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

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Materials Science

Background:

• Establishing structure-property relationships in amorphous solids remains a significant challenge.
• Understanding the propensity for stress-driven shear transformations is crucial for predicting material failure.

Purpose of the Study:

• To introduce a novel local structure representation for amorphous solids.
• To achieve high-fidelity prediction of shear transformation propensity using this representation.
• To leverage deep learning for accurate identification of atoms susceptible to plastic deformation.

Main Methods:

• Development of a rotationally non-invariant local structure representation.
• Integration of the new structure representation with convolutional neural networks (CNNs).
• Application of data-driven models to both 2D and 3D model glasses.

Main Results:

• Unprecedented accuracy in identifying atoms with high shear transformation propensity from static structure.
• Demonstration of model transferability across different processing histories and compositions within the same alloy system.
• Insight into atomic packing features governing local mechanical response and anisotropy.

Conclusions:

• The novel structure representation is essential for accurate prediction of mechanical behavior in amorphous solids.
• Convolutional neural networks combined with this representation offer a powerful tool for materials characterization.
• The findings provide fundamental insights into the atomic-scale origins of mechanical properties in glasses.