Efficient neural network modeling methodology of SiC MOSFET for accurate circuit simulation based on transfer learning

Effective compact models of silicon carbide (SiC) MOSFETs are necessary for the development and optimization of emerging power electronic applications with SiC devices. In this paper, an efficient neural network modeling methodology of a SiC MOSFET is proposed for accurate behavioral modeling in circuit simulation. Instead of empirical formulas, artificial neural networks (ANNs) are used in this article to describe all the device characteristics, including the first- and third-quadrant output characteristics, three interterminal parasitic capacitance characteristics, and device temperature effects. Considering the common absence of measured data in partial operation regions, simulation data are employed to provide ANNs with behavior knowledge of SiC MOSFETs using a transfer learning method for better modeling performance. In addition, the device temperature effects are independently modeled for flexibility. For verification, the simulated and measured switching waveforms of double-pulse tests are compared to validate the accuracy of dynamic characteristics. Inverter simulation is also conducted to validate convergence. The results indicate that the proposed model demonstrates superior performance to comparative models in terms of accuracy. Furthermore, it exhibits satisfactory simulation speed, generalization ability, and convergence ability. Therefore, the proposed method facilitates efficient modeling of SiC MOSFETs and is feasible for accurate simulation of circuit applications.