(99) Heiss, M.; Fontana, Y.; Gustafsson, A.; Wüst, G.; Magen, C.; O’Regan, D. D.; Luo, J. W.;
Ketterer, B.; Conesa-Boj, S.; Kuhlmann, A. V. et al. Self-Assembled Quantum Dots in a
Nanowire System for Quantum Photonics. Nat. Mater. 2013, 12, 439–444.
(100) Goedecker, S. Decay Properties of the Finite-Temperature Density Matrix in Metals. Phys.
Rev. B 1998, 58, 3501–3502.
(101) Aarons, J.; Sarwar, M.; Thompsett, D.; Skylaris, C.-K. Perspective: Methods for Large-
Scale Density Functional Calculations on Metallic Systems. J. Chem. Phys. 2016, 145,
220901.
(102) Ruiz-Serrano, Á.; Skylaris, C.-K. A Variational Method for Density Functional Theory
Calculations on Metallic Systems with Thousands of Atoms. J. Chem. Phys. 2013, 139,
054107.
(103) Verga, L. G.; Aarons, J.; Sarwar, M.; Thompsett, D.; Russell, A. E.; Skylaris, C.-K. DFT
Calculation of Oxygen Adsorption on Platinum Nanoparticles: Coverage and Size Effects.
Faraday Discuss. 2018, 208, 497–522.
(104) Verga, L. G.; Aarons, J.; Sarwar, M.; Thompsett, D.; Russell, A. E.; Skylaris, C.-K. Effect
of Graphene Support on Large Pt Nanoparticles. Phys. Chem. Chem. Phys. 2016, 18,
32713–32722.
(105) Orbital-Free Kinetic-Energy Density Functional Theory. In Theoretical Methods in
Condensed Phase Chemistry; Schwartz, S. D., Ed.; Springer Netherlands; pp 117–184.
(106) Wang, Y. A.; Govind, N.; Carter, E. A. Orbital-Free Kinetic-Energy Functionals for the
Nearly Free Electron Gas. Phys. Rev. B 1998, 58, 13465–13471.
(107) Huang, C.; Carter, E. A. Toward an Orbital-Free Density Functional Theory of Transition
Metals Based on an Electron Density Decomposition. Phys. Rev. B 2012, 85, 045126.
(108) Hung, L.; Carter, E. A. Accurate Simulations of Metals at the Mesoscale: Explicit Treatment
of 1 Million Atoms with Quantum Mechanics. Chem. Phys. Lett. 2009, 475, 163–170.
(109) Boes, J. R.; Groenenboom, M. C.; Keith, J. A.; Kitchin, J. R. Neural Network and ReaxFF
Comparison for Au Properties. Int. J. Quantum Chem. 2016, 116, 979–987.

107
(110) Mueller, T.; Hernandez, A.; Wang, C. Machine Learning for Interatomic Potential Models.
J. Chem. Phys. 2020, 152, 050902.
(111) Behler, J. Perspective: Machine Learning Potentials for Atomistic Simulations. J. Chem.
Phys. 2016, 145, 170901.
(112) Deringer, V. L.; Caro, M. A.; Csányi, G. Machine Learning Interatomic Potentials as
Emerging Tools for Materials Science. Adv. Mater. 2019, 31, 1902765.
(113) Behler, J.; Parrinello, M. Generalized Neural-Network Representation of High-Dimensional
Potential-Energy Surfaces. Phys. Rev. Lett. 2007, 98, 146401.
(114) Natarajan, S. K.; Behler, J. Neural Network Molecular Dynamics Simulations of Solid-
Liquid Interfaces: Water at Low-Index Copper Surfaces. Phys. Chem. Chem. Phys. 2016,
18, 28704–28725.
(115) Shapeev, A. V. Moment Tensor Potentials: A Class of Systematically Improvable
Interatomic Potentials. Multiscale Model. Simul. 2016, 14, 1153–1173.
(116) Bartók, A. P.; Payne, M. C.; Kondor, R.; Csányi, G. Gaussian Approximation Potentials:
The Accuracy of Quantum Mechanics, without the Electrons. Phys. Rev. Lett. 2010, 104,
136403.
(117) Thompson, A. P.; Swiler, L. P.; Trott, C. R.; Foiles, S. M.; Tucker, G. J. Spectral Neighbor
Analysis Method for Automated Generation of Quantum-Accurate Interatomic Potentials.
J. Comput. Phys. 2015, 285, 316–330.
(118) Schlexer Lamoureux, P.; Winther, K. T.; Garrido Torres, J. A.; Streibel, V.; Zhao, M.;
Bajdich, M.; Abild‐Pedersen, F.; Bligaard, T. Machine Learning for Computational
Heterogeneous Catalysis. ChemCatChem 2019, 11, 3581–3601.
(119) Hansen, M. H.; Torres, J. A. G.; Jennings, P. C.; Wang, Z.; Boes, J. R.; Mamun, O. G.;
Bligaard, T. An Atomistic Machine Learning Package for Surface Science and Catalysis.
arXiv:physics/1904.00904 2019, 1–42.
(120) Khorshidi, A.; Peterson, A. A. Amp: A Modular Approach to Machine Learning in
Atomistic Simulations. Comput. Phys. Commun. 2016, 207, 310–324.
(121) Ulissi, Z. W.; Tang, M. T.; Xiao, J.; Liu, X.; Torelli, D. A.; Karamad, M.; Cummins, K.;
Hahn, C.; Lewis, N. S.; Jaramillo, T. F. et al. Machine-Learning Methods Enable Exhaustive
Searches for Active Bimetallic Facets and Reveal Active Site Motifs for CO2 Reduction.
ACS Catal. 2017, 7, 6600–6608.
(122) Bernstein, N.; Csányi, G.; Deringer, V. L. De Novo Exploration and Self-Guided Learning
of Potential-Energy Surfaces. npj Comput. Mater. 2019, 5, 1–9.
(123) Podryabinkin, E. V.; Shapeev, A. V. Active Learning of Linearly Parametrized Interatomic
Potentials. Comput. Mater. Sci. 2017, 140, 171–180.
(124) Kresse, G.; Furthmuller, J.; Furthmüller, J. Efficient Iterative Schemes for Ab Initio Total-
Energy Calculations Using a Plane-Wave Basis Set. Phys. Rev. B 1996, 54, 11169–11186.
(125) Kresse, G.; Furthmuller, J.; Furthmüller, J. Efficiency of Ab-Initio Total Energy
Calculations for Metals and Semiconductors Using a Plane-Wave Basis Set. Comput. Mater.
Sci. 1996, 6, 15–50.
(126) Jinnouchi, R.; Karsai, F.; Kresse, G. On-the-Fly Machine Learning Force Field Generation:
Application to Melting Points. Phys. Rev. B 2019, 100, 014105.
(127) Jinnouchi, R.; Lahnsteiner, J.; Karsai, F.; Kresse, G.; Bokdam, M. Phase Transitions of
Hybrid Perovskites Simulated by Machine-Learning Force Fields Trained on the Fly with
Bayesian Inference. Phys. Rev. Lett. 2019, 122, 225701.
(128) Garrido Torres, J. A.; Jennings, P. C.; Hansen, M. H.; Boes, J. R.; Bligaard, T. Low-Scaling

108
 Algorithm for Nudged Elastic Band Calculations Using a Surrogate Machine Learning
Model. Phys. Rev. Lett. 2019, 122, 156001.
(129) Iype, E.; Hütter, M.; Jansen, A. P. J.; Nedea, S. V.; Rindt, C. C. M. Parameterization of a
Reactive Force Field Using a Monte Carlo Algorithm. J. Comput. Chem. 2013, 34, 1143–
1154.
(130) Van Duin, A. C. T.; Dasgupta, S.; Lorant, F.; Goddard, W. A. ReaxFF: A Reactive Force
Field for Hydrocarbons. J. Phys. Chem. A 2001, 105, 9396–9409.
(131) Senftle, T. P.; Hong, S.; Islam, M. M.; Kylasa, S. B.; Zheng, Y.; Shin, Y. K.; Junkermeier,
C.; Engel-Herbert, R.; Janik, M. J.; Aktulga, H. M. et al. The ReaxFF Reactive Force-Field:
Development, Applications and Future Directions. npj Comput. Mater. 2016, 2, 15011.
(132) Shchygol, G.; Yakovlev, A.; Trnka, T.; Van Duin, A. C. T.; Verstraelen, T. ReaxFF
Parameter Optimization with Monte-Carlo and Evolutionary Algorithms: Guidelines and
Insights. J. Chem. Theory Comput. 2019, 15, 6799–6812.
(133) Sun, W.; Ceder, G. Efficient Creation and Convergence of Surface Slabs. Surf. Sci. 2013,
617, 53–59.
(134) Witman, M.; Ling, S.; Boyd, P.; Barthel, S.; Haranczyk, M.; Slater, B.; Smit, B. Cutting
Materials in Half: A Graph Theory Approach for Generating Crystal Surfaces and Its
Prediction of 2D Zeolites. ACS Cent. Sci. 2018, 4, 235–245.
(135) Novoselov, K. S.; Mishchenko, A.; Carvalho, A.; Castro Neto, A. H. 2D Materials and van
Der Waals Heterostructures. Science 2016, 353, aac9439.
(136) Voloshina, E. N.; Fertitta, E.; Garhofer, A.; Mittendorfer, F.; Fonin, M.; Thissen, A.;
Dedkov, Y. S. Electronic Structure and Imaging Contrast of Graphene Moiré on Metals. Sci.
Rep. 2013, 3, 1072.
(137) Lewandowski, M.; Pabisiak, T.; Michalak, N.; Miłosz, Z.; Babačić, V.; Wang, Y.;
Hermanowicz, M.; Palotás, K.; Jurga, S.; Kiejna, A. On the Structure of Ultrathin FeO Films
on Ag(111). Nanomaterials 2018, 8, 828.
(138) Freund, H. J.; Pacchioni, G. Oxide Ultra-Thin Films on Metals: New Materials for the
Design of Supported Metal Catalysts. Chem. Soc. Rev. 2008, 37, 2224–2242.
(139) Lazić, P. CellMatch: Combining Two Unit Cells into a Common Supercell with Minimal
Strain. Comput. Phys. Commun. 2015, 197, 324–334.
(140) Catlow, R.; Bell, R.; Cora, F.; French, S. A.; Slater, B.; Sokol, A. Computer Modelling of
Inorganic Materials. Annu. Reports Prog. Chem. - Sect. A 2005, 101, 513–547.
(141) Sokol, A. A.; Bromley, S. T.; French, S. A.; Catlow, C. R. A.; Sherwood, P. Hybrid
QM/MM Embedding Approach for the Treatment of Localized Surface States in Ionic
Materials. Int. J. Quantum Chem. 2004, 99, 695–712.
(142) Sinfelt, J. H. Catalysis by Alloys and Bimetallic Clusters. Acc. Chem. Res. 1977, 10, 15–20.
(143) Singh, A. K.; Xu, Q. Synergistic Catalysis over Bimetallic Alloy Nanoparticles.
ChemCatChem 2013, 5, 652–676.
(144) Groß, A. Reactivity of Bimetallic Systems Studied from First Principles. Top. Catal. 2006,
37, 29–39.
(145) Greeley, J.; Mavrikakis, M. Alloy Catalysts Designed from First Principles. Nat. Mater.
2004, 3, 810–815.
(146) Greeley, J.; Mavrikakis, M. Near-Surface Alloys for Hydrogen Fuel Cell Applications.
Catal. Today 2006, 111, 52–58.
(147) Marcinkowski, M. D.; Darby, M. T.; Liu, J.; Wimble, J. M.; Lucci, F. R.; Lee, S.;
Michaelides, A.; Flytzani-Stephanopoulos, M.; Stamatakis, M.; Sykes, E. C. H. Pt/Cu

