Disentangling symmetry and many-body effects in graph neural networks for plastic response prediction in glasses

Abstract Identifying where and when shear transformation will activate in metallic glasses is notoriously difficult, as long-range order is absent and local motifs vary widely. Graph neural networks (GNNs), operating end-to-end on atomic graphs, provide a robust framework for learning structure–property relations in disordered materials. Here, we adapt and systematically benchmark a suite of state-of-the-art GNNs in a prototypical Cu–Zr metallic glass, comparing representative rotationally invariant and equivariant GNN architectures while increasing the many-body interaction order. Across architectures, moving from 2-body to 3-body geometric information consistently improves prediction of the atom-resolved non-affine response, whereas incorporating 4-body terms yields marginal or no gains. For this prediction task, the best performance is obtained with a 3-body invariant GNN model (GemNet-T), although the performance gap between top-performing models is modest. Analysis of the error distribution shows that the model improves in resolving rare structural outliers at both ends of the stability spectrum—the “weakest links” and “stable backbones”. Furthermore, by treating the trained GNN as a differentiable physical probe, we develop a physics-informed interpretability framework. The gradient magnitude of the model prediction with respect to atomic positions serves as a descriptor of local “plastic sensitivity”. This uncovers a hierarchy of plastic deformation, with highly sensitive “triggers” initiating plastic events and less sensitive “followers” subsequently activated. These results provide insights for both accurate prediction and mechanistic understanding of the hierarchical plastic response in glasses.