Researchers from MIT, Harvard, Clemson use machine learning to predict magnetic structure of crystalline materials

A team of researchers at MIT, Harvard University and Clemson University has been using machine learning to predict the magnetic structure of crystalline materials.

According to a report by MIT News, this is critical for applications like data storage, high-resolution imaging, spintronics, superconductivity and quantum computing. Traditional methods for obtaining magnetic structures involve neutron diffraction and scattering studies, which are limited by the availability of time and machines.

“As a result, the magnetic structures of only about 1,500 materials worked out experimentally have been tabulated to date,” the MIT News report stated. “Researchers have also predicted magnetic structures by numerical means, but lengthy calculations are required, even on large, state-of-the-art supercomputers. These calculations, moreover, become increasingly expensive, with power demands growing exponentially, as the size of the crystal structures under consideration goes up.”

The researchers were led by Mingda Li, who is an MIT assistant professor of nuclear science and engineering, as well as Tess Smidt, an MIT assistant professor of electrical engineering and computer science. The first authors are three MIT undergraduates, including Helena Merker, Harry Heiberger, and Linh Nguyen, along with one Ph.D. student, Tongtong Liu.

They found a way to streamline the process through machine learning, and Smidt said it will not only be quicker, but also cheaper. The results of the report were published in iScience.

The team designed a neural network that can predict the magnetic structure of the crystalline materials through the use of “equivariant Euclidean neural networks.” The network ensures that the prediction is not affected by the rotation or translation of the crystal. This is especially helpful in examining 3D materials. The team had a database of about 150,000 substances compiled by the Materials Project at the Lawrence Berkeley National Laboratory. The input was used to assess two properties of given material, including magnetic order and magnetic propagation. 

Students used some of the Materials Project database to train the neural network to find correlations between the crystalline structure and magnetic structure. Optimal results were found when they included information about the atoms’ lattice positions, atomic weight, atomic radius, electronegativity and dipole polarization.

There is a testing phase that comes next where weights are used and predications are compared to previous values. The model had an average accuracy of about 78% and 74%, respectively, for predicting magnetic order and propagation, as reported in IScience. There was 91% accuracy for predicting the order of nonmagnetic materials, even if the material had magnetic atoms.