A Physics-Informed Multiview Collaborative Semisupervised Framework for Battery Lifespan Early Prediction

Accurate early prediction of lithium-ion battery lifespan is essential for optimizing maintenance and accelerating development, yet traditional approaches require extensive labeled data with complete lifetime measurements. While semisupervised learning offers a promising approach by leveraging abundant unlabeled early-cycle data alongside limited labeled examples, purely data-driven implementations often overfit and fail to reflect physical aging mechanisms. This can lead to predictions that lack physical plausibility and interpretability. To address these issues, we present a physics-informed multiview collaborative semisupervised framework that effectively leverages abundant unlabeled early-cycle data alongside a small labeled set. First, we propose a novel physics-guided multiview feature extraction comprising: 1) a mechanistic aging mode view that quantifies loss of active material and lithium inventory via electrode-level open-circuit potential analysis; 2) a phase transition information view that captures underlying phase-change dynamics through Lorentzian-decomposed incremental capacity curves; and 3) an operating condition-related view that encodes temperature, charging time, and other factors under realistic usage conditions. Second, we embed physical priors into an enhanced Gaussian process regression by designing physics-informed kernels for each view, constraining the regressor to physically plausible degradation pathways. Third, we introduce an uncertainty-aware collaborative semisupervised regression strategy that iteratively generates and filters high-confidence pseudolabels via inverse-covariance weighting across views. Extensive experiments on three datasets demonstrate that our method significantly outperforms both supervised and semisupervised baselines, achieving substantial error reduction with as few as 5–15 labeled cells.