Mitigating Multiphysics Interference in Semiconductor Aging via Physics-Embedded Incremental Evolution

Remaining useful life (RUL) prediction for power semiconductor devices such as insulated-gate bipolar transistors (IGBTs) is central to reliable power-electronics operation, yet remains challenging because degradation is non-stationary and electro-thermal precursors are strongly coupled. Here, we propose a physics-informed incremental learning framework (PIILF), which models aging as a latent incremental state-evolution process rather than static trajectory fitting. PIILF integrates an incremental state evolution network (ISEN) for state-wise degradation updates, task-adaptive parameter sharing (TAPS) for mitigating cross-task interference among coupled precursors, and a physics-informed observation decoder (PIOD) that reconstructs observables through electro-thermal coupling relations. On the NASA IGBT accelerated aging dataset, evaluated over 100 random seeds, PIILF achieves lower RMSE and MAE than TimesNet, TimeXer, and DeepHPM, while retaining competitive MAPE, a slightly better $R^2$, and higher parameter efficiency. When the training data are reduced to 50% and 25%, PIILF exhibits smaller error increases than the baselines, indicating greater robustness in data-scarce settings. These findings suggest that embedding physical consistency directly into incremental representation learning provides an effective and efficient route to robust semiconductor RUL prediction.