Physics-Based and Machine-Learning-Assisted Performance Mapping of Ultra-High-Voltage SiC IGBTs and MOSFETs Under Hard- and Soft-Switching Conditions

Abstract

Ultra-high-voltage SiC IGBTs, SiC MOSFETs and se- ries connection Si IGBT and SiC MOSFETs configurations are compared under hard- and soft-switching conditions using a self-consistent electrothermal framework combined with machine- learning-assisted device mapping. All electro-thermal input data are extracted from finite-element device simulations, so that the comparison is based on physics-derived models rather than sim- plified compact-model assumptions. Three different comparisons are considered. First, maximum-current capability maps are calcu- lated to determine the highest current that each device technology can sustain before reaching the imposed thermal limit. Second, a deterministic electrothermal solver selects the minimum-loss feasible device at each operating point. Third, a machine-learning classifier trained from deterministic solutions is used to accelerate superior-device map generation, while deterministic correction is retained near device-transition boundaries. For the 20 kV voltage class considered here, the classifier-only stage reduces map- generation time by up to 3.8× relative to the full deterministic map. When deterministic correction is applied near low-confidence and transition-boundary regions, the hybrid workflow achieves a 2.1× speed-up excluding the one-time training cost.