Every electric vehicle on the road depends on magnets made from rare earth elements that are expensive, difficult to mine, and overwhelmingly processed in one country. A team at the University of New Hampshire thinks AI can find alternatives, and they have already identified 25 candidates.
Published in Nature Communications, the research describes how the team built an AI system that read decades of scientific literature, extracted experimental data on magnetic materials, and compiled the results into the largest searchable database of its kind: 67,573 magnetic compounds, including two dozen that had never been flagged as high-temperature magnets.
Why This Matters Right Now
The timing is not academic. China controls roughly 70% of global rare earth mining, 85% of refined neodymium and praseodymium processing, and 75% of the neodymium-iron-boron (NdFeB) magnet manufacturing that EV motors require. A typical electric vehicle uses 1 to 2 kilograms of these materials in its motor.
In April 2025, China imposed export controls on seven heavy rare earth elements. By October, those controls expanded to cover parts, components, and assemblies containing Chinese-origin rare earths. The effects were immediate: Ford paused Explorer production due to magnet shortages, European prices spiked to six times Chinese domestic rates, and automakers across the West scrambled to find alternatives.
“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,” said Suman Itani, the study’s lead author and a doctoral student in physics at UNH.
How the AI Works
The core problem the team solved is a data one. Thousands of scientific papers over the past 50 years have measured the magnetic properties of various compounds, but that data sits trapped inside PDFs and journal articles in inconsistent formats. No one had compiled it into a single, machine-readable resource.
The UNH team built a system called GPTArticleExtractor that uses large language models with prompt engineering and text tokenization to read scientific papers and pull out specific experimental details: chemical composition, crystal structure, lattice parameters, magnetic phase transition temperatures, coercivity, magnetization, magnetic moment, and more.
The extracted data went into the Northeast Materials Database (NEMAD), now publicly accessible at nemad.org. With 67,573 entries, it is far larger than any previous magnetic materials database.
What They Found
The team then trained machine learning models on the database to classify materials and predict their behavior. The classification model achieved 90% accuracy in sorting materials into three categories: ferromagnetic (attracted to magnets), antiferromagnetic (opposing magnetic alignment), and non-magnetic.
For predicting Curie temperature - the point at which a material loses its permanent magnetism - the regression model achieved a coefficient of determination (R-squared) of 0.87 with a mean absolute error of 56 Kelvin. For Neel temperature (the equivalent threshold for antiferromagnetic materials), R-squared hit 0.83 with a mean absolute error of 38 Kelvin.
These models identified 25 compounds that maintain magnetic properties at high temperatures but had never been recognized as such in existing literature. High operating temperature is a key requirement for EV motors, which generate significant heat during operation.
What This Means
The database itself may be more valuable than any single material it contains. Materials scientists have historically discovered new permanent magnets through intuition and trial-and-error, a process that has not produced a fundamentally new class of permanent magnet in decades despite thousands of known magnetic compounds.
NEMAD changes the search strategy. Instead of testing materials one at a time in a lab, researchers can now screen tens of thousands of candidates computationally, prioritize the most promising ones, and focus experimental work on a shortlist.
The 25 newly identified high-temperature magnets are candidates, not replacements. None has been tested as a permanent magnet for motor applications. The gap between “maintains magnetism at high temperature” and “works as a drop-in replacement for NdFeB in an EV motor” is significant. Permanent magnets need a specific combination of properties - high coercivity, high remanence, high energy product - that few materials achieve.
The Fine Print
This research identifies possibilities, not solutions. Even the most promising of the 25 new compounds would need years of experimental validation, engineering optimization, and manufacturing scale-up before appearing in a vehicle.
The AI extraction system, while achieving 90% classification accuracy, still carries a 10% error rate. Some entries in the database may contain inaccuracies inherited from the source papers or introduced during automated extraction. The Curie temperature predictions have a mean absolute error of 56 Kelvin, which is useful for screening but not precise enough to substitute for laboratory measurement.
The database focuses on magnetic transition temperatures and does not capture all the properties needed to evaluate a material as a permanent magnet candidate. Coercivity data, in particular, is reported inconsistently across the literature and harder to extract reliably.
The research was funded by the U.S. Department of Energy’s Office of Basic Energy Sciences, Division of Materials Sciences and Engineering. The team includes Itani, physics professor Jiadong Zang, and postdoctoral researcher Yibo Zhang.
Still, in a year when China’s export controls have turned rare earth access from a strategic concern into an operational crisis, a public database of 67,573 magnetic materials - and AI tools to search it - is a resource the EV industry did not have before.