A Physics-Informed Deep Learning Framework for Performance Optimization of AlGaN/GaN HEMTs in High-Frequency RF Applications

Abstract

AlGaN/GaN High Electron Mobility Transistors (HEMTs) are essential for high-frequency RF applications. However, designing HEMTs is difficult due to the complexity of the physical relationship between internal charges, device geometry, and reliability. The available standard simulation tools (TCAD) are accurate; however, they are too slow for testing thousands of design variations. In this study, we developed a Physics-Informed Neural Network (PINN) architecture to speed up this process. Initially, a reliable dataset is developed using calibrated TCAD simulations. During the training period, Poisson’s equation residuals and the charge conversions were used. It generates consistent predictions of transconductance, breakdown voltage, cut-off frequency, and drain current. The proposed framework achieves a low prediction error and a significant reduction in the computational cost.