Graph Representation Framework for Accelerating Atomic-Level Semiconductor Device Simulation

Abstract

This article proposes a machine-learning (ML) method to accelerate atomic-level device simulation. The main idea is to utilize graph convolutional network (GCN) to predict the potential distribution of the device, aiming to expedite the time-consuming self-consistent calculation process of the transport-Poisson equation within the nonequilibrium Green’s function (NEGF) method. Using the network-predicted electrostatic potential distribution as the initial solution can accelerate the convergence speed of the above process without compromising the accuracy of NEGF calculations. Most importantly, this network introduces a new method for device representation using graphs. It explicitly encodes the coupling effect between the device’s scattering region (SR) and electrodes while effectively capturing quantum mechanical effects at the atomic level. This method offers a more reliable and physically significant approach to device encoding, providing fresh insights on merging ML with atomic-level device simulations.