Accelerating the discovery of disordered multi-component solid-state electrolytes using machine learning interatomic potentials

Abstract

Machine learning interatomic potentials (MLIPs) are rapidly emerging as powerful tools for materials simulations, offering a promising pathway to explore complex systems beyond the reach of traditional methods. This study investigates the application of MLIPs focused on the MACE architecture to multi-component, disordered solid-state electrolytes (SSEs), a critical class of materials for the next-generation solid-state batteries. We first benchmark the performance of MACE against established SSE families, Na_1+xZr₂Si_xP_3-xO₁₂ and Li_4xGe_xP_1-xO_4-4xS_4x, confirming their general applicability while identifying key considerations for robust potential development in chemically diverse systems. This workflow, emphasizing the selection of representative configurations, provides critical insights for constructing reliable models in complex, multi-components environments. We further demonstrate the predictive power of this approach by constructing a high-performance MLIP for the novel halide system Li₃In_xY_1-xBr_6yCl_6-6y (x, y ∈ [0,1] and x + y ≥ 1), leading to the identification of the Li₃In_0.5Y_0.5Br₃Cl₃ stoichiometry with the most favorable predicted ion transport properties. By analyzing the MD trajectories generated in this work using our MLIP, we identified two distinct Li-ion migration pathways in this material. The trained model facilitates the computational investigation of intricate mixed cation/anion substitutions in halide SSEs, offering new insights into higher-entropy systems incorporating multi-components. Our results underscore the capability of MLIPs to accelerate the discovery cycle of complex functional materials and provide a robust computational framework for designing advanced SSEs.