Crystal Structure Prediction via Deep Learning

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Crystal structure repositories contain vast amounts of crystallographic data.
• Analyzing this data manually is challenging due to its scale and complexity.
• Machine learning offers potential for automated analysis of crystallographic information.

Purpose of the Study:

• To apply deep neural networks (DNNs) for analyzing crystallographic data.
• To train a DNN model to distinguish chemical elements based on their crystallographic environment.
• To predict the likelihood of forming new compounds using known structural templates.

Main Methods:

• Utilized multiperspective atomic fingerprints as input for the DNN model.
• Trained the neural network on a dataset of approximately 50,000 crystal structures.
• Applied the trained model to predict new compound formation based on structural templates.

Main Results:

• The DNN model successfully distinguished chemical elements by their crystallographic environment topology.
• Identified structurally similar atomic sites, revealing trends related to the periodic table.
• The model predicted known elemental compositions with high accuracy on unseen data, with ~30% found in the top 10 predictions.

Conclusions:

• The developed DNN approach effectively analyzes crystallographic data.
• This method can guide synthetic efforts in discovering new materials, particularly complex multi-element systems.
• The findings highlight the power of machine learning in materials discovery.