Robust state-of-health estimation for lithium-ion batteries under stochastic loading: A physics-informed Bayesian continuous-time approach

We estimate state of health (SOH), rather than safety, from five temporally causal 600-s block features: mean voltage, block Ah throughput, voltage variance, mean temperature, and cumulative Ah. The proposed Shadow-Mode CfC uses SPMe simulations only as an initialization prior and is subsequently fine-tuned on measured data. Under the fixed paper protocol with untreated reference-capacity labels and no prediction smoothing, the fixed-protocol Shadow CfC obtains RMSE 0.0320 and MAE 0.0245 on RW11. Under a separate validation-selected protocol, the validation-selected Shadow CfC obtains RMSE 0.0360 in the common baseline comparison. External leave-one-cell-out tests on CALCE and Oxford are reported as dataset-specific stress tests, not as safety certification or universal transfer. Monte Carlo dropout variability is reported descriptively because its empirical coverage is insufficient to support calibrated confidence-interval claims.