Deep learning for automated classification and characterization of amorphous materials.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Quantifying structure-property relationships in complex materials, especially amorphous ones, is challenging due to the lack of long-range order.
• Standard techniques struggle to define reliable structural metrics for amorphous materials, hindering characterization.
• Deep learning offers a promising avenue for analyzing complex material structures.

Purpose of the Study:

• To apply deep learning algorithms for accurate classification and characterization of amorphous materials.
• To compare the performance of convolutional neural networks (CNNs) and message passing neural networks (MPNNs) in distinguishing liquids and glasses.
• To develop novel structural metrics for glass formation based on interpretable deep learning models.

Main Methods:

• Utilized molecular dynamics simulations to generate data for two-dimensional liquids and liquid-cooled glasses.
• Applied convolutional neural networks (CNNs) and message passing neural networks (MPNNs) for material classification.
• Employed a self-attention mechanism within MPNNs to interpret their evaluation of material structures.

Main Results:

• Achieved high classification accuracy (AUC > 0.98) for distinguishing liquids and glasses using both CNNs and MPNNs.
• Demonstrated that MPNNs outperform CNNs in both classification accuracy and interpretability.
• Derived three novel structural metrics from MPNN interpretation that effectively characterize glass formation.

Conclusions:

• Deep learning, particularly MPNNs, provides a powerful and interpretable approach for analyzing amorphous materials.
• The developed methods can identify crucial structural features missed by traditional techniques.
• This work offers new insights into the structural underpinnings of glass formation and material characterization.