Computational Modelling of Nanoparticles

Introduction to modeling nanoclusters and nanoparticles

Scott M. Woodley*,¹; Stefan T. Bromley†,‡,¹    * Kathleen Lonsdale Materials Chemistry, Department of Chemistry, University College London, London, United Kingdom

† Departament de Ciència de Materials i Química Física & Institut de Química Teòrica i Computacional, Universitat de Barcelona, Barcelona, Spain

‡ Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Spain

¹ Corresponding authors: email address: Scott.Woodley@ucl.ac.uk, S.Bromley@ub.edu

Abstract

Although each chapter is self-contained, if you are relatively new to the field of modeling nanoclusters and nanoparticles, or would like to recap important ideas, we first provide an extensive introduction. Here we cover the topic of energy landscapes, where we focus on energy definitions based on analytic forms to describe the interaction between atoms or ions (e.g., Lennard Jones potential)—but clearly these expressions can be replaced with a more full treatment of electronic structure. We also introduce various Monte Carlo approaches typically employed to explore these landscapes. We employ analogies to help with the understanding of how the various algorithms work and give a range of concrete published examples based on the modeling of nanoclusters and nanoparticles of inorganic materials (we leave the coverage of other types of nanoparticles such as metallic clusters and their alloys for later chapters). The development of algorithms for exploring energy landscapes is still very active, where the objectives are to improve their computational efficiency, and/or to expand the range of targeted features/properties. For the latter, historically only the athermal lowest energy minima were first targeted. Afterward, transition points or barriers between minima were also targeted. Nowadays, temperature, pressure, chemical potential, probability flows, and lifetimes of metastable states may all be targeted. For the former objective, reduction in the computational cost of completing a search of an energy landscape may be achieved by a number of means such as (i) employing cheaper to evaluate, yet coarser, measures of the energy landscape before using more computationally expensive refined measures; (ii) employing machine learning algorithms to eventually (once a sufficient training set has been obtained) remove the need to compute the chosen expensive to evaluate the measure of the energy landscape; and (iii) ensuring a search does not get stuck in localized regions through the use of evolutionary ideas (e.g., structural mutations, grouping structures into niches) to maintain or increase the structural diversity in a set of candidate nanoparticle structures. Although new approaches to exploring energy landscapes are continually being proposed, the required foundational background for describing energy landscapes and searching them can be found in numerous text books or reviews on this topic. This introduction is largely adapted from one of these sources [S.M. Woodley, Nanoclusters and nanoparticles, in: S.T. Bromley, M.A. Zwijnenburg (Eds.), Computational Modeling of Inorganic Nanomaterials, CRC Press, Taylor and Francis Group, Boca Raton, USA, 2016, pp. 3–46].

Keywords

Nanoclusters; Energy landscapes; Global optimization; Interatomic potentials; Inorganic materials

1 Introduction

In purely length-scale terms, the nanoscale may be approximately taken to be the size range between 1 and 100 nm (1 nm = 10− 9 m). Accordingly, we may formally define nanoparticles to be any particles with a diameter lying within this range [1]. The lower limit of the nanoscale can be assumed from its name (anything smaller than 1 nm diameter often referred to as subnanometer or molecular), whereas the choice of upper limit is less well demarcated. From a phenomenological perspective, if we take a macroscopic material sample and proceed to remove atoms from it, the upper bound of the nanoscale can be assigned to be the size at which the material begins to display specific, nonbulk-like properties (e.g., atomic ordering, electronic band gap). Clearly, however, this definition is material and property dependent and the crossover size between the bulk macroscopic scale and the upper bound to the nanoscale can vary somewhat above and below 100 nm depending on both. In this book, the focus is on modeling of nanoparticles employing quantum electronic structure and classical atomistic methods. Using modern high-performance computing facilities the former approaches are limited to nanoparticles having diameters below ~ 10 nm. Classical atomistic methods, typically employing empirical force fields, are more computationally efficient but are still typically used to model nanoparticles with a maximum diameter of a few 10s of nanometers. As such, for this introductory chapter and throughout the book we can safely assume that all theoretically treated particles have diameters less than 100 nm.

