Machine Learning Accelerated Atomistic Simulations for 2D Materials With Defects

Abstract

Abstract 2D materials generally show very different physical and chemical properties from 3D materials, which provide them promising applications in cutting-edge technology areas like aerospace, energy storage and electronics. To better understand and illustrate their unique properties, current research heavily relies on atomistic simulations, while successful simulations require the high reliability of interatomic interaction potentials that empirical potentials cannot provide. The ab initio calculations, for example density functional theory (DFT), are able to conduct high-fidelity simulations in essence, but with a high computational cost and largely limited simulation size. Recently machine learning potentials become a trend to interpolate the potential energy surface based on artificial neural networks (ANNs) with reference datasets from first principle calculations. Although machine learning potentials have been developed for many material systems including 2D materials, there is limited work published regarding structural defects. Using graphene as a model material, we demonstrate a new machine learning potential for 2D materials with defects. The reference energies are generated by DFT calculations. ANN with two hidden layers is used for the training, which has been demonstrated to be suitable for developing machine learning potential in our previous work. For ANN input, the atomic structures are coded using perturbation-invariant representations. After training, the weight and bias parameters are exported as our potential and then imported into LAMMPS software to conduct MD simulations. It is expected that our work provides fundamental support on investigating defective graphene and understanding defect effects to develop structure-property relationship. This will also promote the development of machine learning based simulation tools for the study and design of complex materials.