Machine learning predicts the magnetic properties of materials

Magnetic materials are in high demand. These are essential for energy storage innovations that electrification relies on, and for automation power supply systems. It is also found among the more familiar products, from consumer electronics to magnetic resonance imaging (MRI) machines.

As demand continues to grow, current sources and supply chains cannot be maintained. New magnetic materials need to be designed quickly.

Collaboration with Carnegie Mellon University, Lawrence Berkeley National Laboratory, and Fritz Harbor Institute DER MAX-PLANCK-GESELLSCHAFT expands the ability to screen for potential new materials in machine learning models.

The complexity of studying magnetic properties was a major limitation to material discovery. In nonmagnetic materials, the properties vary depending on what atoms are and how they are arranged.

“With magnetic materials, there's another degree of freedom,” says John Kitchin. “Each atoms in a magnetic field have a slightly magnetic vector, and their properties depend on the arrangement of these vectors.”

Even if the same atoms are in the same position, material properties may differ depending on the size and orientation of the magnetic vector.

Existing high-throughput methods for screening new materials do not take into account magnetic properties. For example, a functional theory of density and a fast model of machine learning models trained thereon can calculate energy, force, and thermodynamics. They lack the freedom of spin. Without the additional set of variables needed to predict magnetic properties, existing methods are inaccurate, too slow, or too expensive for the design of magnetic materials.

Researchers have developed a new machine learning model that can predict the magnetic properties of materials by distinguishing the arrangement of magnetic vectors. “This is the first model to explicitly have the freedom to make spin into an input parameter,” says Kitchen, a professor of chemical engineering at Carnegie Mellon.

Published in Proceedings of the National Academy of Sciencesthis work is a collaboration between Wenbin XU at Lawrence Berkeley National Laboratory and Rohan Yuri Sanspure and Adesh Korul, who contributed during his PhD. A chemical engineering student at Carnegie Mellon.

The studies of Kitchin, Xu, Sanspeur, and Kolluru also revealed new methods for data quality analysis. Models like the ones they developed are trained on a dataset consisting of hundreds of thousands or millions of calculations. It is infeasible to inspect all calculations, and these datasets usually contain several outlier points.

“This model allowed us to discover small clusters of unconverged data,” says Kitchin. “No one knew to check this before.” Once you can find an anomaly in your dataset, you can rerun the calculations with Kitchin, Xu, Sanspeur, and Kolluru to get better data and continue training your model.

The cutting-edge prediction accuracy and data efficiency of the models make it more feasible to understand the effects of magnetism on the design and catalysis of magnetic materials. Because calculations can be done cheaply, researchers try different optimization algorithms to sample all the different arrangements of magnetic vectors, and which arrangement calculates the lowest energy?

Models can be used, for example, to screen to create the next supermagnet through possible additives. You can identify which rare earth elements emphasize or reduce the magnetic field.

This model also opens the door to a more complete exploration of the role of magnetism in catalysis. “It's difficult to find the effect of magnetism on catalysts, so people missed out on geometry or adsorption phenomena,” says Kitchin. The calculations are expensive, and the difference in symmetry between the surface of a material and its bulk is often overlooked.

“More arrangements of spins are possible on the surface than they usually do in bulk,” explains Kitchin. “If we assume the surface looks bulky, then the energy arrangement could be the lowest.”

Models from Kitchin, Xu, Sanspeur, and Kolluru may help to find the reaction pathways for these other magnetic states.