Nanoparticles attract a great deal of attention due to their tremendous importance for emerging nanotechnologies. Interest in the synthesis and characterization of nanoparticles has been driven, for example, by their increasing use in electronics, energy conversion, composite materials, and catalysis. Nanoparticles at the lower end of the nanoscale (composed of less than ~ 100 atoms) are often referred to as nanoclusters. Studying small nanoclusters can help us understand and gain insight into the atomic mechanisms of nucleation and the early stages of crystal growth; see, for example, the perspective article on modeling nanoclusters and nucleation by Catlow et al. [2] and the review on experimental and computational studies of zinc sulfide (ZnS) nanostructures by Hamad et al. [3]. The size-dependent changes, in going from the bulk through the nanoscale, are well demonstrated by considering the specific case of ZnS. In the bulk, ZnS can adopt either the sphalerite (cubic) or the wurtzite (hexagonal) crystal structure. The two phases coexist in nature, but sphalerite is the most stable bulk form under ambient conditions. For smaller nanoparticles the wurtzite phase becomes more thermodynamically stable [4]; moreover, nanoparticles with mixed cubic and hexagonal stacking have been synthesized [5]. For small nanoclusters of ZnS, the atomic structures no longer resemble cuts taken from any bulk crystalline phase. Such nanoclusters are readily created by either laser ablation of the bulk structure or by nucleation in solution [6–8]. In Fig. 1, we schematically show this size-dependent evolution of materials properties with respect to a generic property G(R), which depends on the radius R of the material sample. For small R we are in the size range of subnanometer clusters, the G(R) varies dramatically with small changes in R, the every atom counts regime. For large R, the particles are bulk-like and vary in a linear manner with respect to R− α, where α is a constant (typically taken to be 1/3 for reasons explained later). Due to the linear dependency on R− α this regime is often referred to as the scalable regime [9]. The intermediate size range that bridges the subnanometer and bulk-like regimes is the nanoscale, where we find nanoclusters and nanoparticles.

Fig. 1 Schematic representation of the size-dependence of a general property, G ( R ), with respect to an inverse function of the radius, 1/ R α , where α is constant. The region of interest for the modeling studies described in this book (i.e., the nanoscale) is demarked by an orange circle ( light gray in the print version) which encompasses nanoclusters and nanoparticles.

In general, for nanoparticles with radii below ~ 5 nm, clear diffraction patterns are not easy to obtain and accurate structure determination of these small species using standard X-ray diffraction techniques start to fail. Diffraction techniques require a sufficiently sized target (e.g., large single crystals or powders formed of smaller crystals with the same crystalline phase) to be effective. Due to these limitations structure determination of nanoclusters currently relies on computational modeling techniques. As nanocluster atomic structure can be highly sensitive to the number of constituent atoms, these methods should be used anew for every new nanocluster size under investigation. Confidence in structural predictions provided by computational modeling can be obtained by also calculating observable properties (e.g., vibrational spectra) specifically associated with the predicted nanocluster structure, and comparing with the respective experimentally measured property.

Now that the topic of nanoclusters and nanoparticles have been introduced, you can choose to jump straight to the chapter of that is of more interest to you as each chapter is self-contained. If you are relatively new to the field of modeling nanoclusters and nanoparticles, however, we would recommend that you learn or recap important ideas by first reading the rest of this introduction to modeling nanoclusters and nanoparticles. Here we cover the topic of energy landscapes, where we focus on energy definitions based on analytic forms to describe the interaction between atoms or ions (e.g., Lennard Jones potential)—but clearly these expressions can be replaced with a more full treatment of electronic structure. We also introduce various Monte Carlo approaches typically employed to explore these landscapes. We will employ analogies to help with the understanding of how the various algorithms work and give a range of concrete published examples based on the modeling of nanoclusters of inorganic materials (we leave the coverage of other types of nanoparticles such as metallic clusters and their alloys for later chapters). The development of algorithms for exploring energy landscapes is still very active, where the objectives are to improve their computational efficiency, and/or to expand the range of targeted features/properties. For the latter, historically only the athermal lowest energy minima were originally targeted. Afterward, transition points or barriers between minima were also targeted. Nowadays, temperature, pressure, chemical potential, probability flows, and lifetimes of metastable states may all be targeted. For the former objective, reduction in the computational cost of completing a search of an energy landscape may be achieved by a number of means such as (i) employing cheaper to valuate, yet coarser, measures of the energy landscape before using more computationally expensive refined measures; (ii) employing machine learning algorithms to eventually (once a sufficient training set has been obtained) remove the need to compute the chosen expensive to evaluate the measure of the energy landscape; and (iii) ensuring a search does not get stuck in localized regions through the use of evolutionary ideas (e.g., structural mutations, grouping structures into niches) to maintain or increase the structural diversity in a set of candidate nanoparticle structures. Although new approaches to exploring energy landscapes are continually being proposed, the required foundational background for describing energy landscapes and searching them can be found in numerous text books or reviews on this topic. This introduction is largely adapted from one of these sources [S.M. Woodley, Nanoclusters and nanoparticles, in: S.T. Bromley, M.A. Zwijnenburg (Eds.), Computational Modeling of Inorganic Nanomaterials, CRC Press, Taylor and Francis Group, Boca Raton, USA, 2016, pp. 3–46].

