Integrating Atomistic Insights With Circuit Simulations via Transformer-Driven Symbolic Regression

This paper introduces a framework that establishes a cohesive link between first principles-based simulations and circuit-level analyses using a machine learning-based compact modeling platform. Starting with atomistic simulations, the framework examines the microscopic details of material behavior, forming the foundation for later stages. The generated datasets, with molecular insights, are processed using machine learning algorithms to identify complex patterns and relationships. As these machine-learning models develop, they become tools for predicting behaviors beyond the reach of conventional modeling and simulation methods. Applied to circuit simulation, the framework improves understanding of electrical interactions, enhancing accuracy and speeding up design automation. As a proof of concept, we perform first principles-based simulations of the Graphene Nanoribbon Field Effect Transistor (GNRFET), an exploratory device, and create a symbolic-regression-based machine learning model that can readily be integrated into advanced circuit simulation. This framework presents a template offering a unified approach that synergizes the strengths of first principles-based simulations and circuit-level design tools.