Graph convolutional neural networks with global attention for improved materials property prediction

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Materials property prediction (MPP) is crucial for discovering new materials.
• Graph neural networks (GNNs) show promise for MPP but struggle to differentiate atomic contributions.
• Existing GNN models lack the ability to effectively weigh individual atom influences on material properties.

Purpose of the Study:

• To develop a novel graph neural network model, GATGNN, for enhanced materials property prediction.
• To improve the accuracy of inorganic material property predictions by effectively differentiating atomic contributions.
• To provide insights into the correlation between atomic features and material properties.

Main Methods:

• Developed GATGNN, a novel graph neural network incorporating augmented graph-attention (AGAT) layers and a global attention layer.
• AGAT layers capture local atomic relationships, while the global attention layer assesses overall atomic contributions.
• The model was trained and validated on various inorganic material property prediction tasks.

Main Results:

• GATGNN achieved superior prediction performance compared to state-of-the-art GNN models across multiple material properties.
• The model demonstrated an improved ability to differentiate the contributions of individual atoms to material properties.
• Experimental results confirmed the effectiveness of the AGAT and global attention mechanisms.

Conclusions:

• GATGNN offers a significant advancement in machine learning for materials property prediction.
• The model's ability to interpret atomic contributions provides valuable insights for materials design.
• GATGNN represents a powerful new tool for accelerating materials discovery and development.