EOSnet: Embedded Overlap Structures for Graph Neural Networks in Predicting Material Properties.

Area of Science:

• Computational Materials Science
• Machine Learning for Materials

Background:

• Graph Neural Networks (GNNs) are increasingly used for predicting material properties.
• Current GNNs often struggle with capturing many-body interactions and require manual feature engineering.

Purpose of the Study:

• To introduce EOSnet, a novel GNN architecture.
• To address limitations in capturing many-body interactions and reduce manual feature engineering in GNNs for materials science.

Main Methods:

• Developed EOSnet, integrating Gaussian Overlap Matrix (GOM) fingerprints as node features within the GNN.
• GOM fingerprints encode many-body interactions through orbital overlap matrices, offering a rotationally invariant representation.

Main Results:

• EOSnet achieved superior performance in predicting various material properties, especially those sensitive to many-body interactions.
• Achieved a 0.163 eV mean absolute error for band gap prediction, outperforming state-of-the-art models.
• Demonstrated 97.7% accuracy in metal/nonmetal classification and excelled in predicting mechanical properties.

Conclusions:

• Embedding GOM fingerprints into GNN node features significantly enhances the model's ability to capture complex atomic interactions.
• EOSnet represents a powerful advancement for materials discovery and property prediction using machine learning.