Title: Graph representation learning for motif analysis of dense sand

Author (Talk): WaiChing Sun, Columbia University

Abstract:

Understanding the evolving micromechanical behavior of granular materials during deformation requires robust quantification of internal structural changes. In this study, we introduce a novel machine learning framework utilizing graph representation learning to analyze the contact networks of dense sand subjected to triaxial compression, captured via in-situ synchrotron micro-computed tomography (SMT). To overcome the computational bottlenecks of traditional graph-theoretic approacheswhich are typically restricted to analyzing small structural subgraphswe implement a novel order embedding approach coupled with Monte Carlo tree search (MCTS). This framework enables, for the first time, the efficient identification and counting of complex, large-scale structural motifs (up to 15 nodes) within experimental micro-CT data. By mapping these subgraphs into a low-dimensional space, we can quantitatively track the evolution of critical topological features, such as force chains and cycles, and correlate them with grain morphology and shear band formation across different strain regimes. Ultimately, this work establishes order embedding as a scalable, powerful tool for decoding the complex mesoscale rearrangements that dictate the macroscopic strength, stability, and failure mechanisms of granular media.