2 Nanoparticle properties and energy landscapes

What are the properties of nanoparticles? The first and arguably most fundamental property we may want to establish is the atomic structure of a nanoparticle. Once a nanoparticle's atomic structure is known other physical and electronic properties can be calculated. The atomic structure, as with other properties, is dependent on a number of factors, such as the temperature and pressure, whether in vacuum or in a medium, or whether it is capped or not by organic molecules (i.e., chemically passivated). Perhaps, the best place to start is to ask what we mean by the statement: this is the structure of the nanoparticle. One key relevant parameter that has not been mentioned yet is that of time. As the precise atomic coordinates of a nanoparticle fluctuates with time, average coordinates are more relevant. Here, we can introduce the concept of locally ergodic regions [10–13]. Assuming that the atomic structure can be measured and that it takes a period of time tm (often called the observation time) to make this measurement, then over this time period it is important that the measured time average of the coordinates does not change significantly. In particular, we require that further measurements would yield essentially the same average atom coordinates. This range of accessible coordinates is referred to as a locally ergodic region. In a crystal, for example, each atom vibrates about one of the lattice sites and at any specific time the atoms are likely to be slightly displaced from their respective lattice sites in different random directions. The maximum magnitude of these displacements is dependent on a number of factors including the atoms’ local atomic and electronic environment and macroscopically measurable variables, such as the temperature of the crystal and any externally applied pressure. Due to the dynamic nature of the atoms, a measurement of atomic structure is often actually a measurement of the average, or equilibrium position of each atom. The accuracy of such measurements depends on how well the instantaneous positions are sampled, i.e., the time of measurement must be significantly greater than the time, teqR, thus allowing each atom to explore its respective part of the locally ergodic region R. In turn, the time to make such measurements must be shorter than the time required for the system to escape the current locally ergodic region R, the escape time, tesR. Thus, the following inequality must be satisfied [10–13]:

   (1)

One analogy often used in the field of structure prediction is that between the state of the system (a particular atomic structure) and the location on a map of a person who is out exploring a hilly terrain, the so-called walker. The altitude of the walker, or height of the landscape, is analogous to the (potential) energy of the system. The energy landscape can be generated by mapping out the energy of the system, E(r), as a function of atomic coordinates (xi, yi, zi), where the subscript i is used to distinguish each of the N atoms and r is the 3N-component vector, formed by concatenating all atomic coordinates. In general, the number of variables, or atomic coordinates, will typically be greater than two and we can also speak of an energy hypersurface.

It is convenient to assume that the hypersurfaces of relevance to nanocluster configurations and their energies are both smooth and continuous (i.e., there are no discontinuities in the energy or its gradient). The analogous landscape of the walker typically contains undulating hills and valleys to be explored. For the moment, let us assume that the walker is wearing roller skates. The stability of the walker, and the corresponding atomic structure, depends on the local environment of the walker. The speed of the walker increases when traveling downhill and decreases when going uphill. Here, the speed of the walker is related to the kinetic energy of the atoms (assuming no friction). If the local landscape is flat (i.e., the 3N-dimensional gradient in all directions is 0), then the speed of the walker remains constant. In Cartesian coordinates, this can be expressed as

   (2)

where Fi is the force on atom i and, by definition, ri = (xi, yi, zi). Thus, if the initial speed of this walker is 0, then the walker remains there indefinitely. Such points are referred to as stationary points; on an undulating landscape, however, it is important to ascertain whether the particular stationary point represents a stable or unstable point.

