Predicting orientation-dependent plastic susceptibility from static structure in amorphous solids via deep learning

TL;DR

This paper introduces a rotation-variant structure representation combined with deep learning to accurately predict orientation-dependent plastic susceptibility in amorphous solids from static atomic structures, advancing structure-property understanding.

Contribution

It presents a novel rotation-variant structure representation and CNN-based model that significantly improves prediction accuracy of shear transformation propensity in glasses.

Findings

High prediction accuracy for shear transformation propensity.

Model transferability across different compositions and processing histories.

Insight into atomic packing features influencing mechanical response.

Abstract

It has been a long-standing materials science challenge to establish structure-property relations in amorphous solids. Here we introduce a rotation-variant local structure representation that enables different predictions for different loading orientations, which is found essential for high-fidelity prediction of the propensity for stress-driven shear transformations. This novel structure representation, when combined with convolutional neural network (CNN), a powerful deep learning algorithm, leads to unprecedented accuracy for identifying atoms with high propensity for shear transformations (i.e., plastic susceptibility), solely from the static structure - the spatial atomic positions - in both two- and three-dimensional model glasses. The data-driven models trained on samples at one composition and a given processing history are found transferrable to glass samples with different processing histories or at different compositions in the same alloy system. Our analysis of the new structure representation also provides valuable insight into key atomic packing features that influence the local mechanical response and its anisotropy in glasses.