Physics-Guided Dual-Branch Learning for Accurate SiC MOSFET Junction Temperature Estimation

Accurate junction temperature (Tj) estimation is essential for ensuring the reliability and longevity of SiC MOSFETs. However, existing multi-temperature-sensitive electrical parameters (TSEP) fusion approaches often rely on heuristic parameter selection without a rigorous framework, leading to suboptimal accuracy. To address this, this paper introduces a mutual-information-based method to systematically evaluate and select the most informative TSEPs. In addition, to improve robustness against measurement noisewhich is particularly problematic due to the inherently low signal-to-noise ratio (SNR) of TSEPs-synthetic noise of varying levels is injected into experimentally acquired switching waveforms during training. This enables the proposed model to maintain high prediction accuracy under diverse interference conditions. Furthermore, to overcome the sensitivity of certain TSEPs, such as VDS,on, to variations in gate drive voltage, a physics-informed learning architecture is designed to isolate the influence of gate voltage, thereby enabling consistent temperature estimation across different driving conditions. Experimental validation confirms that the proposed method achieves an RMSE < 0.271 under varying gate voltages and retains strong performance under noise-perturbed scenarios, demonstrating both accuracy and adaptability for real-world applications.