Identification of Potential Solid-State Li-Ion Conductors with Semi-Supervised Learning

All-solid-state batteries (ASSBs) are a promising technology to enable enhanced safety, energy density, and power density over conventional Li-ion batteries [1]. Crucial to the operation of ASSBs is a stable Li-ion solid-state electrolyte (SSE) with ionic conductivity comparable to that of conventional liquid electrolytes. Although an SSE with these required properties has yet to be discovered, thousands of known Li-containing materials remain unexplored. In this work, we utilize a semi-supervised learning approach to accelerate the identification of superionic conductors.<br/><br/>We construct a repository containing the experimental ionic conductivities of 1,346 compounds digitized from over 300 publications. Each of the ~26 000 Li-containing materials from the Inorganic Crystal Structure Database and Materials Project is represented using 180 unique descriptor-simplification combinations. Agglomerative clustering is performed, and experimental conductivity values are used as labels to evaluate the effectiveness of each descriptor. A particular spatial descriptor, Smooth Overlap of Atomic Positions, is found to minimize composite intracluster conductivity variance, indicating a strong correlation to ionic conductivity. By examining high-conductivity clusters, 212 previously unexplored candidate materials are identified. Through additional screening using bond-valence site energy and nudged-elastic band calculations, Li₃BS₃ is chosen as a promising compound for experimental validation. We demonstrate that with defects engineered through aliovalent substitution and high-energy ball milling, Li₃BS₃ exhibits a room-temperature conductivity greater than 1 mS cm^-1. The described semi-supervised learning method for assessing descriptor performance and identifying novel materials can be broadly applied for expediting materials discovery.<br/><br/>[1] Janek, J.; Zeier, W. G. A Solid Future for Battery Development. <i>Nat. Energy</i> 2016, <i>1</i> (9), 16141. https://doi.org/10.1038/nenergy.2016.141.