CrystalMELA: a new crystallographic machine learning platform for crystal system determination

Area of Science:

• Crystallography
• Materials Science
• Computational Chemistry

Background:

• Determining crystal systems and space groups is crucial for crystal structure analysis.
• Manual intervention is often required for polycrystalline compounds, creating bottlenecks in material characterization.
• Automating crystal system classification can significantly streamline materials research workflows.

Purpose of the Study:

• To develop and present CrystalMELA, a novel machine learning (ML)-based web platform for automated crystal system classification.
• To provide accessible and user-friendly ML models for crystal system determination.
• To reduce the manual effort and time required in crystal structure analysis.

Main Methods:

• Development of a web platform, CrystalMELA, integrating ML models for crystal system classification.
• Training ML models, including random forest, convolutional neural network, and extremely randomized trees, on simulated powder X-ray diffraction patterns.
• Utilizing a dataset of over 280,000 published crystal structures from the POW_COD database, encompassing diverse compound types.

Main Results:

• Achieved a crystal system classification accuracy of 70% in tenfold cross-validation.
• Improved Top-2 classification accuracy to over 90%.
• Validated the trained ML models against independent experimental data, demonstrating their practical applicability.

Conclusions:

• CrystalMELA offers a powerful and user-friendly solution for automated crystal system classification.
• The platform effectively addresses the bottleneck in material characterization workflows.
• Future development can be community-driven, expanding the platform's capabilities.