Machine-learning-assisted EEHEMT compact modeling of GaN HEMTs

Abstract

We present a machine learning approach that enhances the EEHEMT model for predicting current–voltage (I–V) characteristics of AlGaN/GaN high electron mobility transistors (HEMTs), focusing on the normalized transconductance-to-current characteristic $g_m/I_D$. Instead of replacing model equations, an artificial neural network (ANN) is used to predict key parameters governing the EEHEMT model’s piecewise segments, enabling segmentation adaptive to device geometries and bias conditions. To enhance the accuracy and continuity of first-order derivative of drain current ($g_x=\partial I/\partial V_x$) while maintaining physically reasonable model behavior, we employ a physics-guided hybrid loss function that combines current fitting error, $g_x$ fitting error, a smoothness regularization term on $g_x$, and a monotonic constraint on the breakpoints. The ANN-assisted EEHEMT accurately predicts device characteristics and key metrics such as $V_{th}$, $R_{on}$, and $I_{dmax}$. This approach mitigates derivative discontinuities in $g_x$, preserves model interpretability, and enhances generalization across device geometries. Notably, it accurately predicts $g_m/I_D$, offering valuable insights into GaN HEMT behavior and enabling efficient, design-oriented device optimization.