Small‐signal modeling of microwave transistors using radial basis function artificial neural network‐comparison of different methods for spread constant determined. Issue 6 (15th March 2022)

• Record Type: Journal Article
• Authors: Qi, Junjun; Lu, Hongliang; Yan, Silu; Zhao, Ranran; Tan, Daidao; Zhang, Yuming; Zhang, Yimen
• Is Part Of: International journal of RF and microwave computer-aided engineering. Volume 32:Issue 6(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-03-15
• Subjects: artificial neural network -- equivalent circuit parameters -- radial basis function -- small signal model; Microwave devices -- Computer-aided design -- Periodicals; Computer-aided engineering -- Periodicals; 621.3813
• DOI: 10.1002/mmce.23145
• Languages: English
• ISSNs: 1096-4290

Abstract

Abstract: This article presents a high‐precision modeling method to build a small signal model of GaAs pseudomorphic high electron mobility transistor (pHEMT) by using a radial basis function artificial neural network (RBF ANN). Both the RBF ANN Method 1 that uniformly distributed spread constant (SC) in the given range and Method 2 that increasingly changed SC with a chosen step have been modeled in this work. Compared with the l RBF ANN Method 1, the RBF ANN method 2 can automatically obtain the optimal ANN corresponding to the SC. The RBF ANN method 2 is developed for S ‐parameters model and equivalent circuit parameters (ECPs) model of HEMTs. S ‐parameters model establishes S ‐parameters versus bias, temperature and frequency, while the ECPs model establishes ECPs versus bias and temperature. To validate the capability of the RBF ANN in small signal modeling of GaAs pHEMTs, measured and modeled data of S ‐parameters model and ECPs model of a 4 × 75 μm gate width, 0.15 μm gate length GaAs pHEMT are compared, and very good agreement is achieved up to 50 GHz. For the S ‐parameters model, the average error of the method 2 is improved by about 20% at the temperature of −20°C, 25°C, and 85°C. For the ECPs model, the average error of the method 2 is 80% higher than that of the method 1.