PerCNet: Periodic complete representation for crystal graphs.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Crystals are often represented as graphs, but existing methods primarily capture local atomic information (bond distances, angles).
• This focus on local features leads to non-unique representations, where different crystals map to the same graph structure.
• Capturing both local and global crystal information is crucial for accurate material property prediction.

Purpose of the Study:

• To develop a unique, one-to-one graph representation for crystal materials by incorporating global features.
• To propose a novel computational framework for predicting material properties based on this enhanced representation.
• To address the limitations of existing crystal graph representations that suffer from information loss.

Main Methods:

• Proposed a periodic complete graph representation that includes dihedral angles to capture global structural information.
• Developed a calculation algorithm for infinite extended crystal materials based on the new representation.
• Introduced PerCNet, a neural network with a specialized message-passing mechanism for material property prediction.

Main Results:

• The proposed representation theoretically ensures a one-to-one mapping between crystal structures and their graph models.
• PerCNet achieved state-of-the-art performance on two large-scale material benchmark datasets, outperforming baseline methods.
• The Mean Absolute Error (MAE) was significantly reduced by the proposed method.

Conclusions:

• The novel graph representation effectively captures both local and global crystal information, resolving the many-to-one mapping issue.
• PerCNet demonstrates superior accuracy in predicting crystal material properties compared to existing approaches.
• This work provides a robust and unique graph-based modeling approach for materials science research.