Early-cycle life prediction of mechanically damaged lithium-ion batteries via multi-physics features and machine learning

Lithium-ion batteries are susceptible to mechanical damage during actual service, which accelerates degradation and poses a long-term safety risk. Accurate prediction of the service life of mechanically damaged batteries, especially at early degradation stages, is critical for timely safety warning and protection. This study proposes a data-driven framework that integrates multi-physics features for the early life prediction of mechanically damaged batteries. A dataset resembling realistic service scenarios is constructed through multi-condition indentation experiments and simulations. Physically interpretable features, including early-cycle voltage profiles, impedance, and stress features, are extracted to enable lifespan forecasting. The proposed approach enables accurate life prediction across electrochemical-mechanical domains using only a limited number of initial cycles, overcoming the conventional reliance on long-term aging data. Comparative analyses are performed to evaluate the effects of different machine learning models, feature combinations, and the number of cycles used on prediction performance. Results demonstrate that the proposed approach achieves a mean absolute error of about 50 cycles over the first 200 cycles. The research results provide a reliable technical solution for early health diagnosis and operational warning of batteries in electric vehicles and aviation systems.