Combining Intelligence With Rules for Device Modeling: Approximating the Behavior of AlGaN/GaN HEMTs Using a Hybrid Neural Network and Fuzzy Logic Inference System
oleh: Ahmad Khusro, Saddam Husain, Mohammad S. Hashmi
| Format: | Article |
|---|---|
| Diterbitkan: | IEEE 2024-01-01 |
Deskripsi
This paper uses the Adaptive Neuro-Fuzzy Inference System (ANFIS) to investigate and propose a new alternative behavioral modeling technique for microwave power transistors. Utilizing measured I-V characteristics, associated parameters like transconductance <inline-formula> <tex-math notation="LaTeX">$(g_{\text {m}})$ </tex-math></inline-formula> and output conductance <inline-formula> <tex-math notation="LaTeX">$(g_{\text {ds}})$ </tex-math></inline-formula>, etc., S-parameters characteristics, and RF performance parameters such as unity current gain frequency <inline-formula> <tex-math notation="LaTeX">$(f_{\text {T}})$ </tex-math></inline-formula>, maximum unilateral gain frequency <inline-formula> <tex-math notation="LaTeX">$(f_{\max })$ </tex-math></inline-formula>, ANFIS-based behavioral models are developed for Gallium Nitride (GaN) High Electron Mobility Transistors (HEMTs) and validated. The models have been developed using two distinct devices with dimensions of <inline-formula> <tex-math notation="LaTeX">$10\times 200~\mu m$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$10\times 250~\mu m$ </tex-math></inline-formula> for multi-bias conditions and over a broad frequency range (0.5 to 43.5 GHz). Subsequently, the proposed model performance is validated on devices with geometries of <inline-formula> <tex-math notation="LaTeX">$10\times 220~\mu m$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$4\times 100~\mu m$ </tex-math></inline-formula>, and <inline-formula> <tex-math notation="LaTeX">$2\times 200~\mu m$ </tex-math></inline-formula> to examine the interpolation accuracy, extrapolation potential, and scalability. Here, ANFIS utilizes the subtractive clustering method to process the measurement characteristics by computing the clusters and opts for the best-performing model using error and number of fuzzy rules as criteria. The parameters involved in the fuzzy representation are trained using neural network algorithms, namely gradient-descent and least squares estimate. The proposed models are subsequently incorporated in a commercial circuit simulator (Keysight’s ADS) and the class-F power amplifier’s gain and stability characteristics are computed and studied.