Physics-Inspired Transformer for Remaining Useful Life Prediction of Lithium-Ion Batteries

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is a prerequisite for reliable battery management, predictive maintenance and safe second-life utilization. Existing approaches fall into two families. Physics-based models are interpretable but depend on precisely identified electrochemical parameters and are computationally expensive, whereas data-driven models such as the Transformer capture long-range temporal dependencies yet remain physically unconstrained, offer limited interpretability, and generalize poorly when labeled degradation data are scarce. This paper proposes a physics-inspired Transformer, termed PI-Transformer, in which a shared sequence encoder serves two heads: a primary head that regresses the RUL and an auxiliary head that emits a per-cycle degradation curve on which a first-order decay law is imposed as a soft constraint in the log domain. The constraint therefore regularizes the shared representation through an unsupervised auxiliary task rather than the regression target itself, and it is discarded after training, so that the deployed model carries no additional inference-time cost. Within the BatteryML framework, four variants are compared on the MATR-1 fast-charging dataset using QdLinear features: a Transformer decoder baseline, a Transformer encoder, a causally masked decoder and the proposed physics-inspired decoder. The proposed variant reduces the test RMSE by 20.9% with respect to the unconstrained baseline, from 84.01 to 66.46 cycles; over 10 random seeds it attains 68 ± 8 cycles against 84 ± 9 for the baseline, and the ranking of the four variants is unchanged.