Researchers at the University of New Hampshire have built an AI system that can read scientific papers, extract experimental data, and predict which materials will make good magnets. The result: a database of 67,573 magnetic compounds and the identification of 25 previously unrecognized materials that stay magnetic at high temperatures.
The discovery matters because today’s powerful magnets rely on rare earth elements that are expensive, environmentally damaging to extract, and largely controlled by China. Electric vehicles, wind turbines, and countless other clean energy technologies need alternatives.
How NEMAD Works
The team, led by doctoral student Suman Itani along with physics professor Jiadong Zang and postdoctoral researcher Yibo Zhang, built a system called the Northeast Materials Database (NEMAD) in two stages.
First, they trained a large language model to read scientific papers and extract specific experimental details about magnetic materials: chemical composition, crystal structure, magnetic properties, and phase transition temperatures. This automated what would take human researchers countless hours of manual literature review.
Second, they trained two machine learning models on the extracted data. A classification model determines whether a material is ferromagnetic, antiferromagnetic, or non-magnetic with 90% accuracy. Regression models predict Curie temperatures (the point where a material loses its magnetism) with an R-squared value of 0.87, and Neel temperatures with 0.83.
What They Found
Using these models to scan the Materials Project database, the team identified 25 ferromagnetic candidates with predicted operating temperatures above 500 Kelvin (227 degrees Celsius) and 13 antiferromagnetic candidates stable above 100 Kelvin.
These operating temperatures matter. Permanent magnets in electric vehicle motors get hot. Materials that lose their magnetism at modest temperatures are useless for these applications. The rare earth magnets currently used, like neodymium-iron-boron compounds, work precisely because they maintain their magnetic properties under demanding conditions.
“By accelerating the discovery of sustainable magnetic materials, we can reduce dependence on rare earth elements, lower the cost of electric vehicles and renewable-energy systems, and strengthen the U.S. manufacturing base,” Itani said.
The Fine Print
The research, published in Nature Communications (DOI: 10.1038/s41467-025-64458-z) and funded by the U.S. Department of Energy’s Office of Basic Energy Sciences, represents promising early-stage work. The 25 identified materials are candidates, not proven replacements.
Several hurdles remain before any of these materials could end up in your next EV:
Experimental validation. The AI predicted these materials should work. Someone still needs to synthesize them and test whether the predictions hold.
Performance gaps. Even if the materials perform as predicted, matching the magnetic strength of rare earth magnets is a separate challenge. A material that stays magnetic at high temperatures but produces weak magnetic fields won’t cut it.
Manufacturing scalability. Lab-synthesized materials and mass-produced motor components are different problems entirely.
What’s genuinely valuable here isn’t any single material discovery. It’s the methodology. The NEMAD database and the AI pipeline that created it can accelerate materials science research broadly. The team notes that their large language model approach could extend to other database creation tasks across scientific fields.
The database is publicly available for other researchers to build on. Given that the U.S. currently imports roughly 80% of its rare earth elements, mostly from China, finding domestic alternatives has become both an economic and strategic priority.