Physics-Informed Condition Monitoring of SiC Power Modules

What the math gives to ML

The paper offers a transferable physics-informed mechanism for learning degradation under regime changes and abrupt outliers: expose cumulative damage history as an input, enforce the physically expected monotonic degradation direction through a gradient penalty, and predict a heavy-tailed conditional distribution rather than a point estimate. The strongest neural-network transfer is a monotone prognostics head attached to an MLP, RNN, or transformer, with a Miner-style damage state computed from operating histories and a Student-t likelihood for robustness to wirebond-liftoff-like events. The mechanism makes falsifiable predictions: degradation predictions should be nondecreasing under controlled increases in accumulated damage, and Student-t likelihoods should retain calibration and reduce error specifically on heavy-tailed residuals and out-of-distribution regimes.