TCAD-calibrated ML for trap-induced degradation in AlGaN/GaN HEMTs

This paper presents a TCAD-calibrated machine learning (ML) framework for predicting trap-induced electrical degradation in AlGaN/GaN high electron mobility transistors (HEMTs). The TCAD model is first calibrated and validated against experimentally reported output (Ids–Vds), transfer (Ids–Vgs), and transconductance characteristics from devices with comparable geometry and operating conditions, ensuring physical fidelity. Buffer trap effects are systematically analyzed by varying trap densities from 5.5 × 1015 to 5.5 × 1018 cm−3 at a fixed trap energy of 0.75 eV. The validated TCAD dataset is then used to train Random Forest, Gradient Boosting, Convolutional Neural Network, and Long Short-Term Memory models to reproduce trap-dependent device behavior.

The proposed framework shows strong agreement with experimentally calibrated TCAD simulations, while significantly reducing computational cost. Among the evaluated models, Gradient Boosting achieves the highest accuracy and stability (R2 ≈ 0.9994, MSE ≈ 2.1 × 10−7), effectively capturing current collapse and transconductance degradation. The results establish TCAD-assisted ML as a fast and reliable surrogate for TCAD-based reliability analysis of GaN HEMTs.