Explainable GNN-derived structure-property relationships in interstitial-alloy materials.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Understanding structure-property relationships is crucial for designing novel materials.
• Traditional methods for predicting material properties can be data-intensive and lack interpretability.
• Non-stoichiometric materials and interstitial alloys present unique challenges due to their variable compositions.

Purpose of the Study:

• To develop and apply graph neural networks (GNNs) for predicting properties of non-stoichiometric materials.
• To introduce an interpretable GNN framework for analyzing structure-property relationships.
• To demonstrate the superiority of GNNs over traditional interatomic potential models (IAPs) in accuracy and data efficiency.

Main Methods:

• Application of the crystal graph convolutional network (CGCNet) to predict properties of Mo2C and Ti2C.
• Development of the crystal graph explainer (CGExplainer) for model interpretability.
• Comparison of GNN performance against traditional human-derived interatomic potential models (IAPs).

Main Results:

• CGCNet demonstrated higher prediction accuracy and data efficiency compared to IAPs.
• Significant improvements were observed in the ability of GNNs to extrapolate properties to larger supercells.
• CGExplainer successfully identified key atomic arrangements governing material properties.

Conclusions:

• GNN-based approaches offer a powerful and efficient framework for materials discovery.
• The developed methodology accelerates the design of materials with tailored properties, especially for alloys with variable compositions.
• This work extends the applicability of GNNs to a wider range of complex material systems.