Accelerated degradation framework of lithium-ion batteries with physics-informed domain shift learning

Abstract

This study proposes an accelerated degradation framework for lithium-ion batteries (LIBs) utilized in portable devices. The proposed framework incorporates three key features to reduce experimental time and cost and predict the state of health (SOH) and remaining useful life (RUL) in the design phase under various operational environments. First, the neural network incorporates distinct time-dependent features to reflect the multi-stage charge protocols. This process captures nonlinear degradation characteristics of LIBs under various conditions. Second, a physics-informed training strategy integrates a governing equation that represents battery degradation and a physics loss coefficient that gradually increases over epoch. This strategy promotes physically consistent degradation pathways even with limited data because it facilitates balanced learning from both data and physics supervision. Third, the architecture of the surrogate neural network is addressed with a domain shift learning strategy. This feature enables neural network interpolation and extrapolation capabilities across environments, enhancing adaptability to unseen conditions. Analysis of accelerated degradation experiments using lithium cobalt oxide cells indicates that the framework provides competitive results in predicting SOH and RUL compared to data-driven and empirical methods, particularly under out-of-distribution conditions. The results suggest that experimental time is reduced without significant losses in accuracy, lessening the requirement for long-term degradation experiments. These findings support decision-making in the design phase of LIB applications.