Machine-Learned Quantification of Local Structure and Its Effects in Fracture of Silica Glass and Deformation of Metallic Glass

The variety of local atomic environments found in a glass far exceeds that found in a crystalline material. This makes the task of linking physically important phenomena (e.g. crack nucleation or shear transformation) to local structural risk factors particularly challenging in glassy materials. We present case studies in two simulated materials (silica glass and a binary metallic glass) using unsupervised machine learning techniques (the Gaussian Integral Inner Product Distance with agglomerative clustering and diffusion maps) to extract local structural features, and supervised machine learning to link those features to mechanical behavior. We pinpoint preexisting defects in the as-quenched state as risk factors leading to fracture nucleation in silica, and show how detailed quantification of structural evolution in a metallic glass shear band points the way to improved stateful plastic models. We also consider the comparative merits of human-designed versus machine-learned structural descriptors for glass and look forward to ways to use them synergistically. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525 (SAND2022-8095 A).