Quantifying disorder one atom at a time using an interpretable graph neural network paradigm.

Area of Science:

• Materials Science
• Computational Materials Science
• Data Science

Background:

• Quantifying atomic disorder is crucial for understanding material properties and durability.
• Existing methods may not fully capture the complex local structural environments.

Purpose of the Study:

• To develop a physically interpretable metric for local atomic disorder.
• To establish a continuous spectrum for disorder quantification.
• To apply this metric to various material systems and phenomena.

Main Methods:

• Utilized graph neural networks to define the Structure-Oriented Disorder Analysis Spectrum (SODAS) metric.
• Quantified local atomic configuration diversity against thermal perturbations.
• Applied the SODAS metric to grain boundaries, solid-liquid interfaces, microstructures, and fracture.

Main Results:

• SODAS provides a continuous spectrum of disorder, bridging solid and liquid phases.
• Demonstrated spatio-temporal tracking of interfaces in elemental aluminum.
• Extracted physics-preserved gradients from continuous disorder fields for failure prediction.

Conclusions:

• The SODAS framework offers a generalizable pathway to quantify atomic disorder.
• This metric links complex local atomic structures to macroscopic material behavior.
• Enables better prediction of material performance and failure mechanisms.