Structural defects in granular materials have been difficult to identify and connect to possible deformation and failure patterns. Machine learning has recently been used successfully to identify structural defects in glasses, polycrystals, and 2D granular materials. This success is an important step toward building a microscopic theory of plasticity for granular and amorphous materials. Here, we discuss recent work combining in-situ X-ray measurements of particle-resolved structure and stress with machine learning to locate, quantify, and predict local rearrangements in 3D granular materials. The 3D granular materials we studied include packings of ruby particles and packings of sand grains subjected to uniaxial, hydrostatic, and triaxial loading conditions. The local rearrangements that we studied include non-affine translation, rotation, and volumetric and shear strain. We employed machine learning techniques including Linear Discriminant Analysis and Support Vector Machine, and multivariate regression to identify the location and quantify the magnitude of rearrangements. We found that (1) local rearrangements can be classified as above or below a given magnitude with up to 80% accuracy by considering only a collection of local structural measures, (2) particle-resolved stress did not improve predictions of local rearrangements, and (3) machines trained on samples subjected to one loading path provided robust predictions of rearrangements in samples subjected to another loading path.
We acknowledge support from subcontracts from Lawrence Livermore National Laboratory LDRD program, Award No. LW-17-009, the U.S. National Science Foundation CAREER Award No. CBET-1942096, and synchrotron beamtime at the ESRF under grants ma-3373 and ma-4200. Part of this work was performed under the auspices of the Department of Energy by Lawrence Livermore National Laboratory (LLNL) under award No. DE-AC52-07NA27344.