Data-driven prediction of the mechanical behavior of nanocrystalline graphene using a deep convolutional neural network with PCA

Development of an artificial intelligence model that predicts the mechanical behavior of nanocrystalline graphene by extracting features from randomly arranged grain boundaries according to grain size.The application of data augmentation methods based on periodic boundary conditions.Improvement of the network prediction performance by training principal components mapped to low-dimensional domains.Visualization of image features of randomly formed and arranged nanocrystalline graphene using Grad-CAM.Comparative analysis of prediction performance of pretrained networks using transfer learning. The mechanical properties of nanocrystalline graphene significantly depend on its complex grain boundary configurations and defect distributions, with its inherent nanostructural complexity posing substantial challenges for existing computational methods. This study addresses these challenges by developing an artificial intelligence model that predicts the mechanical behavior of nanocrystalline graphene through the extraction of characteristics from randomly arranged grain boundaries based on grain size. Utilizing Voronoi tessellation, we modeled realistic grain boundaries at the atomic level, while principal component analysis (PCA) was employed to effectively reduce data dimensionality, greatly enhancing the learning efficiency of the convolutional neural network (CNN). By implementing simple yet efficient data augmentation method based on periodic boundary conditions, we substantially expanded the training dataset, providing a robust foundation for model training and validation. The model demonstrated high accuracy in predicting the mechanical responses of nanocrystalline graphene, effectively capturing the crucial impacts of defects and grain boundary distributions. The implementation of PCA proved essential in enhancing prediction accuracy for unseen data, particularly in interpolation and extrapolation scenarios, by concentrating on learning the principal components that govern mechanical behavior. Additionally, by applying explainable artificial intelligence (XAI) tools such as Grad-CAM, we validated the applicability of a pretrained network using minimal data, confirming its ability to identify crucial features impacting material properties.