Magnetic magic: How AI is revolutionizing matter discovery

Ames National Laboratory scientists have devised a machine learning model to predict new magnet materials without using scarce elements. This innovative approach, focusing on the material’s Curie temperature, offers a more sustainable path forward for future technological applications.

Scientists use AI to find new magnetic materials that don’t have key elements.

A team of researchers from the US Department of Energy’s Ames National Laboratory has developed a new material machine learning model for exploring permanent magnet materials without critical elements. The model predicts the Curie temperature of new material combinations. This is an important first step in using artificial intelligence to predict new permanent magnet materials. This model adds to the team’s recently developed capabilities to explore thermodynamically stable rare earth materials.

The importance of high performance magnets

High-performance magnets are essential for technologies such as wind power, data storage, electric vehicles and magnetic refrigeration. These magnets contain important materials such as cobalt and rare earth elements such as Neodymium and Dysprosium. These materials are in high demand but limited supply. This situation is motivating researchers to find ways to design new magnetic materials with reduced amounts of key material.

Magnet image. Source: US Department of Energy’s Ames National Laboratory

The role of machine learning

Machine learning (ML) is a form of artificial intelligence. It is driven by computer algorithms that use data and trial-and-error algorithms to continuously improve its predictions. The team used experimental data on the Curie temperature and a theoretical model to train the ML algorithm. The Curie temperature is the maximum temperature at which a material remains magnetic.

“Finding compounds with high Curie temperatures is an important first step in discovering materials that can Maintains magnetic properties at high temperatures. “This aspect is important for the design of not only permanent magnets but also other functional magnetic materials.”

According to Mudryk, discovering new materials is a challenging activity because the search has traditionally relied on experimentation, which is expensive and time-consuming. However, using ML methods can save time and resources.

Model development

Prashant Singh, a scientist at Ames Laboratory and a member of the research team, explains that a key part of this effort is to develop ML models using basic science. The team trained their ML model using experimentally known magnetic materials. Information about these materials establishes the relationship between certain electronic and atomic structure features and the Curie temperature. These patterns provide computers with a basis for searching for potential candidate documents.

Test and validate the model

To validate the model, the team used Cerium, Zirconium and Iron-based compounds. This idea was proposed by Andriy Palasyuk, a scientist at Ames Laboratory and a member of the research team. He wanted to focus on unknown magnetic materials based on earth-abundant elements. “The next super magnet must not only have outstanding performance, but also rely on abundant domestic components,” Palasyuk said.

Palasyuk worked with Tyler Del Rose, another scientist at Ames Laboratory and a member of the research team, to synthesize and characterize the alloy. They found that the ML model was successful in predicting the Curie temperature of the candidate materials. This success is an important first step in creating a new high-performance permanent magnet design method for future technological applications.

“We are writing physics-based machine learning for a sustainable future,” Singh said.

Reference: “Physics-based machine learning prediction of the Curie temperature and its promise in guiding the discovery of functional magnetic materials” by Prashant Singh, Tyler Del Rose, Andriy Palasyuk and Yaroslav Mudryk , August 2, 2023, Materials chemistry.
DOI: 10.1021/acs.chemmater.3c00892