GraPhAI: Neural Networks for Solving Centrosymmetric Crystal Structures.

Area of Science:

• Crystallography
• Artificial Intelligence
• Materials Science

Background:

• Solving the crystallographic phase problem from low-resolution diffraction data remains a significant challenge.
• Existing methods lack a general-purpose solution for accurate crystal structure determination.

Purpose of the Study:

• To develop an efficient deep learning method for ab initio phasing using low-resolution diffraction data.
• To introduce a novel graph-based representation for diffraction data suitable for neural networks.

Main Methods:

• Development of a graph neural network named GraPhAI.
• Implementation of a new diffraction data representation in graph form for deep learning.
• Training and evaluation of GraPhAI models on centrosymmetric crystal structures with typical unit-cell volumes.

Main Results:

• GraPhAI achieves over 80% success rate for crystal structure determination down to 2 Å resolution.
• The method is particularly effective for structures containing atoms with atomic number Z ≥ 19, including metal-organic frameworks, coordination compounds, and inorganic structures.
• Current success for purely organic crystal structures is limited.

Conclusions:

• GraPhAI offers a promising and efficient approach for ab initio phasing, particularly for inorganic and metal-containing crystalline materials.
• The graph-based data representation is key to the model's performance.
• Further development is needed to extend the method's applicability to purely organic crystal structures.