Using a machine learning approach to determine the space group of a structure from the atomic pair distribution...

Area of Science:

• Crystallography
• Materials Science
• Machine Learning

Background:

• Accurate determination of crystal structure space groups is crucial for understanding material properties.
• Traditional methods like powder diffraction can be limited in scope and require model-based analysis.
• Atomic Pair Distribution Function (PDF) analysis offers an alternative, model-independent approach to structural characterization.

Purpose of the Study:

• To develop and evaluate a machine learning-based method for predicting space groups directly from atomic pair distribution function (PDF) data.
• To assess the performance of a convolutional neural network (CNN) for this prediction task.
• To compare the performance and failure modes of the CNN model against conventional structure indexing algorithms.

Main Methods:

• Development of machine learning models, specifically a convolutional neural network (CNN), trained on a large dataset of over 100,000 atomic PDFs.
• Training data encompassed structures from the 45 most common space groups.
• Testing the CNN model on both calculated and experimental PDF datasets.

Main Results:

• The CNN model correctly identified the space group among the top-6 estimates with 91.9% accuracy for calculated PDFs.
• The model successfully identified space groups for 12 out of 15 experimental PDF datasets.
• Analysis of failed predictions revealed similarities to limitations encountered by conventional indexing algorithms.

Conclusions:

• Machine learning, particularly CNNs, shows significant promise for model-independent space group prediction from PDF data.
• The developed method offers a viable alternative or complementary approach to traditional crystallographic analysis.
• Further development could enable broader application of this technique across various material systems.