Combining Intelligence With Rules for Device Modeling: Approximating the Behavior of AlGaN/GaN HEMTs Using a Hybrid Neural Network and Fuzzy Logic Inference System

Abstract

This research presents a smart modeling technique for high-frequency GaN-based transistors by combining neural networks with fuzzy logic (specifically, an Adaptive Neuro-Fuzzy Inference System — ANFIS). Instead of relying on conventional circuit-based transistor models, the authors develop a data-driven behavioral model that uses real measurement data such as current-voltage (I-V) characteristics, S-parameters, and key RF metrics,

The model is trained using data from two device sizes (10×200 μm and 10×250 μm), and it’s validated on additional devices (e.g., 10×220 μm) to test scalability and generalization. To enhance performance, the ANFIS structure is optimized using subtractive clustering, which helps choose the right number of fuzzy rules. Parameter training is handled by a combination of gradient descent and least squares methods.

Once trained, the model is implemented in the Keysight ADS simulator and evaluated by simulating a class-F power amplifier. The results show the model can accurately predict transistor behavior across different configurations — offering both speed and precision, which makes it well-suited for modern RF circuit design tasks.