New physics-informed AI cuts SiC power module failure prediction errors by 70%

Silicon carbide (SiC) power modules are critical in automotive traction inverters, but their in-field health is notoriously hard to track. Traditional physics-of-failure models aren't real-time, data-driven approaches need large labeled datasets and generalize poorly, and existing physics-informed methods are too heavy for embedded systems. Researchers Mattia Scarpa (Infineon Technologies) and collaborators from the University of Padova and other institutions tackled this with a novel framework that fuses physics priors with neural networks, specifically for SiC MOSFET modules with sintered packaging. This packaging suppresses solder degradation, creating aging behavior distinct from older solder-based modules—with wirebond liftoff events causing abrupt, non-monotonic perturbations in forward voltage drop (V_DS).

Their solution combines three innovations: physics-informed features that encode cumulative damage from junction temperature swings and a Miner rule accumulator, a monotonicity constraint via gradient penalty regularization to embed expected degradation direction, and a heavy-tailed output distribution for calibrated uncertainty that stays robust to liftoff-induced variance. Tested on an industrial power cycling dataset from Infineon under strict cross-validation, the full configuration reduced mean absolute error by roughly 70% over purely data-driven baselines—with stable performance across all folds. The model stays lightweight enough for embedded deployment, making it practical for real-time condition monitoring in EVs. The paper is an extended version of work accepted at PHM 2026.