AI surrogate models for lifetime prediction in power electronic modules

Reliability assessment of power electronic modules (PEMs) traditionally relies on counting methods to decompose complex mission profiles (MPs) into equivalent loading cycles, followed by a cumulative damage rule to estimate fatigue life. While effective for simplified load descriptions, this approach neglects nonlinear load interactions and the underlying physics of degradation, often leading to inaccurate lifetime predictions under realistic operating conditions. A more robust solution is autoregressive damage prediction, where the effect of loading is measured sequentially, thus accounting for cycle interactions, sequence effects, and nonlinear damage accumulation. However, this alternative is computationally inefficient as it requires frequent updates on the module's state, usually via slow and complex simulations. In this paper, we present physics-augmented machine learning (ML) solutions to alleviate the computational burden in realistic remaining useful life (RUL) estimation of PEMs. Specifically, we introduce surrogate models that replicate numerical simulations with high fidelity while being 10^6 times faster. Surrogate model results are thoroughly analyzed, providing guidelines on model selection, the number of simulations needed, and optimal data acquisition strategies. Our framework is designed for practitioners with minimal AI experience, and is supplemented with openly available data and source code to facilitate reproducibility and future research.