Physics-Informed Deep Neural Network for Predictive Modeling and Optimization of ScAlN/GaN HEMTs

Abstract

ScAlN/GaN high-electron-mobility transistors (HEMTs) provide a promising platform for high-power and high-frequency electronics because Sc incorporation can modify polarization, two-dimensional electron-gas (2DEG) density and threshold voltage. However, direct data-driven neural networks can reproduce device characteristics without preserving the physical relationships among composition, barrier thickness, polarization charge, carrier density and transport. This work proposes a physics-informed deep neural network (PIDNN) framework for ScxAl1−xN/GaN HEMTs in which an analytical device model supplies physics constraints while a fully connected neural network learns nonlinear residual behavior. The network uses Sc composition, barrier thickness, gate voltage and drain voltage as principal inputs and predicts sheet carrier density, threshold voltage and drain current. A composite loss combines normalized prediction error with residuals derived from polarization charge, channel charge and current-transport equations. Non-negativity, continuity and monotonicity constraints are introduced to suppress nonphysical predictions. An inverse-design layer identifies composition and barrier-thickness combinations satisfying target current and threshold-voltage specifications. The proposed framework retains the interpretability of analytical modeling while providing the computational efficiency of deep learning for rapid design-space exploration. Illustrative results demonstrate the intended evaluation workflow through parity, error, transfer-characteristic and composition–thickness optimization plots. The framework can subsequently be extended to temperature, self-heating, trapping and RF behavior after calibration against experimental or TCAD data.