Learning Battery Aging Dynamics Using Physics-Informed Transformer

Reliable estimation of lithium-ion battery capacity degradation is essential for electrified powertrain design and control. This work develops a physics-informed transformer model (PITM) that integrates physics-based degradation trends with multivariate cycling data to learn battery aging dynamics for accurate capacity degradation trajectory estimation across diverse battery chemistries and operating conditions. A physics-based reduced-order model (PB-ROM) is evaluated on 17 cells from five aging datasets to identify degradation-relevant variables and quantify its limits. While the PB-ROM captures long-term degradation patterns, its accuracy degrades under strongly nonlinear conditions, with relative standard error of prediction (RSEP) ranging from 5% to 23%. Guided by these insights, the PITM is developed using dynamic features and trained with a composite loss function that incorporates both experimental capacity measurements and PB-ROM-derived degradation increments. Five single-group PITM models and a generalized multigroup variant (G-PITM) are trained across heterogeneous datasets. The proposed models consistently outperform the PB-ROM, reducing RSEP to as low as 2.96% on unseen test cells and accurately capturing complex degradation behavior. Finally, the PITM is integrated into a series-hybrid range-extender powertrain architecture and evaluated under heavy-duty driving cycles. The PITM demonstrates physically consistent trajectory estimation under realistic load dynamics, confirming its suitability for control-oriented applications.