109
 Single-Atom Alloys as Coke-Resistant Catalysts for Efficient C–H Activation. Nat. Chem.
2018, 10, 325–332.
(148) Ruban, A. V.; Skriver, H. L.; Nørskov, J. K. Surface Segregation Energies in Transition-
Metal Alloys. Phys. Rev. B - Condens. Matter Mater. Phys. 1999, 59, 15990–16000.
(149) Nilekar, A. U.; Ruban, A. V.; Mavrikakis, M. Surface Segregation Energies in Low-Index
Open Surfaces of Bimetallic Transition Metal Alloys. Surf. Sci. 2009, 603, 91–96.
(150) Cao, L.; Niu, L.; Mueller, T. Computationally Generated Maps of Surface Structures and
Catalytic Activities for Alloy Phase Diagrams. Proc. Natl. Acad. Sci. U. S. A. 2019, 116,
22044–22051.
(151) Cao, L.; Li, C.; Mueller, T. The Use of Cluster Expansions to Predict the Structures and
Properties of Surfaces and Nanostructured Materials. J. Chem. Inf. Model. 2018, 58, 2401–
2413.
(152) Sadigh, B.; Asta, M.; Ozolinš, V.; Schmid, A. K.; Bartelt, N. C.; Quong, A. A.; Hwang, R.
Q. Short-Range Order and Phase Stability of Surface Alloys: PdAu on Ru(0001). Phys. Rev.
Lett. 1999, 83, 1379–1382.
(153) Wieckhorst, O.; Müller, S.; Hammer, L.; Heinz, K. First-Principles-Based Surface Phase
Diagram of Fully Relaxed Binary Alloy Surfaces. Phys. Rev. Lett. 2004, 92, 195503.
(154) Li, H.; Guo, S.; Shin, K.; Wong, M. S.; Henkelman, G. Design of a Pd-Au Nitrite Reduction
Catalyst by Identifying and Optimizing Active Ensembles. ACS Catal. 2019, 9, 7957–7966.
(155) Tian, L. Y.; Levämäki, H.; Kuisma, M.; Kokko, K.; Nagy, A.; Vitos, L. Density Functional
Theory Description of Random Cu-Au Alloys. Phys. Rev. B 2019, 99, 064202.
(156) Mayrhofer, K. J. J.; Juhart, V.; Hartl, K.; Hanzlik, M.; Arenz, M. Adsorbate-Induced
Surface Segregation for Core-Shell Nanocatalysts. Angew. Chemie - Int. Ed. 2009, 48,
3529–3531.
(157) Svenum, I.-H.; Herron, J. A.; Mavrikakis, M.; Venvik, H. J. Adsorbate-Induced Segregation
in a PdAg Membrane Model System: Pd3Ag(111). Catal. Today 2012, 193, 111–119.
(158) Herron, J. A.; Mavrikakis, M. On the Composition of Bimetallic Near-Surface Alloys in the
Presence of Oxygen and Carbon Monoxide. Catal. Commun. 2014, 52, 65-71.
(159) Liu, S.; Zhao, Z. J.; Yang, C.; Zha, S.; Neyman, K. M.; Studt, F.; Gong, J. Adsorption
Preference Determines Segregation Direction: A Shortcut to More Realistic Surface Models
of Alloy Catalysts. ACS Catal. 2019, 9, 5011–5018.
(160) Tao, F.; Grass, M. E.; Zhang, Y.; Butcher, D. R.; Renzas, J. R.; Liu, Z.; Chung, J. Y.; Mun,
B. S.; Salmeron, M.; Somorjai, G. A. Reaction-Driven Restructuring of Rh-Pd and Pt-Pd
Core-Shell Nanoparticles. Science 2008, 322, 932–934.
(161) Li, L.; Li, X.; Duan, Z.; Meyer, R. J.; Carr, R.; Raman, S.; Koziol, L.; Henkelman, G.
Adaptive Kinetic Monte Carlo Simulations of Surface Segregation in PdAu Nanoparticles.
Nanoscale 2019, 11, 10524–10535.
(162) Greeley, J.; Jaramillo, T. F.; Bonde, J.; Chorkendorff, I. B.; Nørskov, J. K. Computational
High-Throughput Screening of Electrocatalytic Materials for Hydrogen Evolution. Nat.
Mater. 2006, 5, 909–913.
(163) Li, Z.; Wang, S.; Chin, W. S.; Achenie, L. E.; Xin, H. High-Throughput Screening of
Bimetallic Catalysts Enabled by Machine Learning. J. Mater. Chem. A 2017, 5, 24131–
24138.
(164) Mamun, O.; Winther, K. T.; Boes, J. R.; Bligaard, T. High-Throughput Calculations of
Catalytic Properties of Bimetallic Alloy Surfaces. Sci. data 2019, 6, 76.
(165) Mavrikakis, M. Computational Methods: A Search Engine for Catalysts. Nat. Mater. 2006,

110
 5, 847–848.
(166) Khorshidi, A.; Violet, J.; Hashemi, J.; Peterson, A. A. How Strain Can Break the Scaling
Relations of Catalysis. Nat. Catal. 2018, 1, 263–268.
(167) Yan, K.; Maark, T. A.; Khorshidi, A.; Sethuraman, V. A.; Peterson, A. A.; Guduru, P. R.
The Influence of Elastic Strain on Catalytic Activity in the Hydrogen Evolution Reaction.
Angew. Chemie - Int. Ed. 2016, 55, 6175–6181.
(168) Mavrikakis, M.; Hammer, B.; Nørskov, J. Effect of Strain on the Reactivity of Metal
Surfaces. Phys. Rev. Lett. 1998, 81, 2819–2822.
(169) Luo, M.; Guo, S. Strain-Controlled Electrocatalysis on Multimetallic Nanomaterials. Nat.
Rev. Mater. 2017, 2, 1–14.
(170) Hwang, J.; Feng, Z.; Charles, N.; Wang, X. R.; Lee, D.; Stoerzinger, K. A.; Muy, S.; Rao,
R. R.; Lee, D.; Jacobs, R. et al. Tuning Perovskite Oxides by Strain: Electronic Structure,
Properties, and Functions in (Electro)Catalysis and Ferroelectricity. Mater. Today 2019, 31,
100–118.
(171) Xu, Y.; Mavrikakis, M. Adsorption and Dissociation of O2 on Gold Surfaces: Effect of
Steps and Strain. J. Phys. Chem. B 2003, 107, 9298–9307.
(172) Haruta, M.; Kobayashi, T.; Sano, H.; Yamada, N. Novel Gold Catalysts for the Oxidation
of Carbon Monoxide at a Temperature Far Below 0 °C. Chem. Lett. 1987, 16, 405–408.
(173) Haruta, M. Catalysis: Gold Rush. Nature 2005, 437, 1098–1099.
(174) Haruta, M. When Gold Is Not Noble: Catalysis by Nanoparticles. Chem. Rec. 2003, 3, 75–
87.
(175) Oliver-Meseguer, J.; Cabrero-Antonino, J. R.; Dominguez, I.; Leyva-Perez, A.; Corma, A.
Small Gold Clusters Formed in Solution Give Reaction Turnover Numbers of 107 at Room
Temperature. Science 2012, 338, 1452–1455.
(176) Du, Y.; Sheng, H.; Astruc, D.; Zhu, M. Atomically Precise Noble Metal Nanoclusters as
Efficient Catalysts: A Bridge between Structure and Properties. Chem. Rev. 2020, 120, 526–
622.
(177) Zhang, Z.; Zandkarimi, B.; Alexandrova, A. N. Ensembles of Metastable States Govern
Heterogeneous Catalysis on Dynamic Interfaces. Acc. Chem. Res. 2020, 53, 447–458.
(178) Li, Z.; Ji, S.; Liu, Y.; Cao, X.; Tian, S.; Chen, Y.; Niu, Z.; Li, Y. Well-Defined Materials
for Heterogeneous Catalysis: From Nanoparticles to Isolated Single-Atom Sites. Chem. Rev.
2020, 120, 623–682.
(179) Jimenez-Izal, E.; Alexandrova, A. N. Computational Design of Clusters for Catalysis. Annu.
Rev. Phys. Chem. 2018, 69, 377–400.
(180) Gates, B. C.; Flytzani-Stephanopoulos, M.; DIxon, D. A.; Katz, A. Atomically Dispersed
Supported Metal Catalysts: Perspectives and Suggestions for Future Research. Catal. Sci.
Technol. 2017, 7, 4259–4275.
(181) Kleis, J.; Greeley, J.; Romero, N. A.; Morozov, V. A.; Falsig, H.; Larsen, A. H.; Lu, J.;
Mortensen, J. J.; Dułak, M.; Thygesen, K. S. et al. Finite Size Effects in Chemical Bonding:
From Small Clusters to Solids. Catal. Lett. 2011, 141, 1067–1071.
(182) Li, L.; Larsen, A. H.; Romero, N. A.; Morozov, V. A.; Glinsvad, C.; Abild-Pedersen, F.;
Greeley, J.; Jacobsen, K. W.; Nørskov, J. K. Investigation of Catalytic Finite-Size-Effects
of Platinum Metal Clusters. J. Phys. Chem. Lett. 2013, 4, 222–226.
(183) Honkala, K.; Hellman, A.; Remediakis, I. N.; Logadottir, A.; Carlsson, A.; Dahl, S.;
Christensen, C. H.; Nørskov, J. K. Ammonia Synthesis from First-Principles Calculations.
Science 2005, 307, 555–558.

