High-throughput Discovery of Magnetic Rare Earth Transition Metal Alloys

Modern life runs on invisible forces. The electric motors that drive cars, the generators that power wind turbines, and the tiny drives inside computers all depend on permanent magnets. These are materials that hold a magnetic field without needing electricity to keep them active. For decades, the best magnets have relied on a specific family of metals: rare earth elements mixed with transition metals like iron or cobalt. While these materials work incredibly well, they are expensive and their supply chains are fragile. Scientists have long suspected that there are other, better combinations of atoms waiting to be discovered, but the number of possible mixtures is so vast that trying them one by one in a lab would take centuries.

To solve this, researchers have turned to a new kind of search. Instead of mixing chemicals in a beaker, they use powerful computers to imagine millions of new crystal structures. These structures are the repeating patterns in which atoms arrange themselves. By using artificial intelligence to generate these patterns and then testing them with advanced physics simulations, scientists can find promising candidates before ever stepping into a laboratory. This approach allows them to look for materials that are not only magnetic but also stable enough to exist in the real world. The goal is to find magnets that are stronger and more reliable than anything currently available, potentially using more common elements to reduce costs and supply risks.

A team of researchers at the University of North Carolina at Charlotte and the University of Virginia has now applied this high-speed discovery method to the search for new rare-earth magnets. They focused on combinations of rare earth elements, specifically yttrium and samarium, mixed with iron, cobalt, and nickel, along with a few other metals to act as stabilizers. The team started by asking a generative artificial intelligence model to create crystal structures for over three thousand different chemical recipes. The model produced nearly 240,000 unique arrangements of atoms. This was not a random guess; the model was trained on known stable materials and conditioned to create structures that fit specific chemical rules, ensuring the resulting patterns were physically plausible.

Once the computer generated these 240,000 possibilities, the researchers had to filter them down to the most promising few. They used a two-step screening process. First, they ran a fast, machine-learning simulation to check which structures were likely to be stable. This step eliminated the vast majority of the candidates, leaving only a few thousand that looked worth a closer look. Then, they subjected these survivors to a much more rigorous and accurate physics calculation known as density functional theory. This step confirmed which structures were truly stable and calculated their magnetic strength. The result was a list of more than 300 low-energy magnetic candidates. Among these, five were found to be thermodynamically stable, meaning they should be able to exist naturally without falling apart.

The most exciting finding was the strength of the magnets they discovered. The best candidates were rich in iron, the element that provides the bulk of the magnetic power. One specific combination, involving samarium and iron, reached a saturation magnetization of about 1.8 tesla. This is a measure of how strong the magnetic field is when the material is fully magnetized. For comparison, this is significantly stronger than many current commercial magnets. Another top performer, a mix of yttrium, iron, and titanium, also reached nearly 1.8 tesla. The researchers found that the strength of the magnet depended heavily on how many iron atoms were packed into the structure. The more iron they could fit into the crystal lattice, the stronger the magnet became.

The study also revealed how these new magnets are built. Most of the successful ternary compounds—those with three different types of metals—were not entirely new shapes. Instead, they were variations of known binary structures where one type of atom was swapped for another in a very specific way. The researchers found that the new metals, such as manganese or titanium, fit into the crystal by splitting the positions where the original atoms sat. This process, known as Wyckoff site splitting, allowed the new atoms to stabilize the structure without destroying the magnetic alignment. The analysis showed that manganese was a particularly good choice for these substitutions because it aligned its magnetic direction with the iron atoms, adding to the overall strength. In contrast, chromium aligned in the opposite direction, which weakened the magnet. This provides a clear guide for future experiments: if you want a stronger magnet, choose dopants that agree with the iron.

While the computer found these 300 candidates, the researchers noted that the search was limited by the size of the structures the artificial intelligence could generate. The model was restricted to crystals containing no more than 20 atoms. The team suggests that if they can expand this search to larger crystals with more than 20 atoms, they might find even more iron-rich structures with magnetization exceeding 1.8 tesla. The work demonstrates that this generative approach is a powerful tool for navigating the complex landscape of materials science. It moves beyond the slow, trial-and-error methods of the past, offering a systematic path to discovering materials that could one day power the next generation of electric vehicles and green energy technologies. The path forward is clear: refine the models to handle larger structures and test these predicted candidates in the lab to see if they can be made real.