A Bottom-Up Machine-Learning Approach for Efficient Device Simulation

Abstract

The combination of machine-learning (ML) and electronic structure computation has proven effective in studying various properties of molecules and crystals at the atomistic level. However, challenges arise when these molecules or crystals are contacted with external electrodes, complicating the description of quantum transport properties using existing methods. In this study, we propose an attention-based heterogeneous graph neural network to characterize the global field and dynamic features of open systems. Our approach aims to accelerate or bypass the resource-intensive self-consistent iterations of solving Schrödinger and Poisson equations within nonequilibrium Green’s function (NEGF) formalism from the bottom-up, significantly improving the efficiency of quantum transport calculations. Representing the device with a heterogeneous graph largely retains its intrinsic physical characteristics, while the global graph attention network (GAT) effectively captures the propagation of nonlocal physical information, addressing prediction accuracy challenges due to device scaling. The global field heterogeneous graph neural network (GFGNN) demonstrates high accuracy, significant acceleration, and potential transferability at different channel lengths in simulations of p-n junctions (two-terminal with significant tunneling effect) and MOSFETs (three-terminal).