111
(184) Lopez, N. On the Origin of the Catalytic Activity of Gold Nanoparticles for Low-
Temperature CO Oxidation. J. Catal. 2004, 223, 232–235.
(185) Janssens, T. V. W.; Clausen, B. S.; Hvolbæk, B.; Falsig, H.; Christensen, C. H.; Bligaard,
T.; Nørskov, J. K. Insights into the Reactivity of Supported Au Nanoparticles: Combining
Theory and Experiments. Top. Catal. 2007, 44, 15–26.
(186) An, W.; Liu, P. Size and Shape Effects of Pd@Pt Core–Shell Nanoparticles: Unique Role
of Surface Contraction and Local Structural Flexibility. J. Phys. Chem. C 2013, 117, 16144–
16149.
(187) Yan, Z.; Chinta, S.; Mohamed, A.; Fackler, J. P.; Goodman, D. W. The Role of F-Centers
in Catalysis by Au Supported on MgO. J. Am. Chem. Soc. 2005, 127, 1604–1605.
(188) Yoon, B.; Häkkinen, H.; Landman, U.; Wörz, A. S.; Antonietti, J.-M.; Abbet, S.; Judai, K.;
Heiz, U. Charging Effects on Bonding and Catalyzed Oxidation of CO on Au8 Clusters on
MgO. Science 2005, 307, 403–407.
(189) Roduner, E. Catalysis by Metallic Nanoparticles. In Nanoscopic Materials: Size-Dependent
Phenomena; Royal Society of Chemistry: Cambridge, 2006; pp 239–262.
(190) Taylor, K. J.; Pettiette-Hall, C. L.; Cheshnovsky, O.; Smalley, R. E. Ultraviolet
Photoelectron Spectra of Coinage Metal Clusters. J. Chem. Phys. 1992, 96, 3319–3329.
(191) Bergeron, D. E. Al Cluster Superatoms as Halogens in Polyhalides and as Alkaline Earths
in Iodide Salts. Science 2005, 307, 231–235.
(192) Häkkinen, H. Electronic Shell Structures in Bare and Protected Metal Nanoclusters. Adv.
Phys. X 2016, 1, 467–491.
(193) Häkkinen, H. Atomic and Electronic Structure of Gold Clusters: Understanding Flakes,
Cages and Superatoms from Simple Concepts. Chem. Soc. Rev. 2008, 37, 1847–1859.
(194) Kaden, W. E.; Wu, T.; Kunkel, W. A.; Anderson, S. L. Electronic Structure Controls
Reactivity of Size-Selected Pd Clusters Adsorbed on TiO2 Surfaces. Science 2009, 326,
826–829.
(195) Li, S.; Singh, S.; Dumesic, J. A.; Mavrikakis, M. On the Nature of Active Sites for Formic
Acid Decomposition on Gold Catalysts. Catal. Sci. Technol. 2019, 9, 2836–2848.
(196) Xiang, Y.; Sun, D. Y.; Gong, X. G. Generalized Simulated Annealing Studies on Structures
and Properties of Nin (n = 2-55) Clusters. J. Phys. Chem. A 2000, 104, 2746–2751.
(197) Ouyang, R.; Xie, Y.; Jiang, D. E. Global Minimization of Gold Clusters by Combining
Neural Network Potentials and the Basin-Hopping Method. Nanoscale 2015, 7, 14817–
14821.
(198) Zhao, Y.; Chen, X.; Li, J. TGMin: A Global-Minimum Structure Search Program Based on
a Constrained Basin-Hopping Algorithm. Nano Res. 2017, 10, 3407–3420.
(199) Zhao, J.; Shi, R.; Sai, L.; Huang, X.; Su, Y. Comprehensive Genetic Algorithm for Ab Initio
Global Optimisation of Clusters. Mol. Simul. 2016, 42, 809–819.
(200) Jäger, M.; Schäfer, R.; Johnston, R. L. First Principles Global Optimization of Metal
Clusters and Nanoalloys. Adv. Phys. X 2018, 3, 1077–1108.
(201) Huang, R.; Bi, J.-X.; Li, L.; Wen, Y.-H. Basin Hopping Genetic Algorithm for Global
Optimization of PtCo Clusters. J. Chem. Inf. Model. 2020, 60, 2219–2228.
(202) Lykhach, Y.; Kozlov, S. M.; Skála, T.; Tovt, A.; Stetsovych, V.; Tsud, N.; Dvořák, F.;
Johánek, V.; Neitzel, A.; Mysliveček, J. et al. Counting Electrons on Supported
Nanoparticles. Nat. Mater. 2016, 15, 284–288.
(203) Campbell, C. T. The Energetics of Supported Metal Nanoparticles: Relationships to
Sintering Rates and Catalytic Activity. Acc. Chem. Res. 2013, 46, 1712–1719.

112
(204) Gao, M.; Lyalin, A.; Taketsugu, T. Catalytic Activity of Au and Au2 on the h-BN Surface:
Adsorption and Activation of O2. J. Phys. Chem. C 2012, 116, 9054–9062.
(205) Schimmenti, R.; Cortese, R.; Duca, D.; Mavrikakis, M. Boron Nitride-Supported Sub-
Nanometer Pd6 Clusters for Formic Acid Decomposition: A DFT Study. ChemCatChem
2017, 9, 1610–1620.
(206) Schimmenti, R.; Mavrikakis, M. HCOOH Decomposition on Sub-Nanometer Pd6 Cluster
Catalysts: The Effect of Defective Boron Nitride Supports Through First Principles. Appl.
Catal. B Environ. 2021, 280, 119392.
(207) Fujitani, T.; Nakamura, I.; Haruta, M. Role of Water in CO Oxidation on Gold Catalysts.
Catal. Letters 2014, 144, 1475–1486.
(208) Behrens, M.; Studt, F.; Kasatkin, I.; Kühl, S.; Hävecker, M.; Abild-Pedersen, F.; Zander,
S.; Girgsdies, F.; Kurr, P.; Kniep, B.-L. et al. The Active Site of Methanol Synthesis over
Cu/ZnO/Al2O3 Industrial Catalysts. Science 2012, 336, 893–897.
(209) Fu, Q. Active Nonmetallic Au and Pt Species on Ceria-Based Water-Gas Shift Catalysts.
Science 2003, 301, 935–938.
(210) Ghanekar, P.; Kubal, J.; Cui, Y.; Mitchell, G.; Delgass, W. N.; Ribeiro, F.; Greeley, J.
Catalysis at Metal/Oxide Interfaces: Density Functional Theory and Microkinetic Modeling
of Water Gas Shift at Pt/MgO Boundaries. Top. Catal. 2020, 63, 673–687.
(211) Saavedra, J.; Doan, H. A.; Pursell, C. J.; Grabow, L. C.; Chandler, B. D. The Critical Role
of Water at the Gold-Titania Interface in Catalytic CO Oxidation. Science 2014, 345, 1599–
1602.
(212) Green, I. X.; Tang, W.; Neurock, M.; Yates, J. T. Spectroscopic Observation of Dual
Catalytic Sites during Oxidation of CO on a Au/TiO₂ Catalyst. Science 2011, 333, 736–739.
(213) Song, W.; Hensen, E. J. M. A Computational DFT Study of CO Oxidation on a Au Nanorod
Supported on CeO2(110): On the Role of the Support Termination. Catal. Sci. Technol.
2013, 3, 3020–3029.
(214) Daelman, N.; Capdevila-Cortada, M.; López, N. Dynamic Charge and Oxidation State of
Pt/CeO2 Single-Atom Catalysts. Nat. Mater. 2019, 18, 1215–1221.
(215) Yang, S. C.; Pang, S. H.; Sulmonetti, T. P.; Su, W. N.; Lee, J. F.; Hwang, B. J.; Jones, C.
W. Synergy between Ceria Oxygen Vacancies and Cu Nanoparticles Facilitates the
Catalytic Conversion of CO2 to CO under Mild Conditions. ACS Catal. 2018, 8, 12056–
12066.
(216) Arena, F.; Mezzatesta, G.; Zafarana, G.; Trunfio, G.; Frusteri, F.; Spadaro, L. Effects of
Oxide Carriers on Surface Functionality and Process Performance of the Cu-ZnO System
in the Synthesis of Methanol via CO2 Hydrogenation. J. Catal. 2013, 300, 141–151.
(217) Tosoni, S.; Pacchioni, G. Oxide-Supported Gold Clusters and Nanoparticles in Catalysis: A
Computational Chemistry Perspective. ChemCatChem 2019, 11, 73–89.
(218) Sterrer, M.; Risse, T.; Heyde, M.; Rust, H. P.; Freund, H. J. Crossover from Three-
Dimensional to Two-Dimensional Geometries of Au Nanostructures on Thin MgO(001)
Films: A Confirmation of Theoretical Predictions. Phys. Rev. Lett. 2007, 98, 6–9.
(219) Shao, X.; Prada, S.; Giordano, L.; Pacchioni, G.; Nilius, N.; Freund, H. J. Tailoring the
Shape of Metal Ad-Particles by Doping the Oxide Support. Angew. Chemie - Int. Ed. 2011,
50, 11525–11527.
(220) Su, Y. Q.; Liu, J. X.; Filot, I. A. W.; Hensen, E. J. M. Theoretical Study of Ripening
Mechanisms of Pd Clusters on Ceria. Chem. Mater. 2017, 29, 9456–9462.
(221) Chaves, A. S.; Piotrowski, M. J.; Da Silva, J. L. F. Evolution of the Structural, Energetic,

113
 and Electronic Properties of the 3d, 4d, and 5d Transition-Metal Clusters (30 TMn Systems
for n = 2-15): A Density Functional Theory Investigation. Phys. Chem. Chem. Phys. 2017,
19, 15484–15502.
(222) Wang, Y.-G.; Mei, D.; Glezakou, V.-A.; Li, J.; Rousseau, R. Dynamic Formation of Single-
Atom Catalytic Active Sites on Ceria-Supported Gold Nanoparticles. Nat. Commun. 2015,
6, 6511.
(223) Baxter, E. T.; Ha, M. A.; Cass, A. C.; Alexandrova, A. N.; Anderson, S. L. Ethylene
Dehydrogenation on Pt4,7,8 Clusters on Al2O3: Strong Cluster Size Dependence Linked to
Preferred Catalyst Morphologies. ACS Catal. 2017, 7, 3322–3335.
(224) Saleheen, M.; Heyden, A. Liquid-Phase Modeling in Heterogeneous Catalysis. ACS Catal.
2018, 8, 2188–2194.
(225) Chew, A. K.; Walker, T. W.; Shen, Z.; Demir, B.; Witteman, L.; Euclide, J.; Huber, G. W.;
Dumesic, J. A.; Van Lehn, R. C. Effect of Mixed-Solvent Environments on the Selectivity
of Acid-Catalyzed Dehydration Reactions. ACS Catal. 2020, 10, 1679–1691.
(226) Li, G.; Wang, B.; Resasco, D. E. Water-Mediated Heterogeneously Catalyzed Reactions.
ACS Catal. 2020, 10, 1294–1309.
(227) Walker, T. W.; Chew, A. K.; Van Lehn, R. C.; Dumesic, J. A.; Huber, G. W. Rational
Design of Mixed Solvent Systems for Acid-Catalyzed Biomass Conversion Processes Using
a Combined Experimental, Molecular Dynamics and Machine Learning Approach. Top.
Catal. 2020, 63, 649–663.
(228) Lim, H.; Jung, Y. J. Delfos: Deep Learning Model for Prediction of Solvation Free Energies
in Generic Organic Solvents. Chem. Sci. 2019, 10, 8306–8315.
(229) Basdogan, Y.; Groenenboom, M. C.; Henderson, E.; De, S.; Rempe, S. B.; Keith, J. A.
Machine Learning-Guided Approach for Studying Solvation Environments. J. Chem.
Theory Comput. 2020, 16, 633–642.
(230) Pliego, J. R.; Riveros, J. M. Hybrid Discrete-Continuum Solvation Methods. Wiley
Interdiscip. Rev. Comput. Mol. Sci. 2020, 10, 1–25.
(231) Tomasi, J.; Mennucci, B.; Cammi, R. Quantum Mechanical Continuum Solvation Models.
Chem. Rev. 2005, 105, 2999–3093.
(232) Lipparini, F.; Mennucci, B. Perspective: Polarizable Continuum Models for Quantum-
Mechanical Descriptions. J. Chem. Phys. 2016, 144, 160901.
(233) Mennucci, B. Polarizable Continuum Model. Wiley Interdiscip. Rev. Comput. Mol. Sci.
2012, 2, 386–404.
(234) Scalmani, G.; Frisch, M. J. Continuous Surface Charge Polarizable Continuum Models of
Solvation. I. General Formalism. J. Chem. Phys. 2010, 132, 114110.
(235) Andreussi, O.; Marzari, N. Electrostatics of Solvated Systems in Periodic Boundary
Conditions. Phys. Rev. B 2014, 90, 245101.
(236) Andreussi, O.; Dabo, I.; Marzari, N. Revised Self-Consistent Continuum Solvation in
Electronic-Structure Calculations. J. Chem. Phys. 2012, 136, 064102.
(237) Mathew, K.; Kolluru, V. S. C.; Mula, S.; Steinmann, S. N.; Hennig, R. G. Implicit Self-
Consistent Electrolyte Model in Plane-Wave Density-Functional Theory. J. Chem. Phys.
2019, 151, 234101.
(238) Mathew, K.; Sundararaman, R.; Letchworth-Weaver, K.; Arias, T. A.; Hennig, R. G.
Implicit Solvation Model for Density-Functional Study of Nanocrystal Surfaces and
Reaction Pathways. J. Chem. Phys. 2014, 140, 084106.
(239) Sundararaman, R.; Arias, T. A. Efficient Classical Density-Functional Theories of Rigid-

