Abstract
<jats:p>Discrete element method (DEM) simulations provide time-resolved particle-scale information on particle motion, contact forces, rotation, and local deformation. However, these outputs are high-dimensional, making it difficult to directly interpret which particle-level descriptors characterize localized deformation. This study develops a machine-learning workflow for classifying shear-band particles in a reverse-fault-type DEM model. Shear-band labels are defined using a neighborhood-based local equivalent shear strain, whereas the input features are restricted to low-leakage particle-scale descriptors by excluding the target strain, displacement-derived quantities, and local deformation gradients. A gradient-boosting decision-tree classifier is trained and evaluated using particle-time records as an imbalanced binary classification problem. On the test data, the classifier achieves a PR-AUC of 0.906 and a ROC-AUC of 0.995. Spatial prediction maps show that the classifier captures the localized shear-band structure in the sampled test subset. Permutation importance and ablation analyses indicate that the classification is supported not only by particle position and angular velocity, but also by contact-mechanical, force-related, frictional, and velocity descriptors. These results show that interpretable machine-learning analysis can provide a quantitative summary of particle-scale information associated with shear-band classification in DEM datasets. In an additional threshold-sensitivity analysis, the shear-band labels were redefined using the 90th, 95th, 97.5th, and 99th percentiles while fixing the positive rate at 5% in both the training and test subsets. Angular velocity magnitude, x position, and z position consistently appeared among the top-ranked features, indicating that the main feature-importance trend is not strongly dependent on the specific 97.5th-percentile label threshold.</jats:p>