Performance-Based Selection of Machine Learning Interatomic Potentials for Studying Solid-State Electrolytes

High-throughput, accurate prediction of solid-state electrolyte (SSE) properties is essential for advancing all-solid-state batteries (ASSBs). While density functional theory (DFT) can achieve high accuracy, the structural complexity of inorganic superionic conductors makes screening vast compositional spaces computationally prohibitive. Machine learning interatomic potentials (MLIPs) offer comparable accuracy but with orders-of-magnitude lower computational resourcing demands. In this study, we benchmark several pretrained MLIPs on the chemically and structurally complex Li6PS5Cl argyrodite electrolyte, a prototypical SSE. Models are validated against DFT results for structural, energetic, and dynamic properties, and then applied to study the effect of atomic disorder on lithium-ionic transport. Through these studies, three key findings emerge: (1) universal accuracy of MLIP models does not necessarily extend to subtle configurational or compositional changes; (2) nonconservative frameworks, where forces and energies are predicted separately, often fail to capture dynamic behavior; (3) conservative frameworks, which enforce energy-force consistency, better represent physical laws and show superior generalizability beyond the training set. These results provide practical guidance for MLIP model selection by researchers in the ASSB field, and those beyond who investigate highly disordered structures.