Machine learning magnetism classifiers from atomic coordinates.

Area of Science:

• Condensed matter physics
• Materials science
• Computational materials science

Background:

• Determining magnetic structure is a significant challenge in condensed matter physics and materials science.
• Experimental methods like neutron diffraction are resource-intensive and require complex refinement.
• Computational methods such as density functional theory (DFT) often need semi-empirical corrections and are limited to collinear magnetism.

Purpose of the Study:

• To develop a machine learning model for classifying magnetic structures.
• To predict magnetic ordering (ferromagnetic, antiferromagnetic, non-magnetic) and magnetic propagation vectors.
• To overcome limitations of current experimental and computational approaches for magnetic structure determination.

Main Methods:

• Utilized atomic coordinates as input for a machine learning model.
• Developed a Euclidean equivariant neural network that respects crystallographic symmetry.
• Trained the model on materials containing transition metal and rare earth elements.

Main Results:

• Achieved an average accuracy of 77.8% for magnetic structure classification.
• Achieved an average accuracy of 73.6% for predicting magnetic propagation vectors.
• Reached 91% accuracy in identifying non-magnetic ordering, even in the presence of magnetic elements.

Conclusions:

• The machine learning model offers a promising approach for predicting magnetic structures.
• This work represents a step towards the grand challenge of full magnetic structure determination.
• The developed model enhances the predictive capabilities for magnetic materials.