Graph neural network unveils the spatiotemporal evolution of structural defects in sheared granular materials

Abstract

The disordered nature of granular materials poses great difficulty to the accurate characterization of microscopic structures. Despite numerous handcrafted structural indicators, the relationship between particle-scale structure and dynamics, as well as the structural origins of complex constitutive behaviors, remain subjects of debate. In this paper, we utilize a Graph Convolutional Neural Network (GCNN) to establish the structure-property relationship within granular materials. The GCNN model effectively identifies active particles exhibiting intense nonaffine activities based solely on initial particle positions, without relying on handcrafted features. Additionally, we derive a structural indicator called susceptibility from the GCNN output, which quantifies the fragility of local structures to external stimuli and enables the characterization of structural evolution during the shearing process. We demonstrate that structural defects with high susceptibility tend to form spatial clusters, and the distinct failure modes in dense and loose granular assemblies are driven by the differing spatiotemporal evolution of these defect clusters. Our findings suggest that the structural origin of macroscopic yielding in dense granular materials lies in the formation of system-spanning defect clusters, which facilitates the percolation of high-mobility zones and the development of shear bands. Finally, our study indicates that graph-based neural networks are well-suited for modeling and predicting the complex behaviors of granular materials, providing a powerful approach to uncovering underlying mechanisms and deepening our understanding of these materials.