Accelerated electro-thermo-mechanical co-simulation of power modules using reduced-order modeling and neural network-based field reconstruction

Traditional electro-thermo-mechanical co-simulation methods, which couple SPICE-based electrical solvers with finite element thermal and mechanical analyses, are highly accurate but computationally prohibitive, typically requiring several hours per operating point. This limitation makes these approaches difficult to use for real-time health monitoring, iterative design processes, and future mission profile simulations. To address these limitations, we propose a multi-physics reduced-order framework combining three complementary components: the Thermally Adaptive Parameterized Loss Model (TAPLM), a novel temperature-aware polynomial loss model dynamically re-parameterized at each iteration of the electro-thermal coupling loop using full-module Double Pulse Simulations; a Foster network for rapid electro-thermal coupling and per-die junction temperature prediction; and a fully connected neural network trained on FEM datasets to reconstruct high-resolution 2D thermal and mechanical displacement fields directly from per-die power losses. Together, these three components form an end-to-end surrogate chain that, given an electrical operating point, delivers full 2D thermal and mechanical fields without any FEM call at inference time. The framework is validated on a 12-die SiC power module under nominal and degraded operating conditions, achieving maximum field reconstruction errors below ± 1.3 % and a computational speedup exceeding 95% relative to the full SPICE/FEM reference workflow. The contribution of this work lies in the integration of a thermally adaptive power-loss model with an iterative electro-thermal coupling and FEM-trained surrogate reconstruction workflow for rapid prediction of 2D temperature and displacement fields in SiC multi-die power modules. While experimental validation remains an open item for future work, these results demonstrate the accuracy, robustness, and efficiency of the proposed approach for real-time health monitoring, Remaining Useful Life (RUL) estimation, and active thermal management through dynamic load balancing among dies.