High-throughput screening for solid-state Li-ion conductors combining machine learning and first-principles calculations

Beginning with more than 30,000 experimentally reported Li-containing structures drawn from the Inorganic Crystal Structure Database, the Materials Platform for Data Science (Pauling file), and the Crystallography Open Database, we apply a series of automated structural and compositional filters to obtain 1500 unique crystal structures suitable for electronic-structure calculations which yields nearly 1,000 electronic insulators.

We then estimate Li-ion diffusivities for these insulating candidates using molecular dynamics simulations at multiple temperatures. To make simulations computationally feasible at this scale while preserving near first-principles fidelity, we employ a foundational machine-learned interatomic potential which is carefully fine-tuned on relevant Li-chemistries. We discuss the details of the fine-tuning strategies and data-consistency considerations required to obtain a very accurate and robust model.