Leveraging Physics and Machine Learning to Predict Battery Aging

A primary concern with lithium-ion battery technology is the performance deterioration over time. Capacity retention, a vital performance measure, is frequently utilized to assess whether these batteries have approached their end-of-life. Machine learning offers a powerful tool for predicting capacity degradation, leveraging both past data and physical knowledge in the form of simulations or phenomenological models. In this study, we showcase the utility of probabilistic machine learning and transformer-based deep learning modeling in battery health prediction. For our probabilistic models, which operate without pre-training, we employ: i) a structured Gaussian process (GP) - an enhanced version of the standard GP that integrates a phenomenological model as a probabilistic prior mean function, and ii) a multi-fidelity GP, augmented by prior physics-based simulations to boost its predictive power. Concurrently, we deploy time-series transformers pre-trained on existing datasets for forecasting purposes. Our findings juxtapose these models, offering insights into their optimal application scenarios, and the associated codes are open for access.