Strength prediction and design of defective graphene based on machine learning approach

Abstract

Defects are inevitable in two-dimensional (2D) materials. Thus, the strength prediction and design are crucial for practical application of defective 2D materials. Utilizing a dataset from molecular dynamic (MD) simulations, this study aims to predict, as well as design the strength of defective graphene. Through convolutional neural networks (CNN), the constructed residual network ResNet34 can accurately predict the fracture strength directly from the defect configuration of graphene. Meanwhile, ablation class activation map (Ablation-CAM) further identifies the sensitive domains that dominate the fracture strength, in accordance with the crack initiation regions confirmed by MD simulations and experiments. In particular, a new descriptor named sensitive domain factor (SDF) was developed to characterize the important features of sensitive domains. Furthermore, a genetic algorithm (GA) is then applied to strategically optimize the defect configuration under a given defect density, achieving an ideal configuration with the maximum fracture strength. This work pioneers a machine learning framework for the extraction and optimization of defective features in monolayer graphene, offering a novel approach to design the mechanical properties through defect engineering.