Forecasting battery degradation trajectory under domain shift with domain generalization

Abstract

Rechargeable batteries play a pivotal role in the transition towards a carbon-neutral future by electrifying transportation and mitigating the intermittency of renewable energies. Forecasting the degradation of batteries is crucial for maximizing their usage. However, predicting battery degradation is not trivial due to complex failure mechanisms and diverse working conditions and chemistries. To tackle this challenge, we develop a deep learning model by leveraging meta-learning-based and task-driven domain generalization techniques. Our model delivers superior performance in forecasting battery degradation trajectories on 182 cells of different chemistries and conditions with few or no target-domain training data. Moreover, we explore the interpretability of the model and find that the model might gain generalization capability by learning domain gaps that are similar to the differences of degradation modes behind various operational conditions. Collectively, our work not only showcases the promise of the high-reliability data-driven model in diverse conditions and chemistries, but also spotlights the potential interplay between artificial intelligence and domain knowledge.