Abstract
ABSTRACT Artificial neural networks (ANNs) have become integral in the accurate modeling of gallium nitride high electron mobility transistors (GaN HEMTs), but the convergence, robustness, and accuracy of such models are highly sensitive to the tuning of initial values of weights and biases. Optimization algorithms are often utilized for the initialization of parameters to improve the performance of GaN HEMT models. Therefore, this work evaluates and extensively compares hybrid modeling procedures for GaN HEMTs to investigate key aspects such as accuracy, efficiency, and complexity. Specifically, grey wolf optimizer (GWO), black hole optimization (BHO), reptile search algorithm (RSA), and spotted hyena optimizer (SHO) optimization algorithms are utilized with ANN to develop the hybrid approaches. The models are trained and tested across a wide range of operating conditions. A comparative analysis demonstrated that the GWO‐based hybrid optimization approach (GWO‐ANN) consistently outperformed the BHO‐ANN, RSA‐ANN, and SHO‐ANN hybrid optimization approaches in terms of accuracy, complexity, and convergence, and displayed superior alignment between measured and simulated S‐parameters over the full frequency spectrum. The BHO‐ANN based models, while slightly less accurate, manifested reduced computation time due to their simpler implementation. In contrast, the SHO‐ANN based models exhibited the least favorable performance across all metrics, including accuracy, convergence, complexity, and computational time.