A Gray-Box Equivalent Neural Network Circuit Small-Signal Modeling Applied to GaN Transistors

Efficient transistor modeling is an essential step toward improved fabrication processes and reliable circuit design. This places more pressure on the model developer to consider the physical relevance, accuracy, and simulation convergence rate. This paper proposes an improved gray-box modeling technique that combines the accuracy of the neural network (NN) technique with the physical insight of the equivalent circuit (EC) approach. The extrinsic part of the device was characterized by an extended electrical EC to accurately characterize distributed parasitic effects, especially in the mm-wave frequency range. A hybrid technique of EC and NN is used to simulate the intrinsic transistor. Instead of employing a black box, a gray box of interconnected NN models of the intrinsic branches’ admittances was utilized. This approach addresses the challenges of extracting reliable values for the intrinsic capacitances, conductances, and resistances, particularly in the presence of higher measurement uncertainty. Additionally, the NN model provides accurate simulations of the bias and frequency dependence in the active region of the device. The proposed approach was applied to GaN High Electron Mobility Transistors (HEMTs) of various sizes on SiC and Diamond substrates under different bias conditions and across a wide range of frequencies. The simulation results exhibited an excellent agreement with the measurements and verified the validity of the proposed approach for small-signal modeling and linear circuit design. This work could also be extended to develop large-signal models for nonlinear circuit design applications.