Physics-Guided Machine Learning Assisted Characteristic Prediction and Optimization of p-GaN Gate HEMTs

Abstract

In this study, we demonstrate the feasibility of predicting and optimizing GaN-based high-electron-mobility field-effect transistors (GaN HEMTs) devices using the physics-guided machine learning (PGML) method. This paper presents a physics-guided artificial neural network (PG-ANN) comprising three networks: Para-net, Vol-net, and G-net, which are trained on a dataset generated through Technology Computer-Aided Design (TCAD) simulations. Our approach emphasizes the importance of first-order derivative characteristics ($g_x=\partial I/\partial V_x$) correlated with physical parameters for enhanced accuracy of IV curve predictions and employs a physics-based loss function to guide the PG-ANN toward accurate solutions. Using PG-ANN, we present the DerivNet model which accurately predicts device characteristics and captures key performance indicators, the threshold voltage $V_{\mathrm{th}}$, on-state resistance $R_{\mathrm{on}}$, and the maximum drain current $I_{\text{dmax}}$. The PGML method has the potential to significantly expedite device process optimization and is a promising numerical methodology to assist the modeling framework in Design Technology Co-Optimization (DTCO) flow.