Machine learning mechanical properties of defect-engineered hexagonal boron nitride

Defect engineering of 2D materials has exhibited promising potentials for designing new properties beyond their pristine forms. However, the structure–property mapping for defect-engineered 2D materials is usually highly complex due to the vast designing space spanned by multiple types of defects, which may coexist and jointly affect the properties of 2D materials. In this work, we propose a general machine learning framework to efficiently predict the mechanical properties of defect-engineered 2D materials with a generic configuration of coexisting types of defects. The mechanical properties of defect-engineered hexagonal boron nitride are taken as examples for demonstration. Through thousands of molecular dynamics simulations, we first reveal the complex mechanical response of hexagonal boron nitride containing both vacancy defects and substitutional doping defects. Then a machine learning model based on convolutional neural network is established and trained to efficiently predict the simulated mechanical properties. We also demonstrate how the neural network can be applied for finding desirable defected configurations with targeted mechanical properties. The proposed methods are expected be applicable to a variety of properties of defect-engineered 2D materials.