A local maximum on the landscape, or hilltop, represents an unstable stationary point. In such positions, when the walker makes any small movement they will experience a gravitational force pulling them down the hillside, and thus moving them away from such a location. An inflection stationary point is also an unstable point of the landscape. For simplicity, let us consider the one-dimensional (1D) landscape E(x) = x³, which has an inflection stationary point at x = 0. Even if the initial speed of a walker is in the locally uphill direction (toward x > 0), the walker will eventually return to the stationary point but with a velocity in the direction of the local downhill (toward x < 0) and thus moving away from the inflection point. A minimum on the landscape represents a stable stationary point as a small displacement of the walker will always result in the walker experiencing a gravitational force that pulls the walker back down the hill toward the minimum regardless of the initial displacement. Here, the walker oscillates about the local minimum and thus remains in this locally ergodic region, or energy basin. The global minimum (GM) on the energy landscape is the stable point. If the local minimum (LM) is not the lowest point on the landscape, then it is a metastable stationary point—given enough time the walker may escape this locally ergodic region and find a more stable configuration. For a 1D landscape, maxima are transition points between two regions. Assuming the walker is at a LM and ignoring any quantum effects (e.g., low-temperature tunneling through barriers), the energy difference between this LM and a neighboring maximum is the energy barrier that the walker needs to overcome in order to change the average atomic configuration of the system (i.e., to switch locally ergodic regions). If the walker can successfully pass over this barrier (within the observation time of interest), then the locally ergodic region includes both basins. Note that since escape times usually decrease with increasing temperature, the locally ergodic regions for the time scales of interest tend to be larger at elevated temperatures [13,14].

For multidimensional, or N-dimensional landscapes where N > 1, a stationary point can also be an order n saddle point, where n is the number of orthogonal directions along which the stationary point is a 1D maximum. A mountain pass is an oft-used example of a region about a saddle point on a 2D landscape (see Fig. 2A). The highest energy point along the lowest energy pathway between two local minima is a saddle point. Low-order saddle points are the transition points between energy basins, or locally ergodic regions that represent different atomic configurations. The likelihood of escape from an energy basin is not only dependent on the height of the energy barrier but also the width of the pass, how many passes there are and how easy it is to find and navigate through at least one of them, and the probability to return [15]. Locally ergodic regions can thus be one basin or several connected energy basins (a super-basin), or even a nested sequence of marginally ergodic regions [13] (the latter representing, e.g., a glassy material). However, in addition, locally ergodic regions can be zones that are enclosed by entropic barriers [16]. For example, where the path of escape has no energy barrier but is via a very narrow pass that is hard for the system to find. Schematic examples of locally ergodic regions for both a glassy nanoparticle and an entropic barrier are shown in Fig. 2B and C. For further discussion see Refs. [10–14].

Fig. 2 Three example landscapes containing locally ergodic regions: (A) two local energy basins, one either side of a mountain range, (B) seven shallow local energy basins within a super-basin, and (C) an enclosed area containing no local energy minimum. Red ( gray in the print version), green ( gray in the print version), and purple ( dark gray in the print version) lines are the highest, lowest, and second lowest energy contours, respectively, the broken blue ( gray in the print version) trajectories represent paths taken by a walker, and the cross in (A) marks a saddle point. In (A), the time averaged atomic coordinates of the walker within the higher energy basin represents a metastable structure that is predicted to be observed provided Eq. (1) is satisfied before the walker escapes via the mountain pass to the lower basin representing the stable structure. In (B), if Eq. (1) is satisfied for the path within the purple contour, then an average of the structures from each local basin is observed. Eq. (1) can also be satisfied with an entropic barrier, as shown in (C). Note that the paths within each locally ergodic region would be much longer and, for a random walker, the pathway would be more jagged as the walker's momentum is not considered.

3 Predicting nanoparticle structure

To predict the atomic structure of a nanocluster or nanoparticle, one should overcome a number of challenges.

