Beyond Order: Partial Site Occupancies-Informed Machine Learning for Solid-State Electrolytes Design

Crystal structures with partial site occupancies (PSO), a common feature in disordered materials where ions fractionally occupy lattice sites, are crucial for designing solid-state electrolytes (SSEs) with a low activation energy (Ea). However, deciphering the underlying structure–property relationships remains challenging, as traditional ordered-assumption models cannot fully explain PSO effects. Herein, we propose a PSO-informed machine learning (PSO-ML) method for SSEs design, in which PSO knowledge is leveraged by identifying crystal structures from the literature based on identical ionic species and Wyckoff sites criteria, enabling a quantitative correlation between PSO effects and Ea. Applied to trigonal halides Li3YCl6, an Ea prediction model with a determination coefficient (R2) of 93.18% is obtained by partial least-squares analysis with leave-one-out cross-validation. The variable importance in projection analysis identifies the configurational entropy and interlayer distance as key descriptors, both dominated by Y3+ occupancy effects. By optimizing Y3+ occupancy, promising candidates Li3Y4x–3M3–3xCl6 (M: tetravalent cations, 0.75 < x ≤ 0.888) are suggested, where Li3Y0.2Zr0.6Cl6 (room-temperature ionic conductivity of 1.19 mS cm–1) has been experimentally evaluated as an excellent candidate, and other promising compositions are waiting for validation. With high accuracy and interpretability, the PSO-ML method enables the accelerated discovery and design of SSEs and broader disordered materials.