Evaluating and Interpreting the Predictive Power of Features in Battery Lifetime Prediction

Data-driven early prediction of lithium-ion battery lifetime accelerates performance evaluation and management, yet model interpretability remains limited. We evaluate the predictive power of physics-based features using an aging dataset where a single cycling protocol is used and variability is induced by varying the formation protocols. By forward-simulating the open-circuit voltage and resistance-based degradation modes, features generated using early cycles are linked to specific degradation mechanisms. In particular, we find that low-rate features generated from periodic diagnostic cycles are more predictive because they capture the main degradation mode in the dataset, lithium inventory loss. Our findings enhance model interpretability and highlight the necessity of carefully designed diagnostic cycles that capture various degradation behaviors for accurate battery cycle life prediction.