Challenge 1 is to find a suitable analytical expression (or numerical representation) for a cost function (i.e., a function, Ξ(r), which describes the system of interest in a tractable manner). Ideally, the cost function should be robust and computationally cheap to evaluate. An energy-based cost function is typically chosen for determining nanoparticle structure within or near a locally ergodic region. It is convenient to define the region of the energy landscape, Ω, that corresponds to the thermally accessible atomic coordinates. Points outside of Ω would include, for example, structures containing two or more atoms at the same location. As we are only interested in solutions within Ω, the cost function need not be accurately defined outside of Ω and, therefore, in such situations it can be redefined as a surrogate cost function based, for example, only on geometrical parameters without recourse to the actual physical stability. Points outside of Ω are therefore computationally cheap to determine, and a larger proportion of available computer resources can be spent on assessing more feasible structures within Ω—ideally any search should be prevented from leaving Ω, but this is not always easy or straightforward to enforce. An important requirement for such a surrogate cost function is that a higher value is returned for points outside Ω than within Ω, for example, both the surrogate cost function and Ξ(r) should share common minimum super-basins. The uphill gradient of the surrogate cost function outside of Ω should, ideally, point away from Ω. Another advantage of a surrogate cost function (discussed in more detail later) is that the energy for points outside of Ω are sometimes problematic to calculate. Even within Ω, the accuracy does not need to be very high as more accurate evaluations of Ξ(r) can be obtained in later refinement stages.

Successfully completing this first challenge does not immediately provide us with nanoparticle structure as, going back to the walker analogy, we should imagine that the entire landscape is blanketed by a fog. Regions only become visible after the walker has visited every location within the region, that is, Ξ(r) has been computed at every r within the region. Finding the locally ergodic regions on this energy hypersurface is the second challenge. In the first instance, it is common practice to search this hypersurface for local minima. If desired, the locally ergodic region can be explored about each LM as a separate stage. At this point, it is important to note that as the number of atoms in a nanoparticle increases so does the dimensionality of Ξ(r), the number of LM, and hence the difficulty of finding the lowest energy minima on the landscape. Optimally, during the search we do not want to revisit the parts of the landscape, which have already been unveiled previously, or spend excessive time in areas devoid of energy basins. To help navigate around the energy landscapes with the aim of finding the lowest energy minima, numerous global optimization techniques have been developed.

Further challenges, which are not explored in detail in this chapter, include finding the local saddle points between the LM [17], probability flows between LM [15,18] and, finally, residence times of structural solutions of particular type [19]. Saddle points corresponding to energy barriers between local basins can be conveniently mapped as dendritic (tree) graphs [20–22], from which the complexity of the landscape can be visualized. If the dendritic tree representing an energy landscape resembles a downward pointing pine cone (see Fig. 3A) the probability of the system finding the GM stochastically is facilitated through funneling (i.e., local minima becoming deeper and deeper the closer one is to the GM within a single super-basin). The task of locating the GM is much harder if the landscape is characterized by a map resembling a weeping willow tree (see Fig. 3B) or where the GM resides outside a super-basin containing all other minima (Fig. 3C). A plethora of tree graphs and their analysis can be found in the comprehensive book by Wales [23].

Fig. 3 A dendritic tree map of (A) the energy landscape for (MgF 2 ) 3 nanoclusters, where energies are taken from Neelamraju et al. [18], and two fictitious dendritic tree maps based on (A), but modified to demonstrate landscapes for which the global minimum is harder to locate because: (B) there is no funneling; and (C) funneling leads away from the global minimum, which is outside the super-basin containing all other local minima. Note that only the saddle point corresponding to the highest energy point along the lowest energy path between two local energy minima is marked on these maps.

In Sections 5–8, global optimization techniques employed to search for atomic configurations corresponding to the energetic GM are described. Such approaches are readily applied to nanoclusters with relatively low numbers of degrees of freedom. For modeling larger nanoparticles different techniques and assumptions are used as the number of possible configurations to generate (let alone evaluate) rapidly becomes increasingly difficult with increasing size. In Section 9, we show examples illustrating the general inevitable tendency that GM nanoparticle configurations become more bulk-like with increasing size. For such relatively large nanoparticles with a fixed type of bulk-like atomic structure one often assumes that GM nanoparticle type is that which has the lowest surface energy (see discussion later in this section on nanoparticle morphology). Ordering the GM configurations by size, starting from the smallest nanocluster, one often observes a number of size-dependent structural/morphological transitions: for example, planar to nonplanar, nonbulk-like to bulk-like, or a change in the bulk phase that the bulk-like configurations adopt. Each transition can occur over a range of sizes, for example, after a series of nonbulk-like GM, there is a sequence of nonbulk-like and bulk-like GM, before there is only bulk-like GM. With enough atoms, it is also easy to imagine a polycrystalline nanoparticle, that is, a bulk-like configuration that contains grain boundaries. A snapshot from a molecular dynamics simulation of a nanoparticle growing from one or more nucleation points (see Fig. 4) illustrates a grain boundary in a nanoparticle. In nature, there may of course be many defects; for example, Sayle et al. [25] have also modeled microtwinning within a MnO2 nanoparticle formed during a molecular dynamics simulation using a temperature schedule designed to first melt a bulk-like cut and then to encourage recrystallization of a low-energy configuration (see discussion on simulated annealing in Section 6 and the more in-depth insight given in Chapter 3).

