Machine-learned interatomic potentials for battery materials: from fundamental methodology to emerging applications in electrodes, electrolytes, and interfaces.

Abstract

The rapid expansion of battery technologies beyond conventional lithium-ion systems has created an urgent demand for predictive, atomistic-scale simulation tools capable of addressing increasingly complex materials and interfaces. While first-principles calculations have played a central role in elucidating fundamental properties such as redox potentials, phase stability, and ion diffusion, its computational cost severely limits accessible system sizes and timescales. As a consequence, large-scale phenomena including defect-mediated transport, interfacial reactions, and mesoscale structural evolution remain challenging to investigate within a purely first-principles framework. Machine-learned interatomic potentials (MLIPs) have recently emerged as a transformative approach that bridges the gap between quantum-mechanical accuracy and large-scale molecular dynamics simulations. By learning from first-principles reference data, MLIPs enable efficient prediction of energies, forces, and stresses while preserving accuracy comparable to first-principles calculations. In this review, we provide a systematic overview of MLIP methodologies, based on descriptor models (hand-designed and learnable descriptors), and selection of machine learning algorithms. We further examine their growing impact in battery materials research, covering electrodes, solid electrolytes, and their interfaces. The MLIP developments open new possibilities for simulating ion transport, defect chemistry, and interfacial reactivity across diverse chemistries including all solid-state batteries and/or Li, Na, K, and multivalent systems. Finally, we discuss current limitations, validation strategies, and future directions toward robust, universally applicable MLIP frameworks that can accelerate the discovery and rational design of next-generation battery materials.