Optimized Machine Learning-Augmented Hybrid Empirical Models for AlGaN/GaN HEMTs: A Comprehensive Analysis

Precise modeling of gallium nitride (GaN) high-electron mobility transistors (HEMTs) is vital for ensuring reliable and scalable RF circuit design, and efficient characterization of the device behavior. This article presents robust hybrid equivalent circuit (EC)–machine learning (ML) frameworks for significantly streamlining the extraction of small-signal model parameters of AlGaN/GaN HEMTs. The extrinsic and intrinsic parameters of the devices are initially extracted using physics-relevant empirical models in Keysight’s advanced design system. Thereafter, six extensively optimized ML regression models, namely decision tree (DT), ensemble learning (EL), support vector regression (SVR), kernel approximation regression (KAR), Gaussian process regression (GPR), and neural networks (NN) are employed to simulate the intrinsic behavior of GaN HEMTs. The models are trained on GaN HEMTs of geometries 4×100 μm, 10×220 μm, and 10×250 μm, while tested on GaN HEMTs of geometries 2×200 μm and 10×200 μm across diverse biasing and frequency conditions. The input features to models include gate–source voltage (V<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GS</sub>), drain–source voltage (V<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DS</sub>), frequency, number of fingers (NF), unity gate width (W<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</sub>), and effective gate width (W<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">eff</sub>). Finally, a thorough quantitative assessment and detailed comparisons are performed in terms of standard regression tests, mean absolute percentage error, Nash–Sutcliffe efficiency, Kling–Gupta efficiency, training and prediction speed, reliability of model parameters, and simulation agreement with the measured S-parameters. The results demonstrate that among the tested ML models, EL exhibited the lowest mean relative S-parameter errors (2.78–3.75 %), followed by NNs (2.05–6.98 %), DT (2.29–7.35 %), GPR (2.86-8.91 %) SVR (7.83–9.88 %), and KAR (8.26–10.54 %) across diverse GaN HEMTs geometries. This hybrid modeling strategy provides a practical alternative to conventional parameter extraction, offering speed, accuracy, and broader applicability.