Performance Enhancement of Recessed T-Gated Fe-Doped GaN High Electron Mobility Transistor Using an Attributed Multi-Order Graph Convolutional Network Integrated with Frilled Lizard Optimization

Abstract: Enhancing material and structural characteristics in High Electron Mobility Transistors (HEMTs) made with Gallium Nitride (GaN) constitutes an advanced design approach where the layer compositions, lateral dimensions, and gate metal types are optimized simultaneously to enhance performance in both RF and DC applications. The degree of nonlinearity that exists between material structure, physical structure, and electrical parameters produces interactions that challenge accurate performance predictions and optimizations. In the absence of a standard analytical approach, interdependencies between manufacturing structure and performance can again result in suboptimal convergence points, reduced operational stability in use, and lower performance under high-field conditions. To overcome these challenges, this study introduces an intelligent prediction–optimization framework for recessed T-gated Fe-doped AlN/GaN/SiC HEMTs. The proposed hybrid model integrates an Attributed Multi-Order Graph Convolutional Network (AMOGCN) with Frilled Lizard Optimization (FLO), collectively referred to as the AMOGCN-FLO framework. The AMOGCN effectively predicts key device parameters by learning nonlinear correlations among material and structural attributes through multi-order graph feature propagation. Subsequently, the FLO algorithm optimizes the AMOGCN weight parameters via its adaptive exploration-exploitation mechanism inspired by the frilled lizard’s hunting and tree-climbing behavior, ensuring convergence to globally optimal values. The proposed AMOGCN-FLO framework achieves notable performance improvements compared to existing techniques such as Artificial Neural Networks (ANNs), ANN-Particle Swarm Optimization (PSO), Physics-Constrained Neural Network (PCNN), Genetic Algorithm-Extreme Learning Machine (GA–ELM), and Knowledge-Based Neural Network (KBNN). The optimized model achieves a maximum drain current of 1.89 A/mm, transconductance of 435.9 mS/mm and cutoff frequency of 197.4 GHz, exceeding all the comparative models. Such a combined approach offers a strong and smart way to predict and optimize the behavior of GaN-based HEMT devices successfully, leading to the creation of efficient design in future RF and microwave power electronics.