114
 Molecular Fluids and a Simplified Free Energy Functional for Liquid Water. Comput. Phys.
Commun. 2014, 185, 818–825.
(240) Xu, H.; Cheng, D.; Cao, D.; Zeng, X. C. A Universal Principle for a Rational Design of
Single-Atom Electrocatalysts. Nat. Catal. 2018, 1, 339–348.
(241) Goodpaster, J. D.; Bell, A. T.; Head-Gordon, M. Identification of Possible Pathways for C–
C Bond Formation during Electrochemical Reduction of CO2 : New Theoretical Insights
from an Improved Electrochemical Model. J. Phys. Chem. Lett. 2016, 7, 1471–1477.
(242) Garza, A. J.; Bell, A. T.; Head-Gordon, M. Mechanism of CO2 Reduction at Copper
Surfaces: Pathways to C2 Products. ACS Catal. 2018, 8, 1490–1499.
(243) Sundararaman, R.; Schwarz, K. Evaluating Continuum Solvation Models for the Electrode-
Electrolyte Interface: Challenges and Strategies for Improvement. J. Chem. Phys. 2017, 146,
084111.
(244) Sundararaman, R.; Goddard, W. A. The Charge-Asymmetric Nonlocally Determined Local-
Electric (CANDLE) Solvation Model. J. Chem. Phys. 2015, 142, 064107.
(245) Yu, F.; Zhou, H.; Huang, Y.; Sun, J.; Qin, F.; Bao, J.; Goddard, W. A.; Chen, S.; Ren, Z.
High-Performance Bifunctional Porous Non-Noble Metal Phosphide Catalyst for Overall
Water Splitting. Nat. Commun. 2018, 9, 2551.
(246) Kovalenko, A.; Hirata, F. Self-Consistent Description of a Metal-Water Interface by the
Kohn-Sham Density Functional Theory and the Three-Dimensional Reference Interaction
Site Model. J. Chem. Phys. 1999, 110, 10095–10112.
(247) Chandler, D.; Andersen, H. C. Optimized Cluster Expansions for Classical Fluids. II.
Theory of Molecular Liquids. J. Chem. Phys. 1972, 57, 1918–1929.
(248) Fujita, T.; Yamamoto, T. Assessing the Accuracy of Integral Equation Theories for Nano-
Sized Hydrophobic Solutes in Water. J. Chem. Phys. 2017, 147, 014110.
(249) Howard, J. J.; Perkyns, J. S.; Choudhury, N.; Pettitt, B. M. Integral Equation Study of the
Hydrophobic Interaction between Graphene Plates. J. Chem. Theory Comput. 2008, 4,
1928–1939.
(250) Truchon, J. F.; Pettitt, B. M.; Labute, P. A Cavity Corrected 3D-RISM Functional for
Accurate Solvation Free Energies. J. Chem. Theory Comput. 2014, 10, 934–941.
(251) Beveridge, D. L.; DiCapua, F. M. Free Energy via Molecular Simulation: Applications to
Chemical and Biomolecular Systems. Annu. Rev. Biophys. Biophys. Chem. 1989, 18, 431–
492.
(252) Fogolari, F.; Corazza, A.; Esposito, G. Free Energy, Enthalpy and Entropy from Implicit
Solvent End-Point Simulations. Front. Mol. Biosci. 2018, 5, 11.
(253) Stoltz, G. Computation of Free Energy Differences. Encycl. Appl. Comput. Math. 2015, 85,
254–260.
(254) Laio, A.; Parrinello, M. Escaping Free-Energy Minima. Proc. Natl. Acad. Sci. 2002, 99,
12562–12566.
(255) Laio, A.; Gervasio, F. L. Metadynamics: A Method to Simulate Rare Events and
Reconstruct the Free Energy in Biophysics, Chemistry and Material Science. Reports Prog.
Phys. 2008, 71, 126601.
(256) Gono, P.; Ambrosio, F.; Pasquarello, A. Effect of the Solvent on the Oxygen Evolution
Reaction at the TiO2-Water Interface. J. Phys. Chem. C 2019, 123, 18467–18474.
(257) Herron, J. A.; Morikawa, Y.; Mavrikakis, M. Ab Initio Molecular Dynamics of Solvation
Effects on Reactivity at Electrified Interfaces. Proc. Natl. Acad. Sci. 2016, 113, E4937–
E4945.

115
(258) Ludwig, T.; Gauthier, J. A.; Brown, K. S.; Ringe, S.; Nørskov, J. K.; Chan, K. Solvent-
Adsorbate Interactions and Adsorbate-Specific Solvent Structure in Carbon Dioxide
Reduction on a Stepped Cu Surface. J. Phys. Chem. C 2019, 123, 5999–6009.
(259) Heenen, H. H.; Gauthier, J. A.; Kristoffersen, H. H.; Ludwig, T.; Chan, K. Solvation at
Metal/Water Interfaces: An Ab Initio Molecular Dynamics Benchmark of Common
Computational Approaches. J. Chem. Phys. 2020, 152, 144703.
(260) Kristoffersen, H. H.; Chang, J. H. Effect of Competitive Adsorption at the Interface between
Aqueous Electrolyte and Solid Electrode. In Computational Photocatalysis: Modeling of
Photophysics and Photochemistry at Interfaces ACS; 2019; pp 225–238.
(261) Calle-Vallejo, F.; De Morais, R. F.; Illas, F.; Loffreda, D.; Sautet, P. Affordable Estimation
of Solvation Contributions to the Adsorption Energies of Oxygenates on Metal
Nanoparticles. J. Phys. Chem. C 2019, 123, 5578–5582.
(262) Garcia-Ratés, M.; García-Muelas, R.; López, N. Solvation Effects on Methanol
Decomposition on Pd(111), Pt(111), and Ru(0001). J. Phys. Chem. C 2017, 121, 13803–
13809.
(263) Sicinska, D.; Paneth, P.; Truhlar, D. G. How Well Does Microsolvation Represent
Macrosolvation? A Test Case: Dynamics of Decarboxylation of 4-Pyridylacetic Acid
Zwitterion. J. Phys. Chem. B 2002, 106, 2708–2713.
(264) Simm, G. N.; Türtscher, P. L.; Reiher, M. Systematic Microsolvation Approach with a
Cluster-Continuum Scheme and Conformational Sampling. J. Comput. Chem. 2020, 41,
1144–1155.
(265) Dumesic, J. A.; Rudd, D. F.; Aparicio, L. M.; Rekoske, J. E.; Treviño, A. A. The
Microkinetics of Heterogeneous Catalysis; American Chemical Society: Washington DC,
1993.
(266) Gokhale, A.; Kandoi, S.; Greeley, J. Molecular-Level Descriptions of Surface Chemistry in
Kinetic Models Using Density Functional Theory. Chem. Eng. Sci. 2004, 59, 4679–4691.
(267) Ovesen, C. V.; Clausen, B. S.; Hammershøi, B. S.; Steffensen, G.; Askgaard, T.;
Chorkendorff, I.; Nørskov, J. K.; Rasmussen, P. B.; Stoltze, P.; Taylor, P. A Microkinetic
Analysis of the Water–Gas Shift Reaction under Industrial Conditions. J. Catal. 1996, 158,
170–180.
(268) Stoltze, P.; Nørskov, J. K. The Surface Science Based Ammonia Kinetics Revisited. Top.
Catal. 1994, 1, 253–263.
(269) Nakao, T.; Tada, T.; Hosono, H. First-Principles and Microkinetic Study on the Mechanism
for Ammonia Synthesis Using Ru-Loaded Hydride Catalyst. J. Phys. Chem. C 2020, 124,
2070–2078.
(270) Singh, A. R.; Montoya, J. H.; Rohr, B. A.; Tsai, C.; Vojvodic, A.; Nørskov, J. K.
Computational Design of Active Site Structures with Improved Transition-State Scaling for
Ammonia Synthesis. ACS Catal. 2018, 8, 4017–4024.
(271) Xu, D.; Wu, P.; Yang, B. Essential Role of Water in the Autocatalysis Behavior of Methanol
Synthesis from CO2 Hydrogenation on Cu: A Combined DFT and Microkinetic Modeling
Study. J. Phys. Chem. C 2019, 123, 8959–8966.
(272) Maulana, A. L.; Putra, R. I. D.; Saputro, A. G.; Agusta, M. K.; Nugraha, N.; Dipojono, H.
K. DFT and Microkinetic Investigation of Methanol Synthesis via CO2 Hydrogenation on
Ni(111)-Based Surfaces. Phys. Chem. Chem. Phys. 2019, 21, 20276–20286.
(273) Janse van Rensburg, W.; Petersen, M. A.; Datt, M. S.; van den Berg, J.-A.; van Helden, P.
On the Kinetic Interpretation of DFT-Derived Energy Profiles: Cu-Catalyzed Methanol