Fig. 4 Structural model of a Li 2 MnO 3 nanoparticle containing a grain boundary obtained from a molecular dynamic simulations that was initialized from a cut out of the bulk phase [24]. The blue (light gray in the print version) and yellow (white in the print version) transparent layers highlight the two parts of the nanoparticle that are on either side of the grain boundary, and green (light gray in the print version), blue (dark gray in the print version), and red (gray in the print version) spheres represent Li+, Mn⁴ +, and O² − ions, respectively.

For large pure-phase single crystals, the morphology (crystal shape) can be determined by the requirement for the overall surface energy to attain a minimum. Typically, the equilibrium morphology is determined using the Wulff construction, where the distance from the center of the crystal to the center of each facetted surface is proportional to its surface energy. Thus, low-energy surfaces are realized and dominate the morphology (see Fig. 5). Computationally, the bulk surface energy of each candidate facet (chosen plane along which the bulk phase is cleaved) is readily determined using periodic surface models (e.g., take a single well-isolated slab with 2D periodic boundary conditions and increase the number of atomic layers parallel to the surface until the surface energy has converged). For nanoclusters, the assumption that each surface ends a semiinfinite number of atomic layers is usually wrong, although energy minima of nanoclusters can resemble cuts from the bulk phase (see cuboid examples in Section 9, which can be regarded as cuts from the NaCl rock salt structure). Hence, for nanoclusters, global optimization techniques are applied to search E(ri) for all atoms, i. As the number of atoms of the cluster increase, there must come a point where the interior atoms of the energy minimum structure resemble the bulk phase. Hence, the number of variables, and the space to search over, can be dramatically reduced by fixing or initializing these to bulk-like sites and then applying Monte Carlo (MC)-based global optimization or molecular dynamics to find the location of the outermost atoms.

Fig. 5 ) morphology predicted using the Wulff construction with the surface energies stated below each crystal. The size of the exposed surface increases (decreases) as the relative surface energy decreases (increases).

In the following section, we address the first challenge and describe the most commonly used cost functions, including energy definitions. The second challenge will be addressed next under the assumptions that only a local minimum is sought (rather than the whole locally ergodic region) for nanoclusters in vacuum in the athermal limit, and that all atomic coordinates are (initially) unknown variables. However, the methods described can be extended to include regions of atoms that are fixed.

4 Energy functions

To assess the quality of a candidate nanoparticle structure, a cost function, Ξ(r), is evaluated, where lower values imply a better fit to the target requirements. For nanoclusters, an energy-based cost function is typically employed and the target requirements may also include m physical properties, λmtarget, for example, observed infrared frequencies [26,27], or, perhaps, a specific energy difference between the lowest unoccupied and highest occupied electronic states [28]:

(3)

In this expression, wm is the relative weight, with respect to the cluster's energy, given to the deviation of the calculated value(s) of property m (i.e., λmcalc.) from its target. In the field of structure prediction, the atomic coordinates, r, are typically sought that only minimize the energy term, and thus discussions later are restricted to energy-based cost functions. Optimization algorithms presented in later sections are, however, also applicable to the more general expression in Eq. (3), although the derivatives with respect to r for the physical properties may be more demanding to compute.

