Machine learning accelerated search of the strongest graphene/h-BN interface with designed fracture properties

Abstract

Two-dimensional lateral heterostructures exhibit novel electronic and optical properties that are induced by their in-plane interface for which the mechanical properties of the interface are important for the stability of the lateral heterostructure. Therefore, we performed molecular dynamics simulations and developed a convolutional neural network-based machine learning model to study the fracture properties of the interface in a graphene/hexagonal boron nitride lateral heterostructure. The molecular dynamics (MD) simulations show that the shape of the interface can cause an 80% difference in the fracture stress and the fracture strain for the interface. By using 11 500 training samples obtained with help of high-cost MD simulation, the machine learning model is able to search out the strongest interfaces with the largest fracture strain and fracture stress in a large sample space with over 150 000 structures. By analyzing the atomic configuration of these strongest interfaces, we disclose two major factors dominating the interface strength, including the interface roughness and the strength of the chemical bond across the interface. We also explore the correlation between the fracture properties and the thermal conductivity for these lateral heterostructures by examining the bond type and the shape of the graphene/hexagonal boron nitride interface. We find that interfaces comprised of stronger bonds and smoother zigzag interfaces can relieve the abrupt change of the acoustic velocity, leading to the enhancement of the interface thermal conductivity. These findings will be valuable for the application of the two-dimensional lateral heterostructure in electronic devices.