116
 Synthesis. Catal. Lett. 2015, 145, 559–568.
(274) Tang, Q.-L.; Zou, W.-T.; Huang, R.-K.; Wang, Q.; Duan, X.-X. Effect of the Components’
Interface on the Synthesis of Methanol over Cu/ZnO from CO2/H2 : A Microkinetic
Analysis Based on DFT + U Calculations. Phys. Chem. Chem. Phys. 2015, 17, 7317–7333.
(275) Ye, J.; Liu, C.; Mei, D.; Ge, Q. Methanol Synthesis from CO2 Hydrogenation over a
Pd4/In2O3 Model Catalyst: A Combined DFT and Kinetic Study. J. Catal. 2014, 317, 44–
53.
(276) Grabow, L. C.; Mavrikakis, M. Mechanism of Methanol Synthesis on Cu through CO2 and
CO Hydrogenation. ACS Catal. 2011, 1, 365–384.
(277) van Helden, P.; Berg, J.-A. van den; Petersen, M. A.; Janse van Rensburg, W.; Ciobîcă, I.
M.; van de Loosdrecht, J. Computational Investigation of the Kinetics and Mechanism of
the Initial Steps of the Fischer–Tropsch Synthesis on Cobalt. Faraday Discuss. 2017, 197,
117–151.
(278) Filot, I. A. W.; van Santen, R. A.; Hensen, E. J. M. The Optimally Performing Fischer-
Tropsch Catalyst. Angew. Chemie Int. Ed. 2014, 53, 12746–12750.
(279) Mirwald, J. W.; Inderwildi, O. R. Unraveling the Fischer–Tropsch Mechanism: A
Combined DFT and Microkinetic Investigation of C–C Bond Formation on Ru. Phys. Chem.
Chem. Phys. 2012, 14, 7028–7031.
(280) Campbell, C. T. Future Directions and Industrial Perspectives Micro- and Macro-Kinetics:
Their Relationship in Heterogeneous Catalysis. Top. Catal. 1994, 1, 353–366.
(281) Salciccioli, M.; Stamatakis, M.; Caratzoulas, S.; Vlachos, D. G. A Review of Multiscale
Modeling of Metal-Catalyzed Reactions: Mechanism Development for Complexity and
Emergent Behavior. Chem. Eng. Sci. 2011, 66, 4319–4355.
(282) Gokhale, A. A.; Kandoi, S.; Greeley, J. P.; Mavrikakis, M.; Dumesic, J. A. Molecular-Level
Descriptions of Surface Chemistry in Kinetic Models Using Density Functional Theory.
Chem. Eng. Sci. 2004, 59, 4679–4691.
(283) Chen, B. W. J.; Mavrikakis, M. How Coverage Influences Thermodynamic and Kinetic
Isotope Effects for H2/D2 Dissociative Adsorption on Transition Metals. Catal. Sci. Technol.
2020, 10, 671–689.
(284) García-Diéguez, M.; Hibbitts, D. D.; Iglesia, E. Hydrogen Chemisorption Isotherms on
Platinum Particles at Catalytic Temperatures: Langmuir and Two-Dimensional Gas Models
Revisited. J. Phys. Chem. C 2019, 123, 8447–8462.
(285) Collinge, G.; Yuk, S. F.; Nguyen, M.-T.; Lee, M.-S.; Glezakou, V.-A.; Rousseau, R. Effect
of Collective Dynamics and Anharmonicity on Entropy in Heterogenous Catalysis:
Building the Case for Advanced Molecular Simulations. ACS Catal. 2020, 10, 9236–9260.
(286) De Wispelaere, K.; Vanduyfhuys, L.; Van Speybroeck, V. Entropy Contributions to
Transition State Modeling. In Modelling and Simulation in the Science of Micro- and Meso-
Porous Materials; Elsevier, 2018; pp 189–228.
(287) Campbell, C. T.; Sellers, J. R. V. The Entropies of Adsorbed Molecules. J. Am. Chem. Soc.
2012, 134, 18109–18115.
(288) Grimme, S. Supramolecular Binding Thermodynamics by Dispersion-Corrected Density
Functional Theory. Chem. - A Eur. J. 2012, 18, 9955–9964.
(289) Sprowl, L. H.; Campbell, C. T.; Árnadóttir, L. Hindered Translator and Hindered Rotor
Models for Adsorbates: Partition Functions and Entropies. J. Phys. Chem. C 2016, 120,
9719–9731.
(290) Piccini, G.; Alessio, M.; Sauer, J. Ab Initio Calculation of Rate Constants for Molecule-

117
 Surface Reactions with Chemical Accuracy. Angew. Chemie Int. Ed. 2016, 55, 5235–5237.
(291) Bajpai, A.; Mehta, P.; Frey, K.; Lehmer, A. M.; Schneider, W. F. Benchmark First-
Principles Calculations of Adsorbate Free Energies. ACS Catal. 2018, 8, 1945–1954.
(292) Li, H.; Paolucci, C.; Schneider, W. F. Zeolite Adsorption Free Energies from Ab Initio
Potentials of Mean Force. J. Chem. Theory Comput. 2018, 14, 929–938.
(293) Paolucci, C.; Di Iorio, J. R.; Schneider, W. F.; Gounder, R. Solvation and Mobilization of
Copper Active Sites in Zeolites by Ammonia: Consequences for the Catalytic Reduction of
Nitrogen Oxides. Acc. Chem. Res. 2020, 53, 1881–1892.
(294) Alexopoulos, K.; Lee, M.-S.; Liu, Y.; Zhi, Y.; Liu, Y.; Reyniers, M.-F.; Marin, G. B.;
Glezakou, V.-A.; Rousseau, R.; Lercher, J. A. Anharmonicity and Confinement in Zeolites:
Structure, Spectroscopy, and Adsorption Free Energy of Ethanol in H-ZSM-5. J. Phys.
Chem. C 2016, 120, 7172–7182.
(295) Xu, L.; Stangland, E. E.; Mavrikakis, M. A DFT Study of Chlorine Coverage over Late
Transition Metals and Its Implication on 1,2-Dichloroethane Hydrodechlorination. Catal.
Sci. Technol. 2018, 8, 1555–1563.
(296) Ojeda, M.; Nabar, R.; Nilekar, A. U.; Ishikawa, A.; Mavrikakis, M.; Iglesia, E. CO
Activation Pathways and the Mechanism of Fischer-Tropsch Synthesis. J. Catal. 2010, 272,
287–297.
(297) Bhandari, S.; Rangarajan, S.; Mavrikakis, M. Combining Computational Modeling with
Reaction Kinetics Experiments for Elucidating the In Situ Nature of the Active Site in
Catalysis. Acc. Chem. Res. 2020, 53, 1893–1904.
(298) Grabow, L. C.; Gokhale, A. A.; Evans, S. T.; Dumesic, J. A.; Mavrikakis, M. Mechanism
of the Water Gas Shift Reaction on Pt: First Principles, Experiments, and Microkinetic
Modeling. J. Phys. Chem. C 2008, 112, 4608–4617.
(299) Getman, R. B.; Schneider, W. F. DFT-Based Coverage-Dependent Model of Pt-Catalyzed
NO Oxidation. ChemCatChem 2010, 2, 1450–1460.
(300) Singh, S.; Li, S.; Carrasquillo-Flores, R.; Alba-Rubio, A. C.; Dumesic, J. A.; Mavrikakis,
M. Formic Acid Decomposition on Au Catalysts: DFT, Microkinetic Modeling, and
Reaction Kinetics Experiments. AIChE J. 2014, 60, 1303–1319.
(301) Rubert-Nason, P.; Mavrikakis, M.; Maravelias, C. T.; Grabow, L. C.; Biegler, L. T.
Advanced Solution Methods for Microkinetic Models of Catalytic Reactions: A Methanol
Synthesis Case Study. AIChE J. 2014, 60, 1336–1346.
(302) Rangarajan, S.; Maravelias, C. T.; Mavrikakis, M. Sequential-Optimization-Based
Framework for Robust Modeling and Design of Heterogeneous Catalytic Systems. J. Phys.
Chem. C 2017, 121, 25847–25863.
(303) Stein, M. Large Sample Properties of Simulations Using Latin Hypercube Sampling.
Technometrics 1987, 29, 143–151.
(304) McKay, M. D.; Beckman, R. J.; Conover, W. J. A Comparison of Three Methods for
Selecting Values of Input Variables in the Analysis of Output from a Computer Code.
Technometrics 1979, 21, 239–245.
(305) Bhandari, S.; Rangarajan, S.; Maravelias, C. T.; Dumesic, J. A.; Mavrikakis, M. Reaction
Mechanism of Vapor-Phase Formic Acid Decomposition over Platinum Catalysts: DFT,
Reaction Kinetics Experiments, and Microkinetic Modeling. ACS Catal. 2020, 10, 4112–
4126.
(306) Norton, P. R.; Davies, J. A.; Jackman, T. E. Absolute Coverages of CO and O on Pt(111);
Comparison of Saturation CO Coverages on Pt(100), (110) and (111) Surfaces. Surf. Sci.

118
 1982, 122, L593–L600.
(307) Ovesen, C. V.; Stoltze, P.; Nørskov, J. K.; Campbell, C. T. A Kinetic Model of the Water
Gas Shift Reaction. J. Catal. 1992, 134, 445–468.
(308) Nakamura, J.; Campbell, J. M.; Campbell, C. T. Kinetics and Mechanism of the Water-Gas
Shift Reaction Catalysed by the Clean and Cs-Promoted Cu(110) Surface: A Comparison
with Cu(111). J. Chem. Soc. Faraday Trans. 1990, 86, 2725–2734.
(309) Koryabkina, N. A.; Phatak, A. A.; Ruettinger, W. F.; Farrauto, R. J.; Ribeiro, F. H.
Determination of Kinetic Parameters for the Water-Gas Shift Reaction on Copper Catalysts
under Realistic Conditions for Fuel Cell Applications. J. Catal. 2003, 217, 233–239.
(310) Campbell, C. T.; Daube, K. A. A Surface Science Investigation of the Water-Gas Shift
Reaction on Cu(111). J. Catal. 1987, 104, 109–119.
(311) Campbell, C. T. Surface Science Studies of the Water–Gas Shift Reaction on a Model
Cu(111) Catalyst. J. Vac. Sci. Technol. A Vacuum, Surfaces, Film. 1987, 5, 810–813.
(312) van Herwijnen, T.; de Jong, W. A. Kinetics and Mechanism of the CO Shift on Cu ZnO. 1.
Kinetics of the Forward and Reverse CO Shift Reactions. J. Catal. 1980, 63, 83–93.
(313) Crapper, M. D.; Riley, C. E.; Woodruff, D. P.; Puschmann, A.; Haase, J. Determination of
the Adsorption Structure for Formate on Cu(110) Using SEXAFS and NEXAFS. Surf. Sci.
1986, 171, 1–12.
(314) Fujitani, T.; Choi, Y.; Sano, M.; Kushida, Y.; Nakamura, J. Scanning Tunneling
Microscopy Study of Formate Species Synthesized from CO2 Hydrogenation and Prepared
by Adsorption of Formic Acid over Cu(111). J. Phys. Chem. B 2000, 104, 1235–1240.
(315) Gokhale, A. A.; Dumesic, J. A.; Mavrikakis, M. On the Mechanism of Low-Temperature
Water Gas Shift Reaction on Copper. J. Am. Chem. Soc. 2008, 130, 1402–1414.
(316) Madon, R. J.; Braden, D.; Kandoi, S.; Nagel, P.; Mavrikakis, M.; Dumesic, J. A.
Microkinetic Analysis and Mechanism of the Water Gas Shift Reaction over Copper
Catalysts. J. Catal. 2011, 281, 1–11.
(317) Clay, J. P.; Greeley, J. P.; Ribeiro, F. H.; Nicholas Delgass, W.; Schneider, W. F. DFT
Comparison of Intrinsic WGS Kinetics over Pd and Pt. J. Catal. 2014, 320, 106–117.
(318) Carrasquillo-Flores, R.; Gallo, J. M. R.; Hahn, K.; Dumesic, J. A.; Mavrikakis, M. Density
Functional Theory and Reaction Kinetics Studies of the Water-Gas Shift Reaction on Pt-Re
Catalysts. ChemCatChem 2013, 5, 3690–3699.
(319) Grasemann, M.; Laurenczy, G. Formic Acid as a Hydrogen Source – Recent Developments
and Future Trends. Energy Environ. Sci. 2012, 5, 8171–8181.
(320) Yi, N.; Saltsburg, H.; Flytzani-Stephanopoulos, M. Hydrogen Production by
Dehydrogenation of Formic Acid on Atomically Dispersed Gold on Ceria. ChemSusChem
2013, 6, 816–819.
(321) Bulushev, D. A.; Beloshapkin, S.; Ross, J. R. H. Hydrogen from Formic Acid
Decomposition over Pd and Au Catalysts. Catal. Today 2010, 154, 7–12.
(322) Bi, Q. Y.; Lin, J. D.; Liu, Y. M.; He, H. Y.; Huang, F. Q.; Cao, Y. Gold Supported on
Zirconia Polymorphs for Hydrogen Generation from Formic Acid in Base-Free Aqueous
Medium. J. Power Sources 2016, 328, 463–471.
(323) Zacharska, M.; Chuvilin, A. L.; Kriventsov, V. V.; Beloshapkin, S.; Estrada, M.; Simakov,
A.; Bulushev, D. A. Support Effect for Nanosized Au Catalysts in Hydrogen Production
from Formic Acid Decomposition. Catal. Sci. Technol. 2016, 6, 6853–6860.
(324) Bi, Q.-Y.; Du, X.-L.; Liu, Y.-M.; Cao, Y.; He, H.-Y.; Fan, K.-N. Efficient Subnanometric
Gold-Catalyzed Hydrogen Generation via Formic Acid Decomposition under Ambient

