(186a) Automated Solid State Electrolyte Conductivity Predictions Via Probability Density Analysis

The ionic conductivity is the paramount property of the solid state electrolyte (SSE) to enable solid state lithium batteries, but is often multiple orders of magnitude lower than in liquid electrolytes. Current machine learning (ML) models predicting ionic conductivity directly from the crystal structure are limited by the low number of training points and offer little mechanistic insight into diffusion mechanisms for a given crystal. Here, we train a ML model to predict the lithium ion probability density from crystal structure (Figure 1). The probability density qualitatively predicts the ionic conduction pathways and topological analysis provides quantitative predictions of diffusion activation energy barriers and ionic conductivity. We train a convolutional neural network with a Gaussian representation of the crystal structure electron density as the input, and a binarized probability density as the output. The newly proposed model is trained on a dataset of 100 materials with a final training accuracy of 94% and a test accuracy of 91%. By screening the ICSD, we rediscover Li₆NBr₃ which has an experimentally measured ionic conductivity of 0.0035 mS/cm at 150°C and propose a series of conduction pathways which would not have been feasible through density functional theory due to the large number of atoms. Additional crystalline families were found such as LiGaS₂, and Li₂NaB(PO₄)₂ whose ionic conductivities are further optimized through aliovalent substitution and are being verified with ab initio molecular dynamics. Our resulting ML algorithm is applicable over a wide range of compositions, scales efficiently to hundreds of atoms, and can analyze disorder systems making it a powerful new tool to analyze, optimize, and screen for new SSEs.