GCPNet: An interpretable Generic Crystal Pattern graph neural Network for predicting material properties.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting material properties from crystal structures is crucial for materials discovery.
• Existing graph neural networks often lack microstructure information and have limitations in feature extraction.

Purpose of the Study:

• Introduce a novel graph neural network, Generic Crystal Pattern graph neural Network (GCPNet), for accurate material property prediction.
• Enhance the extraction of structural features from crystalline materials.

Main Methods:

• Developed GCPNet based on crystal pattern graphs.
• Employed Graph Convolutional Attention Operator (GCAO) and a two-level update mechanism.
• Validated the model on five public datasets.

Main Results:

• GCPNet achieved higher precision in material property prediction compared to existing networks.
• Demonstrated robustness and ease of use in real-world applications.
• Showcased model interpretability, improving perovskite screening efficiency by 32%.

Conclusions:

• GCPNet provides an effective solution for screening and discovering ideal crystals.
• Offers an efficient alternative to current neural networks for material property prediction.
• Facilitates accelerated materials design and development.