119
 Conditions. J. Am. Chem. Soc. 2012, 134, 8926–8933.
(325) Ojeda, M.; Iglesia, E. Formic Acid Dehydrogenation on Au-Based Catalysts at near-
Ambient Temperatures. Angew. Chemie Int. Ed. 2009, 48, 4800–4803.
(326) Stamatakis, M.; Vlachos, D. G. Unraveling the Complexity of Catalytic Reactions via
Kinetic Monte Carlo Simulation: Current Status and Frontiers. ACS Catal. 2012, 2, 2648–
2663.
(327) Voter, A. F. Introduction to the Kinetic Monte Carlo Method. In Radiation Effects in Solids;
Sickafus, K. E., Kotomin, E. A., Uberuaga, B. P., Eds.; Springer: Netherlands, 2007; pp 1–
23.
(328) Chatterjee, A.; Vlachos, D. G. An Overview of Spatial Microscopic and Accelerated Kinetic
Monte Carlo Methods. J. Comput. Mater. Des. 2007, 14, 253–308.
(329) Schulze, T. P. Efficient Kinetic Monte Carlo Simulation. J. Comput. Phys. 2008, 227, 2455–
2462.
(330) Stamatakis, M.; Vlachos, D. G. A Graph-Theoretical Kinetic Monte Carlo Framework for
on-Lattice Chemical Kinetics. J. Chem. Phys. 2011, 134, 214115.
(331) Nielsen, J.; D’Avezac, M.; Hetherington, J.; Stamatakis, M. Parallel Kinetic Monte Carlo
Simulation Framework Incorporating Accurate Models of Adsorbate Lateral Interactions. J.
Chem. Phys. 2013, 139, 224706.
(332) Hoffmann, M. J.; Matera, S.; Reuter, K. Kmos: A Lattice Kinetic Monte Carlo Framework.
Comput. Phys. Commun. 2014, 185, 2138–2150.
(333) Jørgensen, M.; Grönbeck, H. MonteCoffee: A Programmable Kinetic Monte Carlo
Framework. J. Chem. Phys. 2018, 149, 114101.
(334) Hjorth Larsen, A.; Jørgen Mortensen, J.; Blomqvist, J.; Castelli, I. E.; Christensen, R.;
Dułak, M.; Friis, J.; Groves, M. N.; Hammer, B.; Hargus, C. et al. The Atomic Simulation
Environment—a Python Library for Working with Atoms. J. Phys. Condens. Matter 2017,
29, 273002.
(335) Stamatakis, M.; Chen, Y.; Vlachos, D. G. First-Principles-Based Kinetic Monte Carlo
Simulation of the Structure Sensitivity of the Water À Gas Shift Reaction on Platinum
Surfaces. J. Phys. Chem. C 2011, 115, 24750–24762.
(336) Sendner, C.; Sakong, S.; Groß, A. Kinetic Monte Carlo Simulations of the Partial Oxidation
of Methanol on Oxygen-Covered Cu(110). Surf. Sci. 2006, 600, 3258–3265.
(337) Hess, F.; Over, H. Rate-Determining Step or Rate-Determining Configuration? The Deacon
Reaction over RuO2(110) Studied by DFT-Based KMC Simulations. ACS Catal. 2017, 7,
128–138.
(338) Silverberg, M.; Ben-Shaul, A. Adsorbate Islanding in Surface Reactions: A Combined
Monte Carlo-Lattice Gas Approach. J. Chem. Phys. 1987, 87, 3178–3194.
(339) Chen, B. W. J.; Stamatakis, M.; Mavrikakis, M. Kinetic Isolation between Turnovers on Au
18 Nanoclusters: Formic Acid Decomposition One Molecule at a Time. ACS Catal. 2019,
9, 9446–9457.
(340) Hess, F.; Krause, P. P. T.; Rohrlack, S. F.; Hofmann, J. P.; Farkas, A.; Over, H. One-
Dimensional Confinement in Heterogeneous Catalysis: Trapped Oxygen on RuO2(110)
Model Catalysts. Surf. Sci. 2012, 606, L69–L73.
(341) Sanchez, J. M.; De Fontaine, D. Theoretical Prediction of Ordered Superstructures in
Metallic Alloys. In Structure and Bonding in Crystals; Navrotsky, A., O’Keeffe, M., Eds.;
Academic Press, Inc: New York, 1981; Vol. 2, pp 117–132.
(342) Hansen, E. W.; Neurock, M. First-Principles-Based Monte Carlo Simulation of Ethylene

120
 Hydrogenation Kinetics on Pd. J. Catal. 2000, 196, 241–252.
(343) Hess, F.; Over, H. Kinetic Monte Carlo Simulations of Heterogeneously Catalyzed
Oxidation Reactions. Catal. Sci. Technol. 2014, 4, 583–598.
(344) Bray, J. M.; Schneider, W. F. First-Principles Analysis of Structure Sensitivity in NO
Oxidation on Pt. ACS Catal. 2015, 5, 1087–1099.
(345) Wu, C.; Schmidt, D. J.; Wolverton, C.; Schneider, W. F. Accurate Coverage-Dependence
Incorporated into First-Principles Kinetic Models: Catalytic NO Oxidation on Pt (1 1 1). J.
Catal. 2012, 286, 88–94.
(346) Ångqvist, M.; Muñoz, W. A.; Rahm, J. M.; Fransson, E.; Durniak, C.; Rozyczko, P.; Rod,
T. H.; Erhart, P. ICET – A Python Library for Constructing and Sampling Alloy Cluster
Expansions. Adv. Theory Simulations 2019, 2, 1900015.
(347) Troppenz, M.; Rigamonti, S.; Draxl, C. Predicting Ground-State Configurations and
Electronic Properties of the Thermoelectric Clathrates Ba8AlxSi46-x and Sr8AlxSi46-X. Chem.
Mater. 2017, 29, 2414–2424.
(348) Vignola, E.; Steinmann, S. N.; Vandegehuchte, B. D.; Curulla, D.; Stamatakis, M.; Sautet,
P. A Machine Learning Approach to Graph-Theoretical Cluster Expansions of the Energy
of Adsorbate Layers. J. Chem. Phys. 2017, 147, 054106.
(349) Hess, F. Efficient Implementation of Cluster Expansion Models in Surface Kinetic Monte
Carlo Simulations with Lateral Interactions: Subtraction Schemes, Supersites, and the
Supercluster Contraction. J. Comput. Chem. 2019, 40, 2664–2676.
(350) Jørgensen, M.; Grönbeck, H. The Site-Assembly Determines Catalytic Activity of
Nanoparticles. Angew. Chemie - Int. Ed. 2018, 57, 5086–5089.
(351) Nikbin, N.; Mpourmpakis, G.; Vlachos, D. G. A Combined DFT and Statistical Mechanics
Study for the CO Oxidation on the Au10–1 Cluster. J. Phys. Chem. C 2011, 115, 20192–
20200.
(352) Nikbin, N.; Austin, N.; Vlachos, D. G.; Stamatakis, M.; Mpourmpakis, G. Catalysis at the
Sub-Nanoscale: Complex CO Oxidation Chemistry on a Few Au Atoms. Catal. Sci. Technol.
2015, 5, 134–141.
(353) Hansen, P. L.; Wagner, J. B.; Helveg, S.; Rostrup-Nielsen, J. R.; Clausen, B. S.; Topsøe, H.
Atom-Resolved Imaging of Dynamic Shape Changes in Supported Copper Nanocrystals.
Science (80-. ). 2002, 295, 2053–2055.
(354) Noussiou, V. K.; Provata, A. Surface Reconstruction in Reactive Dynamics: A Kinetic
Monte Carlo Approach. Surf. Sci. 2007, 601, 2941–2951.
(355) Hoffmann, M. J.; Scheffler, M.; Reuter, K. Multi-Lattice Kinetic Monte Carlo Simulations
from First Principles: Reduction of the Pd(100) Surface Oxide by CO. ACS Catal. 2015, 5,
1199–1209.
(356) Papanikolaou, K. G.; Darby, M. T.; Stamatakis, M. Engineering the Surface Architecture of
Highly Dilute Alloys: An Ab Initio Monte Carlo Approach. ACS Catal. 2020, 10, 1224–
1236.
(357) Xia, Y.; Xiong, Y.; Lim, B.; Skrabalak, S. E. Shape-Controlled Synthesis of Metal
Nanocrystals: Simple Chemistry Meets Complex Physics? Angew. Chemie - Int. Ed. 2009,
48, 60–103.
(358) Mostafa, S.; Behafarid, F.; Croy, J. R.; Ono, L. K.; Li, L.; Yang, J. C.; Frenkel, A. I.; Cuenya,
B. R. Shape-Dependent Catalytic Properties of Pt Nanoparticles. J. Am. Chem. Soc. 2010,
132, 15714–15719.
(359) Cao, S.; Tao, F.; Tang, Y.; Li, Y.; Yu, J. Size- and Shape-Dependent Catalytic Performances

