Surrogate-Assisted Multi-Objective Optimization of GaN HEMTs Using a TCAD DoE Dataset

Abstract

GaN HEMTs enable higher switching frequency and power density in modern power converters, but practical device optimization must balance breakdown robustness and conduction loss while maintaining stable threshold behavior compatible with gate-drive constraints. Although TCAD provides physics-based insight, iterative exploration across multiple design knobs is computationally expensive and limits design turn-around time. This paper presents an end-to-end workflow that couples a space-filling TCAD design of experiments dataset with hybrid surrogate models and surrogate-assisted NSGA-II optimization to rapidly map the breakdown voltage and on-resistance trade-off under realistic electrical constraints. A diverse design of experiments is generated via Latin hypercube sampling and greedy maximin selection, and TCAD-extracted metrics are used to train surrogates using neural networks for continuous targets and a tree ensemble to capture the non-linear breakdown behavior. The optimized Pareto front identifies a breakdown optimum, an on-resistance optimum, and a balanced knee point, which are re-validated in TCAD. Re-validation of these representative recommendations confirms close agreement across key metrics with discrepancies within approximately two percent, demonstrating a reliable and efficient framework for rapid wide bandgap device screening.