UltraFlux IDM-Learn: Decoupling operating-condition variability from intrinsic battery degradation under non-stationary fast charging

Reliable state-of-health (SOH) tracking for lithium-ion batteries under dynamic operating regimens provide critical information in overcoming the bottlenecks for sustainable electrification. Conventional data-driven prognostics approaches often lack physical interpretability, leading to spurious correlations between statistical operational features. This limitation can conflate irreversible thermodynamic aging processes with reversible kinetic polarization effects—a distinction that is essential for accurate degradation modeling. Such conflation may induce non-physical oscillations in predicted health trajectories, undermining prognostic reliability. Here, we present the intrinsic degradation manifold learning (IDM-Learn) framework to overcome this structural vulnerability. Controlled temperature comparisons, independent degradation measurements, and perturbation–recovery tests support interpreting the model-derived components as degradation-associated and condition-responsive within the evaluated experimental setting. It achieves effective pathway homogenization. Compared with the baseline, IDM-Learn reduces the loss from 0.0283 to 0.0081, RMSE from 0.0557 to 0.0071, MAE from 0.0444 to 0.0051, and RPE from 0.3571 to 0.0079, corresponding to reductions of 71.4 %, 87.3 %, 88.5 %, and 97.8 %, respectively. These results support the predictive value of IDM-Learn for SOH estimation within the evaluated datasets and operating conditions. Further independent validation is required to assess its applicability across battery chemistries, temperature domains, and deployment environments.