Machine Learning Stress Overshoot of Amorphous Solids

When undergoing quasistatic shear, slow-quenched glasses exhibit an overshoot in shear stress, which decreases with the increase of quench rate and eventually evolves into a smooth crossover. However, the structure of glasses does not exhibit significant quench rate dependence. It thus remains unclear what determines the emergence of stress overshoot. Here, inspired by image recognition, we propose that stress overshoot of amorphous solids can be predicted from structures of solids using machine learning methods. Our results show that pure geometrical quantities such as local coordination number, bond orientational order and voronoi cell volume, which seem to weakly correlate with soft spots, have displayed very high predictive power of stabilities of amorphous solids, when combined with coarse-grain and machine learning methods. Besides, the stress overshoot in pinned systems suggests that our recently defined order parameter from the normal modes of vibration is a more general structure descriptor to identify stress overshoot. The high accuracy in identifying stress overshoot may imply that machine learning methods successfully capture the spatial correlation of the descriptor in some complicated way.