121
 of Oxidation and Reduction Reactions on Nanocatalysts. Chem. Soc. Rev. 2016, 45, 4747–
4765.
(360) Zhang, L.; Roling, L. T.; Wang, X.; Vara, M.; Chi, M.; Liu, J.; Choi, S.-I.; Park, J.; Herron,
J. A.; Xie, Z.; Mavrikakis, M.; Xia, Y. Platinum-Based Nanocages with Subnanometer-
Thick Walls and Well-Defined, Controllable Facets. Science 2015, 349, 412–416.
(361) Barmparis, G. D.; Lodziana, Z.; Lopez, N.; Remediakis, I. N. Nanoparticle Shapes by Using
Wulff Constructions and First-Principles Calculations. Beilstein J. Nanotechnol. 2015, 6,
361–368.
(362) Lai, K. C.; Han, Y.; Spurgeon, P.; Huang, W.; Thiel, P. A.; Liu, D. J.; Evans, J. W.
Reshaping, Intermixing, and Coarsening for Metallic Nanocrystals: Nonequilibrium
Statistical Mechanical and Coarse-Grained Modeling. Chem. Rev. 2019, 119, 6670–6768.
(363) Lai, K. C.; Evans, J. W. Reshaping and Sintering of 3D Fcc Metal Nanoclusters: Stochastic
Atomistic Modeling with Realistic Surface Diffusion Kinetics. Phys. Rev. Mater. 2019, 3,
026001.
(364) Li, L.; Plessow, P. N.; Rieger, M.; Sauer, S.; Sánchez-Carrera, R. S.; Schaefer, A.; Abild-
Pedersen, F. Modeling the Migration of Platinum Nanoparticles on Surfaces Using a Kinetic
Monte Carlo Approach. J. Phys. Chem. C 2017, 121, 4261–4269.
(365) Kim, M. J.; Alvarez, S.; Chen, Z.; Fichthorn, K. A.; Wiley, B. J. Single-Crystal
Electrochemistry Reveals Why Metal Nanowires Grow. J. Am. Chem. Soc. 2018, 140,
14740–14746.
(366) Trochet, M.; Mousseau, N.; Béland, L. K.; Henkelman, G. Off-Lattice Kinetic Monte Carlo
Methods. In Handbook of Materials Modeling; Springer International Publishing: Cham,
2019; pp 1–29.
(367) Andersen, M.; Panosetti, C.; Reuter, K. A Practical Guide to Surface Kinetic Monte Carlo
Simulations. Front. Chem. 2019, 7, 202.
(368) Xu, L.; Henkelman, G. Adaptive Kinetic Monte Carlo for First-Principles Accelerated
Dynamics. J. Chem. Phys. 2008, 129, 114104.
(369) Trushin, O.; Karim, A.; Kara, A.; Rahman, T. S. Self-Learning Kinetic Monte Carlo
Method: Application to Cu(111). Phys. Rev. B 2005, 72, 115401.
(370) Henkelman, G.; Jónsson, H. A Dimer Method for Finding Saddle Points on High
Dimensional Potential Surfaces Using Only First Derivatives. J. Chem. Phys. 1999, 111,
7010–7022.
(371) Bussi, G.; Laio, A. Using Metadynamics to Explore Complex Free-Energy Landscapes. Nat.
Rev. Phys. 2020, 2, 200–212.
(372) Henkelman, G. Atomistic Simulations of Activated Processes in Materials. Annu. Rev.
Mater. Res. 2017, 47, 199–216.
(373) Greeley, J. Theoretical Heterogeneous Catalysis: Scaling Relationships and Computational
Catalyst Design. Annu. Rev. Chem. Biomol. Eng. 2016, 7, 605–635.
(374) Abild-Pedersen, F.; Greeley, J.; Studt, F.; Rossmeisl, J.; Munter, T.; Moses, P.; Skúlason,
E.; Bligaard, T.; Nørskov, J. Scaling Properties of Adsorption Energies for Hydrogen-
Containing Molecules on Transition-Metal Surfaces. Phys. Rev. Lett. 2007, 99, 016105.
(375) Jones, G.; Studt, F.; Abild-Pedersen, F.; Nørskov, J. K.; Bligaard, T. Scaling Relationships
for Adsorption Energies of C2 Hydrocarbons on Transition Metal Surfaces. Chem. Eng. Sci.
2011, 66, 6318–6323.
(376) Studt, F.; Abild-Pedersen, F.; Bligaard, T.; Sørensen, R. Z.; Christensen, C. H.; Nørskov, J.
K. Identification of Non-Precious Metal Alloy Catalysts for Selective Hydrogenation of

122
 Acetylene. Science 2008, 320, 1320–1322.
(377) Medford, A. J.; Lausche, A. C.; Abild-Pedersen, F.; Temel, B.; Schjødt, N. C.; Nørskov, J.
K.; Studt, F. Activity and Selectivity Trends in Synthesis Gas Conversion to Higher
Alcohols. Top. Catal. 2014, 57, 135–142.
(378) Medford, A. J.; Vojvodic, A.; Hummelshøj, J. S.; Voss, J.; Abild-Pedersen, F.; Studt, F.;
Bligaard, T.; Nilsson, A.; Nørskov, J. K. From the Sabatier Principle to a Predictive Theory
of Transition-Metal Heterogeneous Catalysis. J. Catal. 2015, 328, 36–42.
(379) Calle-Vallejo, F.; Martínez, J. I.; Rossmeisl, J. Density Functional Studies of Functionalized
Graphitic Materials with Late Transition Metals for Oxygen Reduction Reactions. Phys.
Chem. Chem. Phys. 2011, 13, 15639.
(380) Montoya, J. H.; Tsai, C.; Vojvodic, A.; Nørskov, J. K. The Challenge of Electrochemical
Ammonia Synthesis: A New Perspective on the Role of Nitrogen Scaling Relations.
ChemSusChem 2015, 8, 2180–2186.
(381) Elnabawy, A. O.; Herron, J. A.; Scaranto, J.; Mavrikakis, M. Structure Sensitivity of Formic
Acid Electrooxidation on Transition Metal Surfaces: A First-Principles Study. J.
Electrochem. Soc. 2018, 165, J3109–J3121.
(382) Hong, X.; Chan, K.; Tsai, C.; Nørskov, J. K. How Doped MoS2 Breaks Transition-Metal
Scaling Relations for CO2 Electrochemical Reduction. ACS Catal. 2016, 6, 4428–4437.
(383) Bronsted, J. N. Acid and Basic Catalysis. Chem. Rev. 1928, 5, 231–338.
(384) Evans, M.; Polanyi, M. Inertia and Driving Force of Chemical Reactions. Trans. Faraday
Soc. 1938, 34, 11–24.
(385) Nørskov, J. K.; Bligaard, T.; Logadottir, A.; Bahn, S.; Hansen, L. B.; Bollinger, M.;
Bengaard, H.; Hammer, B.; Sljivancanin, Z.; Mavrikakis, M. et al. Universality in
Heterogeneous Catalysis. J. Catal. 2002, 209, 275–278.
(386) Ferrin, P.; Simonetti, D.; Kandoi, S.; Kunkes, E.; Dumesic, J. A.; Nørskov, J. K.;
Mavrikakis, M. Modeling Ethanol Decomposition on Transition Metals: A Combined
Application of Scaling and Brønsted-Evans-Polanyi Relations. J. Am. Chem. Soc. 2009, 131,
5809–5815.
(387) Liu, C.; Cundari, T. R.; Wilson, A. K. CO2 Reduction on Transition Metal (Fe, Co, Ni, and
Cu) Surfaces: In Comparison with Homogeneous Catalysis. J. Phys. Chem. C 2012, 116,
5681–5688.
(388) Liu, B.; Greeley, J. Decomposition Pathways of Glycerol via C–H, O–H, and C–C Bond
Scission on Pt(111): A Density Functional Theory Study. J. Phys. Chem. C 2011, 115,
19702–19709.
(389) Huang, S.-C.; Lin, C.-H.; Wang, J.-H. Trends of Water Gas Shift Reaction on Close-Packed
Transition Metal Surfaces. J. Phys. Chem. C 2010, 114, 9826–9834.
(390) Fajín, J. L. C.; Cordeiro, M. N. D. S.; Illas, F.; Gomes, J. R. B. Descriptors Controlling the
Catalytic Activity of Metallic Surfaces toward Water Splitting. J. Catal. 2010, 276, 92–100.
(391) Sutton, J. E.; Guo, W.; Katsoulakis, M. A.; Vlachos, D. G. Effects of Correlated Parameters
and Uncertainty in Electronic-Structure-Based Chemical Kinetic Modelling. Nat. Chem.
2016, 8, 331–337.
(392) Nørskov, J. K.; Rossmeisl, J.; Logadottir, A.; Lindqvist, L.; Kitchin, J. R.; Bligaard, T.;
Jónsson, H. Origin of the Overpotential for Oxygen Reduction at a Fuel-Cell Cathode. J.
Phys. Chem. B 2004, 108, 17886–17892.
(393) Norskov, J. K.; Abild-Pedersen, F.; Studt, F.; Bligaard, T. Density Functional Theory in
Surface Chemistry and Catalysis. Proc. Natl. Acad. Sci. 2011, 108, 937–943.

