Computational and Data-Driven Discovery of Li Solid-State Electrolytes: From Representation to Experimental Realization

Improvements in energy storage are required to facilitate the transition to renewable energy and the electrification of transport. Lithium-ion batteries (LIBs) are a promising solution, but the current leading chemistry, consisting of a layered oxide cathode and a graphite anode separated by a liquid electrolyte, has been optimized to near-theoretical limits. Replacing the graphitic carbon with Li metal would significantly improve energy density but the instability of the Li metal-electrolyte interface introduces performance and safety challenges. Using a solid-state electrolyte (SSE) to construct an all-solid-state battery (ASSB) could mitigate these issues. However, an ideal SSE material has yet to be identified. Thousands of known Li-containing materials have not yet been evaluated as SSEs. Data-driven methods could prioritize materials for experimental study but have historically lacked sufficient data and optimal representations. Chapter 2 presents the largest structure-ionic conductivity database to date and uses semi-supervised learning to determine the highest-performing descriptors. From ~26,000 Li-containing materials, 212 candidates are identified and screened using semi-empirical and first-principles calculations. Li3BS3 exhibits ionic conductivity above 10-3 S cm-1 with defect engineering through substitution and mechanical milling. Chapter 3 explores Cl, Al, and Si substitution in Li3BS3 to reveal mechanisms of ionic conductivity enhancement. At low substitution levels, conductivity improvements are driven by disordered environments from reduced crystallinity and microstructural effects. For Cl and Al, higher substitution generates fully amorphous phases with ionic conductivity above 10-4 S cm-1. Sufficient Si substitution produces novel crystalline phases with conductivities exceeding 10-3 S cm-1. Previous approaches, such as that in Chapter 2, could not represent disordered compounds, excluding much of the training data and candidate materials. This is particularly significant given the importance of disorder highlighted in Chapters 2 and 3 and the prevalence of disorder in known superionic conductors. Chapter 4 implements a transfer-learned graph representation compatible with disordered structures. A larger database is curated and used to train models for screening all known Li-containing materials. Experimental validation of superionic conductivity in an identified candidate demonstrates the utility of this graph-based approach for discovering experimentally relevant, high-performance materials.