Machine Learning Potentials for Graphene

Abstract

Graphene has been one of the most researched material in the world for the past two decades due to its unique combination of mechanical, thermal and electrical properties. Graphene exists in a stable two dimensional (2D) structure with hexagonal carbon rings. This special 2D structure of graphene enables it to exhibit a wide range of peculiar material properties like high Young’s modulus, high specific strength, and electrical conductivity etc. However, it is extremely challenging and costly to investigate graphene solely based on experimental tests. Atomistic simulations are powerful computational techniques for characterizing materials at small length and time scales with a fraction of cost relative to experimental testing. High fidelity atomistic simulations like Density Functional Theory (DFT) simulations, and ab initio molecular dynamic simulations have higher accuracy in predicting 2D material properties but are computationally expensive. Classic molecular dynamics (MD) simulations adopt empirical interatomic potentials which drastically reduce the computational time but has lower simulation accuracy. To bridge the gap between these two type of simulation techniques, a new artificial neural network potential is developed, for graphene in this study, to enable the characterization of 2D materials using classic MD simulations with a comparable accuracy of first principles simulation. This is expected to accelerate the discovery and design of novel graphene based functional materials. In the present study mechanical and thermal properties of graphene are investigated using the machine learning potentials by conducting MD simulations. To validate the accuracy of machine learning potentials mechanical properties such as Young’s modulus, ultimate tensile strength and thermal properties such as coefficient of thermal expansion and lattice parameter are evaluated for graphene and compared with existing literature.