Temperature excavation to boost machine learning battery thermochemical predictions

Abstract

Advancing battery technologies requires precise predictions of thermochemical reactions among multiple components to efficiently exploit the stored energy and conduct thermal management. Recently, machine learning (ML) promised to address this complex thermochemical prediction task; however, it failed due to the huge gap between high problem complexity and extremely limited experimental data available for model training. Here, we innovate and validate the temperature excavation (TE) method that interprets the kinetic preferences of thermochemical reactions within minimal experiments into millions of training data. With the help of the TE method, we build the first universally applicable battery thermal runaway model, which achieves high prediction accuracy across a 500°C range on 15 distinct commercial and advanced chemistries with different battery formats and covers all normal working conditions. The TE method also demonstrates broad adaptability and training stability on various ML algorithms, opening new interdisciplinary opportunities for ML in thermochemistry and all thermal-related studies.