When do short-range atomistic machine-learning models fall short?

Abstract

We explore the role of long-range interactions in atomistic machine-learning models by analyzing the effects on fitting accuracy, isolated cluster properties, and bulk thermodynamic properties. Such models have become increasingly popular in molecular simulations given their ability to learn highly complex and multi-dimensional interactions within a local environment; however, many of them fundamentally lack a description of explicit long-range interactions. In order to provide a well-defined benchmark system with precisely known pairwise interactions, we chose as the reference model a flexible version of the Extended Simple Point Charge (SPC/E) water model. Our analysis shows that while local representations are sufficient for predictions of the condensed liquid phase, the short-range nature of machine-learning models falls short in representing cluster and vapor phase properties. These findings provide an improved understanding of the role of long-range interactions in machine learning models and the regimes where they are necessary.

Machine learning potentials for extended systems: a perspective

“…There have been recent successes in materials where a priori one perhaps would not expect that, e.g. some bulk oxides [27,28] and even bulk liquid water [29,30].…”

“…This is well justified and introduces only very small errors e.g. for systems without significant charge transfer like in elemental compounds [47,48] or in the presence of efficient screening, like in bulk liquid water [29,30]. Still, any interaction present beyond the cutoff limits the accuracy that can be achieved, because for the ML algorithm such interactions result in inconsistent data that cannot be represented with the available information about the local atomic environments.…”

“…Assessing the relevance of long-range interactions is thus an important task and has been done e.g. for the cases of amorphous carbon [48], carbon clusters [50], water clusters [29], organic molecules [49,51] and the adsorption of metal clusters at oxides [51]. For instance, it has been shown in a locality test for carbon [48] how long-range interactions can be quantified by sampling different atomic positions outside the cutoff sphere as illustrated in Fig.…”

On the role of long-range electrostatics in machine-learned interatomic potentials for complex battery materials

“…34,[39][40][41] However, the importance of long-range effects will clearly depend on the material and property of interest and thus demands more systematic scrutiny. 42 For example, ionic diffusion in electrolytes may lead to the transient local accumulation and depletion of charges, which can break the isotropy of the electrostatic environment. Even more critically, grainboundaries, interfaces, and defects may lead to a permanent localized polarization of materials.…”

MB-Fit: Software Infrastructure for Data-Driven Many-Body Potential Energy Functions

“…52 On the other hand, ML PEFs trained on condensed-phase data are able to closely reproduce the corresponding ab initio simulations of liquid and solid phases but are not, in general, directly transferable to molecular clusters or air/solid and air/liquid interfaces. 53 In this context, it should be noted that since gas-phase training data are generated for molecular systems with a handful of atoms, they can be computed at relatively higher levels of theory, often coupled cluster with single, double, and perturbative triple excitations, i.e., CCSD(T), the "gold standard" for molecular interactions, 54 compared to training data for condensed-phase systems which are effectively limited to density functional theory (DFT) calculations. 17 An alternative ML approach to the development of accurate multidimensional PEFs, which are transferable from the gas to the condensed phase, can be rigorously derived from the manybody expansion (MBE) of the energy.…”