Physics-Guided Active Learning and Geometry-Aware Electrothermal Screening of Semiconductor Devices: A GaN Case Study

Abstract

Exploring high-dimensional semiconductor device design spaces under coupled electrothermal constraints remains challenging, particularly in 3-D heterogeneous systems with tightly coupled electrical, thermal, and integration requirements. Conventional TCAD-driven exploration is physically accurate but computationally expensive and offers limited insight into governing parameters. This work presents physics-guided active learning trigate optimization (PALTO), a physics-guided active-learning framework for data-efficient design space exploration, combining ensemble surrogate modeling, uncertainty-driven sampling, and diversity-aware selection to reduce reliance on exhaustive TCAD simulations while preserving predictive fidelity. Using multichannel trigate GaN devices as a case study, PALTO identifies Pareto-optimal configurations. A geometry-aware electrothermal scaling methodology is further introduced to evaluate multifin scalability using compact metrics capturing electrostatic activation, thermal vulnerability, and Joule-stress degradation. This analysis—corroborated by feature-importance ranking—identifies GaN channel thickness as a dominant parameter governing electrothermal scalability. Geometry-derived indicators are further used to organize Pareto-selected configurations according to relative electrothermal scaling tendencies without additional high-fin-count TCAD simulations. Results show that stronger barrier activation improves single-fin current but causes earlier electrothermal saturation, whereas GaN-dominant structures exhibit improved high-fin scalability.