Modeling Advanced Electronic Device Characterization Using a Unified Physics-Informed Machine Learning Framework
Abstract
Advancements in nanoscale semiconductor technologies, wide-bandgap materials, and emerging 2D devices have significantly increased the complexity of device characterization, making traditional TCAD simulations and compact modeling increasingly time-consuming and insufficiently flexible. Physics-Informed Machine Learning (PIML) offers a promising pathway by integrating physical laws with data-driven modeling to enhance accuracy, interpretability, and generalization. This paper presents a unified PIML framework for advanced electronic device characterization that embeds semiconductor transport physics, electrostatic constraints, and compact-model priors directly into neural architectures and loss functions. The proposed approach leverages hybrid PINN-based solvers, residual-learning compact models, and uncertainty-aware training to model I–V, C–V, and RF characteristics across nanosheet FETs, GaN HEMTs, and 2D-material THz transistors. Experimental results demonstrate that PIML models reduce prediction error by up to 35% compared to purely data-driven models and achieve substantially improved physical consistency, particularly in charge conservation and monotonic device behavior. Moreover, the differentiable structure of the framework enables efficient inverse design and parameter extraction, significantly accelerating device optimization workflows. Overall, the study establishes physics-informed ML as a scalable and robust methodology for next-generation electronic device modeling, bridging the gap between high-fidelity physics solvers and fast, design-oriented surrogate models.