On the Role of Long-Range Electrostatics in Machine-Learned Interatomic Potentials for Complex Battery Materials

Carsten G Staacke,Hendrik H Heenen,Christoph Scheurer,Karsten Reuter,Johannes T Margraf,Gábor Csányi

doi:10.1021/acsaem.1c02363

Abstract

Modeling complex energy materials such as solid-state electrolytes (SSEs) realistically at the atomistic level strains the capabilities of state-of-the-art theoretical approaches. On one hand, the system sizes and simulation time scales required are prohibitive for first-principles methods such as the density functional theory. On the other hand, parameterizations for empirical potentials are often not available, and these potentials may ultimately lack the desired predictive accuracy. Fortunately, modern machine learning (ML) potentials are increasingly able to bridge this gap, promising first-principles accuracy at a much reduced computational cost. However, the local nature of these ML potentials typically means that long-range contributions arising, for example, from electrostatic interactions are neglected. Clearly, such interactions can be large in polar materials such as electrolytes, however. Herein, we investigate the effect that the locality assumption of ML potentials has on lithium mobility and defect formation energies in the SSE Li7P3S11. We find that neglecting long-range electrostatics is unproblematic for the description of lithium transport in the isotropic bulk. In contrast, (field-dependent) defect formation energies are only adequately captured by a hybrid potential combining ML and a physical model of electrostatic interactions. Broader implications for ML-based modeling of energy materials are discussed.

Highlights

IntroductionThe expectation here is that flexible machine learning (ML) potentials can overcome long-standing limitations of empirical potentials that use simple fixed functional forms.[7]
The development of new analytical approximation frameworks is currently leading to an unparalleled surge of machine learning (ML) approaches in all areas of chemistry and materials science.[1−5] Here, ML is typically considered a universal approach for learning a complex relationship y = f(x) without explicitly knowing the physical form of f.6,7 In the context of interatomic potentials, this means establishing the relationship between a system’s atomistic structure and its total energy E = f({Z, R}), where Z are the atomic numbers and R are the position vectors of the constituting atoms
A hallmark of many-body ML potentials is the assumption that the total energy can be described as a sum of local atomic contributions, which corresponds to a complete neglect of long-range interactions

Summary

Introduction

The expectation here is that flexible ML potentials can overcome long-standing limitations of empirical potentials that use simple fixed functional forms.[7] Such limitations are especially acute when covalent bonds are formed or broken, when atoms vary their hybridization or charge state, and generally when large changes in chemical environments occur. All these aspects apply prominently to the simulation of operando energy conversion systems in general and battery materials in particular.[3,8−15] With the structural and compositional complexity of contemporary battery materials severely limiting direct first-principles-based simulations, there is considerable hope that ML potentials trained with first-principles data will enable simulations at unprecedented length and time scales and a predictive quality matching that of electronic structure methods.[16−23].

Objectives

Methods

Results

Conclusion