Physics-Constrained Deep Operator Network for Remaining Useful Life Prediction of Li-ion Batteries with Early-Cycle Data

Authors:

Abstract
Accurate remaining useful life (RUL) prediction of lithium-ion batteries (LIBs) is critical for safe, reliable, and cost-effective operation of battery-powered systems. Existing data-driven approaches often require long-term cycling histories and may produce physically inconsistent predictions, while physics-based models can be limited by complexity and parameter-identification challenges. This paper proposes a physics-constrained deep operator network (PC-DON) for early-cycle battery RUL prediction. The framework uses a branch network to learn cell-specific degradation patterns from early-cycle incremental capacity curves and a trunk network to represent cycle-level degradation states. Three physics-based constraints are designed and evaluated to enhance physical consistency and generalization. The method is tested on 124 commercial LIB cells and compared with a benchmark physics-informed neural network. Results demonstrate improved accuracy and robustness. Additional analyses of physics-loss combinations, early-cycle selection, and SHAP-based feature importance confirm the framework’s effectiveness, interpretability, and potential for early-stage battery health prognosis.

This publication has 17 references indexed in Scilit: