Scalable deeper graph neural networks for high-performance materials property prediction

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning, particularly graph neural networks (GNNs), shows promise for materials discovery by predicting properties from crystal structures.
• Existing GNN models face challenges with scalability, complex hyperparameter tuning, and performance limitations due to over-smoothing.

Purpose of the Study:

• To develop a scalable and high-performance GNN model for accurate materials property prediction.
• To address the limitations of existing GNNs in terms of scalability and performance.

Main Methods:

• Proposed DeeperGATGNN, a scalable global graph attention neural network incorporating differentiable group normalization (DGN) and skip connections.
• Conducted systematic benchmark studies on multiple datasets to evaluate model performance and scalability.

Main Results:

• DeeperGATGNN achieved state-of-the-art prediction results on five out of six datasets, outperforming existing GNN models by up to 10%.
• Demonstrated superior scalability, enabling the training of very deep networks (>30 layers) without significant performance degradation.

Conclusions:

• DeeperGATGNN offers a scalable and effective solution for high-performance materials property prediction.
• The model advancements facilitate deeper network architectures, pushing the boundaries of materials discovery.