EXTENDED ABSTRACT: In this study, we compiled existing data on battery materials, constructed a database, and used machine learning to predict lithium compounds with high ionic conductivity. Initially, data on solid electrolyte materials collected from literature were used to establish the AtomWork Battery (AWB) database [1], which includes ion conductivity data for 2,721 solid electrolyte materials. Employing an originally developed periodic descriptor set [2] along with AWB data, we utilized five machine learning models—GPR (Gaussian Process Regression), NGB (NGBoost), XGB (XGBoost), and RF (Random Forest)—to screen approximately 8,000 known Li compounds, identifying 60 candidate compounds with high ionic conductivity. The accuracy and reliability of these proposals were validated using experimental data on ion conductivity published in literature, as illustrated in Fig. 1. Approximately half of these compounds, 28 types, were confirmed to have an ionic conductivity exceeding 10⁻ ⁴ S/m. Additionally, for compounds lacking experimental data, ion conductivity was estimated using the MLFF (Machine Learning Force Field) method. Notably, the fluorite crystal structure of Li1.75Ge0.25As0.75Se0.25 demonstrated a diffusion rate comparable to the known high ionic conductor Li10GeP2S12 at temperatures above 650K.

Acknowledgments: This research was supported by the JST Collaborative Creation Program for Advanced Battery Research
and Development Hub JPMJPF2016.
Keywords: data-driven exploration, Li ionic conductor, battery material database, machine learning, MLFF calculation
REFERENCES:
[1] Y. Xu, Y. J. Wu, H. Li, L. Fang, S. Hayashi, A. Oishi, N. Shimizu, R. Caputo, P. Villars, Science and Technology of Advanced Materials, 25, 2024. DOI: 10.1080/14686996.2024.2403328
[2] Ken Sakaguchi, Yibin Xu, Wenqin Peng, Patent Application Number 2024-12367