Magnetic Field–Enhanced Digital Twin for Battery SoC/SoH Estimation

Abstract

This study adopts a two-step approach for battery state estimation in commercial large-format NMC pouch cells. First, a validated baseline for state of charge (SoC), state of health (SoH), and remaining useful life (RUL) is established using physics-based modeling, conventional electrical and thermal measurements, electrochemical characterization, and machine-learning techniques. Second, spatially resolved magnetic-field measurements are introduced as an additional sensing modality to investigate whether they improve observability of internal current-distribution heterogeneity and reduce estimation uncertainty. A physics-constrained recurrent neural network is used for SoH estimation from cycle-level electrical, thermal, and electrochemical health indicators and is evaluated using leave-one-cell-out cross-validation across 15 cells. The validated baseline achieves a per-cell SoH mean absolute percentage error of 0.94%, compared with 1.83% for a non-physics-informed recurrent baseline. A recurrent SoC estimator using time-resolved voltage, current, and temperature measurements achieves a mean absolute error of 1.12 percentage points of SoC. RUL estimation remains less mature because only five cells have reached the defined end-of-life threshold, with the current point-RUL model achieving 26.97% mean absolute percentage error. In parallel, a spatial magnetic-sensor array provides time-synchronized magnetic-field measurements that are incorporated as additional input features. Ongoing analysis compares the established electrical, thermal, and electrochemical baseline with the same estimators augmented by magnetic-field descriptors under an equivalent validation protocol. The work therefore establishes a rigorously validated baseline and an integrated magnetic-field feature pathway, against which the incremental diagnostic value of magnetic sensing is being assessed through a controlled ablation study.