Mining ionic conductivity descriptors of antiperovskite electrolytes for all-solid-state batteries via machine learning

Lithium-rich and sodium-rich antiperovskites (X3BA, X = Li, Na) have been explored as promising inorganic electrolytes for all-solid-state batteries in recent years. To accelerate the design and discovery of high room temperature ionic conductivity antiperovskites, in this work, we apply the machine learning (ML) method to mine material descriptors characterizing ionic conductivity directly. Experimental samples are collected firstly from previous research to construct a small dataset (106 samples) supporting the data-driven strategy. After rough classification learning and exact symbolic regression learning, a simple and comprehensive descriptor t/η is proposed showing negative relationships with logarithmic ionic conductivity, where t and η are the tolerance factor and atomic packing factor, respectively. As a case study, we screen candidates in the family of lithium-based nitro-halide double antiperovskites, in which Li6NClBr2, Li6NBrBr2, and Li6NBrI2 are identified as good conductors by the descriptor and showing over 1 × 10−4 S·cm−1 room temperature bulk ionic conductivity in the ab initio molecular dynamics simulations. The descriptor would be convenient to guide the experimental study of antiperovskite electrolytes. Also, our mining strategy is effective in understanding a latent structure-activity relationship from small and complicated data.