Physics-informed machine learning for compact modeling and variability prediction of 2D material-based THz transistors

Abstract The growing demand for ultra-high-speed electronics and communication systems has intensified the search for advanced modeling techniques to support the next generation of beyond-CMOS devices. Two-dimensional (2D) materials such as transition metal dichalcogenides (TMDs) and graphene have demonstrated exceptional electrical properties, making them strong candidates for terahertz (THz) transistors. However, accurately predicting device behavior, variability, and reliability remains challenging due to complex physical interactions at the nanoscale and the lack of robust, generalizable compact models. In this work, we propose a novel physics-informed machine learning (PIML) framework for compact modeling and variability prediction of 2D material-based THz transistors. By integrating fundamental semiconductor physics with data-driven neural network architectures, the proposed framework enhances prediction accuracy and model interpretability while maintaining computational efficiency. Extensive simulation experiments validate the framework using open-source device datasets and custom-generated synthetic data for 2D TMD transistors operating in the THz regime. Results demonstrate significant improvements over conventional empirical models in terms of prediction error, generalization across device geometries, and resilience to process-induced variability. This work bridges the gap between physics-based modeling and modern machine learning, providing a practical toolset for high-speed circuit designers. The proposed approach supports advanced design automation flows for emerging THz integrated circuits, contributing to the development of reliable, high-performance electronics for future wireless communication and computing infrastructures.