MICROWAVE TRANSISTOR MODELING USING MACHINE LEARNING TECHNIQUES

Abstract

The current research assesses machine learning methods for modeling GaN HEMT, specifically interpolation and extrapolation issues within higher-frequency circuits. S-parameters prediction based on GaN HEMT data is evaluated by comparing Neural Networks (ANN) and Tree-Based Models (Extra Trees - ET, Random Forests - RF). The methodology involved splitting data by VDS bias and comparing validation strategies (80/20 split vs 5-fold CV) and various hyperparameter optimizers. Results show the 80/20 split is significantly faster (∼9×) with comparable or better performance than CV=5. ET models generally outperformed RF. ET GA outperformed in interpolation, whereas ANN Optuna had superior stability and robustness with extrapolation. Models based on trees are even simpler and much quicker to apply and adjust (ET GA tuning ∼14 times quicker compared to ANN Optuna). The research confirms the trade-offs between types of models and validation strategies for accurate and efficient microwave transistor modeling.