Accelerating solid-state battery design: predicting ionic conductivity with machine learning potentials

Abstract

Achieving high ionic conductivity in solid-state electrolytes (SSEs) is critical for next-generation batteries, yet exploring the vast chemical space is hindered by the computational expense of first-principles methods. To accelerate this discovery process, we developed a high-fidelity machine learning potential (MLP) for the Li–P–S–Ge–Sn system. The MLP demonstrates excellent agreement with DFT calculations for both energies and atomic forces. We then evaluated its predictive performance across multiple tasks, including diffusion barrier calculations and diffusion coefficient extraction, and obtained consistently accurate results. Critically, large-scale molecular dynamics (MD) simulations using this potential predicted room-temperature ionic conductivities for eight prototypical SSEs that correlate strongly with experimental values (R² = 0.93). Leveraging this predictive power, we identified two key strategies for performance enhancement: optimizing lithium vacancy concentrations and tuning the host framework via partial Ge/Sn substitution. Our simulations show that these modifications enhance conductivity by creating more favorable diffusion pathways and increasing structural disorder. This work establishes an efficient computational engine for the rapid and accurate screening of SSE compositions, providing a powerful tool for the rational, data-driven design of next-generation solid-state batteries.