123
(394) Sabatier, P. Hydrogénations et Déshydrogénations Par Catalyse. Berichte der Dtsch. Chem.
Gesellschaft 1911, 44, 1984–2001.
(395) Zhao, Z. J.; Liu, S.; Zha, S.; Cheng, D.; Studt, F.; Henkelman, G.; Gong, J. Theory-Guided
Design of Catalytic Materials Using Scaling Relationships and Reactivity Descriptors. Nat.
Rev. Mater. 2019, 4, 792–804.
(396) Pérez-Ramírez, J.; López, N. Strategies to Break Linear Scaling Relationships. Nat. Catal.
2019, 2, 971–976.
(397) Jensen, K. D.; Tymoczko, J.; Rossmeisl, J.; Bandarenka, A. S.; Chorkendorff, I.; Escudero-
Escribano, M.; Stephens, I. E. L. Elucidation of the Oxygen Reduction Volcano in Alkaline
Media Using a Copper–Platinum(111) Alloy. Angew. Chemie - Int. Ed. 2018, 57, 2800–
2805.
(398) Nesselberger, M.; Ashton, S.; Meier, J. C.; Katsounaros, I.; Mayrhofer, K. J. J.; Arenz, M.
The Particle Size Effect on the Oxygen Reduction Reaction Activity of Pt Catalysts:
Influence of Electrolyte and Relation to Single Crystal Models. J. Am. Chem. Soc. 2011,
133, 17428–17433.
(399) Andersson, M. P.; Bligaard, T.; Kustov, A.; Larsen, K. E.; Greeley, J.; Johannessen, T.;
Christensen, C. H.; Nørskov, J. K. Toward Computational Screening in Heterogeneous
Catalysis: Pareto-Optimal Methanation Catalysts. J. Catal. 2006, 239, 501–506.
(400) Frey, K.; Schmidt, D. J.; Wolverton, C.; Schneider, W. F. Implications of Coverage-
Dependent O Adsorption for Catalytic NO Oxidation on the Late Transition Metals. Catal.
Sci. Technol. 2014, 4, 4356–4365.
(401) Tang, Y.; Roberts, C. A.; Perkins, R. T.; Wachs, I. E. Revisiting Formic Acid
Decomposition on Metallic Powder Catalysts: Exploding the HCOOH Decomposition
Volcano Curve. Surf. Sci. 2016, 650, 103–110.
(402) Quaino, P.; Juarez, F.; Santos, E.; Schmickler, W. Volcano Plots in Hydrogen
Electrocatalysis-Uses and Abuses. Beilstein J. Nanotechnol. 2014, 5, 846–854.
(403) Zeradjanin, A. R.; Grote, J. P.; Polymeros, G.; Mayrhofer, K. J. J. A Critical Review on
Hydrogen Evolution Electrocatalysis: Re-Exploring the Volcano-Relationship.
Electroanalysis 2016, 28, 2256–2269.
(404) Peterson, A. A.; Nørskov, J. K. Activity Descriptors for CO2 Electroreduction to Methane
on Transition-Metal Catalysts. J. Phys. Chem. Lett. 2012, 3, 251–258.
(405) Szécsényi, Á.; Khramenkova, E.; Chernyshov, I. Y.; Li, G.; Gascon, J.; Pidko, E. A.
Breaking Linear Scaling Relationships with Secondary Interactions in Confined Space: A
Case Study of Methane Oxidation by Fe/ZSM-5 Zeolite. ACS Catal. 2019, 9, 9276–9284.
(406) Doyle, A. D.; Montoya, J. H.; Vojvodic, A. Improving Oxygen Electrochemistry through
Nanoscopic Confinement. ChemCatChem 2015, 7, 738–742.
(407) Li, H.; Shin, K.; Henkelman, G. Effects of Ensembles, Ligand, and Strain on Adsorbate
Binding to Alloy Surfaces. J. Chem. Phys. 2018, 149, 174705.
(408) Hook, A.; Celik, F. E. Predicting Selectivity for Ethane Dehydrogenation and Coke
Formation Pathways over Model Pt-M Surface Alloys with Ab Initio and Scaling Methods.
J. Phys. Chem. C 2017, 121, 17882–17892.
(409) Liu, J.; Lucci, F. R.; Yang, M.; Lee, S.; Marcinkowski, M. D.; Therrien, A. J.; Williams, C.
T.; Sykes, E. C. H.; Flytzani-Stephanopoulos, M. Tackling CO Poisoning with Single-Atom
Alloy Catalysts. J. Am. Chem. Soc. 2016, 138, 6396–6399.
(410) Darby, M. T.; Réocreux, R.; Sykes, E. C. H.; Michaelides, A.; Stamatakis, M. Elucidating
the Stability and Reactivity of Surface Intermediates on Single-Atom Alloy Catalysts. ACS

124
 Catal. 2018, 8, 5038–5050.
(411) Darby, M. T.; Stamatakis, M.; Michaelides, A.; Sykes, E. C. H. Lonely Atoms with Special
Gifts: Breaking Linear Scaling Relationships in Heterogeneous Catalysis with Single-Atom
Alloys. J. Phys. Chem. Lett. 2018, 9, 5636–5646.
(412) Greiner, M. T.; Jones, T. E.; Beeg, S.; Zwiener, L.; Scherzer, M.; Girgsdies, F.; Piccinin,
S.; Armbrüster, M.; Knop-Gericke, A.; Schlögl, R. Free-Atom-like d States in Single-Atom
Alloy Catalysts. Nat. Chem. 2018, 10, 1008–1015.
(413) Thirumalai, H.; Kitchin, J. R. Investigating the Reactivity of Single Atom Alloys Using
Density Functional Theory. Top. Catal. 2018, 61, 462–474.
(414) Sun, G.; Zhao, Z.-J.; Mu, R.; Zha, S.; Li, L.; Chen, S.; Zang, K.; Luo, J.; Li, Z.; Purdy, S.
C. et al. Breaking the Scaling Relationship via Thermally Stable Pt/Cu Single Atom Alloys
for Catalytic Dehydrogenation. Nat. Commun. 2018, 9, 4454.
(415) Zhang, L.; Kim, H. Y.; Henkelman, G. CO Oxidation at the Au–Cu Interface of Bimetallic
Nanoclusters Supported on CeO 2 (111). J. Phys. Chem. Lett. 2013, 4, 2943–2947.
(416) van Deelen, T. W.; Hernández Mejía, C.; de Jong, K. P. Control of Metal-Support
Interactions in Heterogeneous Catalysts to Enhance Activity and Selectivity. Nat. Catal.
2019, 2, 955–970.
(417) Chen, Z.; Mao, Y.; Chen, J.; Wang, H.; Li, Y.; Hu, P. Understanding the Dual Active Sites
of the FeO/Pt(111) Interface and Reaction Kinetics: Density Functional Theory Study on
Methanol Oxidation to Formaldehyde. ACS Catal. 2017, 7, 4281–4290.
(418) Sandberg, R. B.; Hansen, M. H.; Nørskov, J. K.; Abild-Pedersen, F.; Bajdich, M. Strongly
Modified Scaling of CO Hydrogenation in Metal Supported TiO Nanostripes. ACS Catal.
2018, 8, 10555–10563.
(419) Rodriguez, J. A.; Ma, S.; Liu, P.; Hrbek, J.; Evans, J.; Perez, M. Activity of CeOx and TiOx
Nanoparticles Grown on Au(111) in the Water-Gas Shift Reaction. Science. 2007, 318,
1757–1760.
(420) Andersen, M.; Medford, A. J.; Nørskov, J. K.; Reuter, K. Analyzing the Case for
Bifunctional Catalysis. Angew. Chemie - Int. Ed. 2016, 55, 5210–5214.
(421) Andersen, M.; Medford, A. J.; Nørskov, J. K.; Reuter, K. Scaling-Relation-Based Analysis
of Bifunctional Catalysis: The Case for Homogeneous Bimetallic Alloys. ACS Catal. 2017,
7, 3960–3967.
(422) Choksi, T.; Majumdar, P.; Greeley, J. P. Electrostatic Origins of Linear Scaling
Relationships at Bifunctional Metal/Oxide Interfaces: A Case Study of Au Nanoparticles
on Doped MgO Substrates. Angew. Chemie - Int. Ed. 2018, 57, 15410–15414.
(423) Mehta, P.; Greeley, J.; Delgass, W. N.; Schneider, W. F. Adsorption Energy Correlations at
the Metal–Support Boundary. ACS Catal. 2017, 7, 4707–4715.
(424) Erlebacher, J.; Aziz, M. J.; Karma, A.; Dimitrov, N.; Sieradzki, K. Evolution of
Nanoporosity in Dealloying. Nature 2001, 410, 450–453.
(425) Wittstock, A.; Zielasek, V.; Biener, J.; Friend, C. M.; Baumer, M. Nanoporous Gold
Catalysts for Selective Gas-Phase Oxidative Coupling of Methanol at Low Temperature.
Science 2010, 327, 319–322.
(426) Zielasek, V.; Jürgens, B.; Schulz, C.; Biener, J.; Biener, M. M.; Hamza, A. V.; Bäumer, M.
Gold Catalysts: Nanoporous Gold Foams. Angew. Chemie Int. Ed. 2006, 45, 8241–8244.
(427) Xie, Y.; Wang, Z. W.; Zhu, T. Y.; Shu, D. J.; Hou, Z. F.; Terakura, K. Breaking the Scaling
Relations for Oxygen Reduction Reaction on Nitrogen-Doped Graphene by Tensile Strain.
Carbon N. Y. 2018, 139, 129–136.

125
(428) Kropp, T.; Mavrikakis, M. Effect of Strain on the Reactivity of Graphene Films. J. Catal.
2020, 390, 67–71.
(429) Rodrigues, J. A.; Goodman, D. W. The Nature of the Metal-Metal Bond in Bimetallic
Surfaces. Science 1992, 257, 897–903.
(430) Nilsson, A.; Pettersson, L. G. M.; Hammer, B.; Bligaard, T.; Christensen, C. H.; Nørskov,
J. K. The Electronic Structure Effect in Heterogeneous Catalysis. Catal. Letters 2005, 100,
111–114.
(431) Hammer, B.; Nørskov, J. K. Why Gold Is the Noblest of All the Metals. Nature 1995, 376,
238–240.
(432) Kibler, L. A.; El-Aziz, A. M.; Hoyer, R.; Kolb, D. M. Tuning Reaction Rates by Lateral
Strain in a Palladium Monolayer. Angew. Chemie Int. Ed. 2005, 44, 2080–2084.
(433) Xin, H.; Vojvodic, A.; Voss, J.; Nørskov, J. K.; Abild-Pedersen, F. Effects of d-Band Shape
on the Surface Reactivity of Transition-Metal Alloys. Phys. Rev. B 2014, 89, 115114.
(434) Xin, H.; Linic, S. Communications: Exceptions to the d -Band Model of Chemisorption on
Metal Surfaces: The Dominant Role of Repulsion between Adsorbate States and Metal d -
States. J. Chem. Phys. 2010, 132, 221101.
(435) Toyoda, E.; Jinnouchi, R.; Hatanaka, T.; Morimoto, Y.; Mitsuhara, K.; Visikovskiy, A.;
Kido, Y. The D-Band Structure of Pt Nanoclusters Correlated with the Catalytic Activity
for an Oxygen Reduction Reaction. J. Phys. Chem. C 2011, 115, 21236–21240.
(436) Zhou, W. P.; Lewera, A.; Larsen, R.; Masel, R. I.; Bagus, P. S.; Wieckowski, A. Size Effects
in Electronic and Catalytic Properties of Unsupported Palladium Nanoparticles in
Electrooxidation of Formic Acid. J. Phys. Chem. B 2006, 110, 13393–13398.
(437) Visikovskiy, A.; Matsumoto, H.; Mitsuhara, K.; Nakada, T.; Akita, T.; Kido, Y. Electronic
d-Band Properties of Gold Nanoclusters Grown on Amorphous Carbon. Phys. Rev. B 2011,
83, 165428.
(438) Van Bokhoven, J. A.; Miller, J. T. D Electron Density and Reactivity of the d Band As a
Function of Particle Size in Supported Gold Catalysts. J. Phys. Chem. C 2007, 111, 9245–
9249.
(439) Grant, J. T.; Carrero, C. A.; Goeltl, F.; Venegas, J.; Mueller, P.; Burt, S. P.; Specht, S. E.;
McDermott, W. P.; Chieregato, A.; Hermans, I. Selective Oxidative Dehydrogenation of
Propane to Propene Using Boron Nitride Catalysts. Science 2016, 354, 1570–1573.
 

Table of Contents Figure

126
127