Decoupled Multiplicative Physics-Guided Neural Network for Extrapolative Modeling of GaN HEMT

Abstract

Rapid direct-current (DC) surrogate modeling of gallium nitride (GaN) high-electron-mobility transistors (HEMTs) requires accurate prediction of the current-capability, conduction-loss, and self-heating indicators used in geometry screening. This work presents an electrothermal decoupled multiplicative physics-guided neural network (ET-DM-PGNN) for this task. It combines a gray-box backbone, separate high- and low-field polarization channels, a latent DC self-heating update, and bounded target-specific multiplicative residuals. A two-stage strategy first calibrates the complete surrogate and then refines the current branch while preserving resistance and temperature. On 32 interstitial samples, ET-DM-PGNN achieves R2>0.997, with mean absolute percentage errors of 0.13%, 0.15%, and 0.04% for Imax, Ron, and Tmax, respectively. The corresponding errors on nine external cases are 1.46%, 0.94%, and 0.74%. A leakage-controlled, three-seed level-held-out audit yields errors of 0.44–6.40% in four composition and lateral-geometry directions, whereas holding out the 0.015 µm barrier level produces errors of 12.41–21.21% under the tested training protocol. Across 365 geometries, 4 updates differ from a 12-step reference by at most 0.0076%. Numerically validated at Vgs=3 V using TCAD DC sweeps extending to Vds=15 V, the surrogate supports rapid fixed-protocol geometry screening and identifies the ultrathin-barrier direction as the least reliable tested extrapolation regime.