Abstract:
This thesis investigates and compares the ensemble modeling methods and neural networks approaches for small-signal modeling of Gallium Nitride High Electron Mobility Transistors (GaN HEMTs). Specifically, ensemble methods are represented by Random Forests and eXtreme Gradient Boosting (XGBoost) algorithms, while neural networks techniques consist of Adaptive Neuro-Fuzzy Inference Systems (ANFIS) and Feed-forward Artificial Neural Networks (FFANN). To carry out the research to a higher standard, this work utilizes two distinct GaN HEMT devices. The first, a GaN HEMT grown on Diamond, is characterized by a smaller dataset and fewer modeling parameters. Conversely, the second device, a GaN HEMT on Silicon, possesses a larger dataset and a greater number of training parameters.
Furthermore, the model performance is meticulously evaluated using Mean Square Error
(MSE), Mean Absolute Error (MAE), and the coefficient of determination (R2). Findings suggest that ensemble models exhibit enhanced stability and greater robustness against overfitting. While the neural networks-based models demonstrate superior accuracy and a more streamlined development process. This research provides critical guidance for researchers and engineers in selecting the most suitable modeling approach for certain GaN HEMT devices. The choice hinges on a careful balance between prioritizing accuracy, mitigating overfitting, and managing the complexities inherent in model development.