Assessment and Application of Universal Machine Learning Interatomic Potentials in Solid-State Electrolyte Research

Abstract

High-performance solid-state electrolytes (SSEs) are crucial for next-generation lithium batteries. However, conventional methods like density functional theory and empirical force fields face challenges in computational cost, scalability, and transferability across diverse systems. Machine learning interatomic potentials (MLIPs) offer a promising alternative by balancing accuracy and efficiency. Nevertheless, their performance and applicability for SSEs remain poorly defined, limiting reliable model selection. In this study, we benchmark 12 MLIPs─including GRACE, DPA, MatterSim, MACE, SevenNet, CHGNet, TensorNet, M3GNet, and ORB─across energies, forces, phonons, electrochemical stability, thermodynamic properties, elastic moduli, and Li⁺ diffusivity. GRACE-2L-OAM, MACE-MPA, MatterSim, DPA-3.1-3M, and SevenNet-MF-ompa show superior accuracy. Using MatterSim, we study Li₃YCl₆ and Li₆PS₅Cl, revealing that ∼40–50% S/Cl anion disorder enhances Li⁺ migration connectivity in Li₆PS₅Cl, while higher Li⁺ content in Li₃Ycl₆ expands conduction channels and reduces energy barriers. These insights highlight the power of MLIP-driven simulations for mechanistic understanding and rational design of high-conductivity SSEs.