Machine learning model predicts magnetic properties of materials

Magnetic materials are in high demand. They're essential to the energy storage innovations on which electrification depends and to the robotics systems powering automation. They're also inside more familiar products, from consumer electronics to magnetic resonance imaging (MRI) machines.

Current sources and supply chains won't be able to keep up as demand continues to grow. We need to design new magnetic materials, and quickly.

A collaboration between Carnegie Mellon University, Lawrence Berkeley National Laboratory, and the Fritz-Haber-Institut der Max-Planck-Gesellschaft is broadening capabilities to screen potential new materials with machine learning models.

The complexity of studying magnetic properties has been a big limitation for materials discovery. In nonmagnetic materials, the properties depend on what kind of atoms are there and how they are arranged. "With magnetic materials, there's one more degree of freedom," says John Kitchin. "On each atom that's magnetic, there is a little magnetic vector, and the properties depend on the arrangement of those vectors." Even when the same atoms are in the same positions, the material properties can differ depending on the magnitude and orientation of the magnetic vectors.

Existing high-throughput methods for screening new materials don't account for magnetic properties. Density functional theory and the faster machine learning models trained on it, for example, can compute energy, forces, and thermodynamics. They lack spin degrees of freedom. Without that additional set of variables needed to predict magnetic properties, existing methods are inaccurate, too slow, or too expensive for designing magnetic materials.

This is the first model that explicitly has the degrees of freedom for allowing the spin to be an input parameter.

John Kitchin, Professor, Chemical Engineering

Researchers developed a new machine learning model that can predict the magnetic properties of materials by differentiating the arrangement of magnetic vectors. "This is the first model that explicitly has the degrees of freedom for allowing the spin to be an input parameter," says Kitchin, professor of chemical engineering. Published in Proceedings of the National Academy of Sciences, the work is a collaboration with Wenbin Xu at Lawrence Berkeley National Laboratory and Rohan Yuri Sanspeur and Adeesh Kolluru, who contributed while they were Ph.D. students in chemical engineering at Carnegie Mellon.

Kitchin, Xu, Sanspeur, and Kolluru's research also revealed a new method for data quality analysis. Models like the one they developed are trained on datasets consisting of hundreds of thousands or even millions of calculations. It's infeasible to inspect every calculation, and these datasets typically contain a few outlier points.

"This model allowed us to discover little clusters of the data that were not converged," says Kitchin. "Nobody knew to check for this before." Once they could find the anomalies in the dataset, Kitchin, Xu, Sanspeur, and Kolluru could re-run the calculations and get better data to continue training their model.

With the model's state-of-the-art prediction accuracy and data efficiency, designing magnetic materials and understanding the effects of magnetism in catalysis are now more feasible. Because calculations can be done cheaply, researchers can try different optimization algorithms to sample all the different arrangements of magnetic vectors and quickly calculate which arrangement has the lowest energy.

The model could be used, for example, to screen through possible additives to make the next supermagnet. It could identify which rare earth elements accentuate a magnetic field or which reduce it.

The model also opens the door to a fuller exploration of the role of magnetism in catalysis. "People have missed geometries or adsorption phenomena because the effects of magnetism in catalysis are hard to find," says Kitchin. The calculations are expensive, and the symmetry differences between a material's surface and its bulk are often overlooked. "More arrangements of the spins are possible at a surface than are in the bulk normally," explains Kitchin. "If you assume the surface looks like the bulk, you may miss the lowest energy arrangement." Kitchin, Xu, Sanspeur, and Kolluru's model may help to find reaction pathways in these other magnetic states.

For media inquiries, please contact Lauren Smith at lsmith2@andrew.cmu.edu.