From equation-centric modeling to transferable representation learning in battery lifetime prediction

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is fundamental to safety assurance, lifecycle management, and rational resource utilization. Over the past four decades, battery lifetime prediction has undergone a structural methodological evolution. This review reveals a transition from equation-centric degradation modeling toward representation-driven and transferable learning frameworks. Early studies relied on handcrafted electrochemical descriptors and statistical pattern recognition for qualitative screening. The subsequent shift to quantitative prediction established three foundational equation-based paradigms: empirical cycle-capacity extrapolation, discharge-event damage accumulation under dynamic duty cycles, and electrochemical-parameter-based simulations grounded in transport processes and side-reaction kinetics. Although physically interpretable, these approaches were largely static and constrained by parameter identifiability and stress-dependent variability. Later developments integrated Bayesian filtering and data-driven learning to reformulate degradation prediction as an adaptive state-estimation problem with uncertainty quantification. More recent advances emphasize representation learning, where high-dimensional electrochemical signals are transformed into transferable degradation features, and cross-domain transfer mechanisms mitigate data scarcity and distribution shifts across heterogeneous cells and operating conditions. We argue that the central challenge has shifted from deriving explicit degradation equations to constructing robust degradation representations that remain consistent with electrochemical principles while generalizing across conditions. Future progress will depend on field-grounded fleet-scale representation learning, physics-informed frameworks that reconcile mechanistic fidelity with data efficiency, adaptive and privacy-preserving transfer under distribution shift, and trustworthy deployment enabled by calibrated uncertainty, active learning, and mechanism-consistent explainability.