As in other areas of molecular simulation, energies of nanoclusters can be calculated using methods based either on classical interatomic potentials or by employing quantum mechanical electronic structure techniques. Calculating the energy using electronic structure techniques can in itself be a challenge, particularly when the nanocluster does not resemble a reasonable chemical structure, or is a candidate structure that is outside of Ω (as defined in Section 3). For such candidate structures we may encounter problems of nearly degenerate electronic terms, which are difficult to converge in the electronic space, and the computational cost of structural relaxation is typically greater than that for candidate structures within Ω. Moreover, the resulting configuration of such a calculation is not necessarily relevant as it could be a high-energy local minimum or even be a fragmented structure. Thus, evaluation of such candidates using electronic structure approaches should be avoided as it is not just a significant computational cost but often does not aid the search for the global or other low-energy local minimum structures. Exploration of energy landscapes based on electronic structure methods has been successfully applied to generating low-energy minimum, atomic structures of nanoclusters (see, for example, Refs. [18,26,29–34]). However, some form of prescreening or filtering [30] using a surrogate cost function is usually employed to avoid a detailed assessment of non-Ω candidate structures. The application of machine learning is an alternative approach to reducing the cost of direct calculation of the electronic structure (apart from that required in the initial training) or, indeed, any other targeted property. Herein, we focus of classical atomistic approaches to modeling nanocluster energy landscapes. A recent review of quantum mechanical studies of moderately sized nanoparticles can be found in Ref. [35].

Nanocluster energies calculated using methods based on interatomic potentials are far less computationally expensive than electronic structure-based approaches. Therefore, larger numbers and sizes of clusters can readily be explored. These methods do, however, suffer from a number of limitations—interatomic potential parameters may depend on the coordination of the atoms; the oxidation state of each atom is required to be predefined; and, in some cases, key electronic effects (not modeled) are important for obtaining accurate atomic structures and energy ranking of minima—and consequently may in some cases give inaccurate results. Some limitations can also be advantageous, for example, for ionic oxide nanoclusters, predefining all oxide anions with a fixed charge automatically avoids unwanted minima structures containing peroxides (i.e., O–O bonding) or fragmented clusters with isolated oxygen atoms.

A fruitful approach is to combine both classical and quantum approaches, with the interatomic potential-based methods being used in a first stage during the exploration of a range of cluster sizes and structures, while the electronic structure methods are used in a second stage to refine the energy ordering and structures of a selected subset of clusters (e.g., those calculated to be low in energy in the first stage). In the second stage, electronic structure methods can be also employed to test the reliability of the first stage calculations by probing cluster structures, which are likely to be sensitive to the parameterization of the interatomic potential used. Typically potential parameters are tuned to reproduce bulk structures and properties as these are more readily available. However, potential parameters have been refined to reproduce atomic structures of nanoclusters that have been optimized using electronic structure methods [36], especially when the local atomic structure of the nanocluster differs significantly from that in the bulk phases.

The total energy of the system, based on interatomic potentials and a rigid ion model, is written as a function of the atomic coordinates. For strongly ionic systems, such as alkali halides and alkaline earth metal oxides, the Born model provides a good basis, allowing the energy to be written as a sum of Coulomb and short-range terms:

   (4)

where the cluster is composed of N atoms, rij, and Eij are the distance and short-range pairwise interaction between atoms i and j, respectively, and qk is the charge on the ion k. Here, zero energy is chosen to represent the system when all interatomic distances are infinite. In other words, Eq. (4) gives the energy to form the cluster from its constituent ions. We note that the charges qk are fixed and thus any variability in the electron transfer between ions is not taken into account. Moreover, single-body terms are ignored as no internal electronic energies are taken into account and, given we are interested in energy differences between two configurations with identical constituents, this term will then cancel. The short-range interactions, Eij, are included in order to account for: (i) the strong Pauli repulsion between ions at short interionic distances, and (ii) the attractive Van de Waals interactions, or dispersion (London forces) between ions that is more noticeable at longer distances. Short-range interactions are typically modeled using one of the following three potentials: the Buckingham potential [37],

   (5)

the Lennard-Jones potential [38],

   (6)

or the Morse potential [39,40],

(7)

The potential parameters, Aij, Bij, Cij, Dij, ρij, and aij are species dependent (hence the use of ij subscripts). When refining these potential parameters it is important to remember what each term in Eij or parameter represents, for example, Cij controls the strength of the dispersion between atoms i and j, whereas aij is the equilibrium bond distance for a diatomic molecule composed of atoms i and j.

The shell model [41,42] can also be employed to describe the electronic polarizability of individual ions. In this approach, each ion, i, is represented by two point charges, a core with mass, Mi, and charge, Xi, and a massless shell with charge, Yi, which are coupled together via a harmonic spring (with spring constant ki). Note that the short-range potentials given above are typically applied between shells, that Xi + Yi is the charge of ion, and that the polarizability of the ion is proportional to Yi²/ki. Using the shell model, the high-frequency dielectric constants can also be computed assuming only the positions of the shells have a chance to respond to a changing electric