Predicting the failure of two-dimensional silica glasses

Area of Science:

• Materials Science
• Computational Materials Science
• Artificial Intelligence

Background:

• Predicting material failure is crucial for industrial applications and component monitoring.
• Deep learning advances enable failure prediction in disordered solids, but lack physical interpretability.
• Interpreting complex machine learning models remains a significant challenge in materials science.

Purpose of the Study:

• To develop an interpretable machine learning approach for predicting material failure in silica glasses.
• To bridge the gap between accurate deep learning predictions and physical understanding.
• To enable the transferability of predictive models to different sample geometries and experimental data.

Main Methods:

• Utilized machine learning models to predict failure in simulated 2D silica glasses based on initial structure.
• Employed Gradient-weighted Class Activation Mapping (Grad-CAM) to generate attention maps for model predictions.
• Validated the physical interpretability of attention maps in relation to topological defects and local potential energies.

Main Results:

• Attention maps derived from Grad-CAM provided physically meaningful interpretations of failure predictions.
• The predictive models demonstrated transferability to silica glass samples of varying shapes and sizes.
• The approach successfully applied interpretable predictions to experimental images of silica glasses.

Conclusions:

• Artificial neural networks can provide interpretable failure predictions for materials.
• Grad-CAM facilitates the physical interpretation of machine learning models in materials science.
• This strategy enables the application of simulation-trained models to experimental data for reliable material behavior prediction.