Abstract
This paper thoroughly analyzes six different architectures of Artificial Neural Network (ANN) used in the development of small-signal model of Gallium Nitride (GaN) High Electron Mobility Transistors (HEMTs). At the outset, multilayer perceptron, cascade-forward, nonlinear autoregressive with exogenous inputs (NARX) in series-parallel and parallel configurations, distributed layer network, and layer recurrent architectures are used to develop GaN HEMT models for simulating the behavior of the device. Subsequently, comparison of the proposed architecture is carried out in terms of ease of implementation, simulation time, computational efficiency, fitting curves, mean squared error, mean absolute error, and coefficient of determination at distinct bias conditions. It is identified that the NARX series-parallel architecture based model is the most effective small-signal model among all the other ANN based models. It is computationally efficient, simple to implement, and possess the best generalization capability. It is also observed that the multilayer perceptron and cascade-forward exhibit analogous performance but the latter has a little edge. The NARX-parallel and feedback delay exhibit similar performance whereas the layer recurrent architecture is found to be the least suitable for the modelling of GaN HEMTs.