Improving Fast Charging Safety With Core Temperature Estimation Via Kolmogorov-Arnold Network

What the math gives to ML

The paper offers a transferable robust control-barrier-function mechanism rather than merely a battery-specific controller: a learned estimate of an unmeasured safety state is inserted into a higher-order barrier constraint, and a quadratic program minimally modifies the nominal control while accounting for bounded estimation and model errors. The key guarantee is that replacing the unknown disturbance term by a certified lower bound preserves forward safety when the robust barrier inequality holds. This can be transferred to neural state-space models, world models, and safe reinforcement learning as an inference-time safety projection layer, with an explicit falsifiable prediction: safety violations should remain absent whenever the empirical residual disturbance stays above the certified bound, while violations should begin when that bound is underestimated.