Navigating materials chemical space to discover new battery electrodes using machine learning

Abstract

Investigating the role of electrodes' physiochemical properties on their output voltage can be beneficial in developing high-performance batteries. To this end, this study uses a two-step machine learning (ML) approach to predict new electrodes and analyze the effects of their physiochemical properties on the voltage. The first step utilizes an ML model to curate an informative feature space that elucidates the relationship between physiochemical properties and voltage output. The second step trains an active learning model on the informative feature space using Bayesian optimization to screen potential battery electrodes from a dataset of 3656 materials. This strategy successfully identified 41 electrode materials that exhibit good electronic conductivity and host highly electronegative anions. This work provides an efficient strategy to discover novel electrode materials while integrating domain knowledge of chemistry and material science with ML in materials research.