Physics-Informed Intelligent Thermal Model for SiC Power Modules With High Generalization

This paper proposes a novel physics informed intelligent thermal model (PIITM) for silicon carbide (SiC) power modules, offering interpretability and generalization compared to traditional ITM. This proposed PIITM framework integrates the thermal transfer logic of each component to provide good interpretability. A residual convolutional neural network (CNN) is employed to extract features of thermal cross coupling (TCC) effect. This algorithm maps junction temperature information directly into the feature space while the attention mechanism (AM) adjusts feature weights, thereby improving training efficiency and learning performance. Moreover, imbalanced training data under high-frequency and high-temperature conditions limits traditional ITM generalization, addressed by the wide-temperature dataset and transfer learning (TL)-integrated algorithm. Based on a wide-temperature dataset derived on a finite element model (FEM), this trained model can accurately describe the nonlinear temperature-dependent effect of SiC power module under high temperature exceeding 150°C. Furthermore, the proposed method achieves excellent generality across different operating frequencies by using TL to fine tune the neural network parameters with few-shot samples. Detailed benchmarks regarding the construction process, accuracy, efficiency, and application scenarios of FEM, lumped thermal model (LTM), and PIITM are provided at the end to emphasize the superiority of the proposed interpretable and generalized PIITM.