On artificial crystal structure generation for solving the phase problem with deep learning

Area of Science:

• Crystallography
• Materials Science
• Artificial Intelligence

Background:

• Solving the crystallographic phase problem is crucial for determining crystal structures.
• Neural networks offer a promising approach for phase retrieval.
• Generating realistic artificial crystal structures is essential for training these networks.

Purpose of the Study:

• To present and discuss methods for generating artificial crystal structures for neural network training.
• To evaluate the performance of the PhAI neural network on experimental data after retraining with novel artificial datasets.

Main Methods:

• Structure generation involves sampling unit-cell parameters and atom placement.
• Lattice basis vectors are generated from sampled unit-cell volumes.
• Molecule-like fragments are generated using database data to guide atom placement, complementing random methods.

Main Results:

• The PhAI neural network was benchmarked and retrained using various artificial datasets.
• Retraining PhAI with a new type of artificial data demonstrated significant improvements.
• The improved model showed enhanced generalization for solving the phase problem in larger unit-cell structures.

Conclusions:

• Artificial data generation is a viable strategy for enhancing neural network performance in crystallography.
• The developed methods enable scalable generation of crystal structures for training.
• The study highlights the potential of AI in advancing crystallographic structure determination.