Improved Verilog-A Based Artificial Neural Network Modeling Applied to GaN HEMTs

This study presents a novel approach to implementing an artificial neural network (ANN) model for simulating high electron mobility transistors (HEMTs) in Keysight ADS through integrating Verilog-A coding. It streamlines the realization of ANN models characterized by diverse complexities and layer structures. The proposed method is demonstrated by developing nonlinear models for GaN HEMT on two distinct substrates. GaN-on-Si and GaN-on-SiC with respective 10x200 mu m$10\times 200\nobreakspace \umu{\rm m}$ and 8x125 mu m$8\times 125\nobreakspace \umu{\rm m}$ gate widths are characterized by S-parameters at a grid of gate and drain bias conditions. The intrinsic gate capacitance and conductances are extracted from the de-embedded S-parameters, which are then integrated to find the gate charges and currents. The drain current with the inherent self-heating and trapping effects is modeled based on the pulsed IV measurement at well-defined quiescent voltages. Subsequently, the related ANN models of these nonlinear elements are interconnected to form the intrinsic part of the large-signal model. This intrinsic part with all ANN sub-models is then completely implemented using a Verilog-A-based code. The whole ANN large-signal model is then validated by single- and two-tone radio frequency large-signal measurements, which shows a perfect fitting with a high convergence rate. The overall simulation time is five times reduced when the developed Verilog-A-based ANN is used instead of the table-based model. Overall, the large-signal Verilog-A-based ANN model exhibits an improved performance enhancement compared to the conventional table-based models. This indicates the practical viability of the Verilog-A integration technique in modeling the nonlinear GaN HEMTs.