A computer-driven materials search has identified 344 candidate rare-earth-transition-metal magnets, with the highest modeled saturation magnetization reaching 1.837 tesla in SmFe12. Yet just five structures lay on the study's calculated DFT hull, its benchmark for thermodynamic stability, and experimental stability was not assessed.
A funnel from structures to candidates
The work, identified as arXiv:2608.25270v1 and dated 26 August 2026, asks whether diffusion-based crystal-structure generation can be combined with layered screening to find materials that pair high magnetization with calculated thermodynamic stability.
Researchers enumerated 186 unique binary compositions and 2,952 unique ternary compositions after normalizing them to primitive forms. The search required a minimum transition-metal fraction of 0.75 in binary systems and 0.8 in ternary systems, and it capped each unit cell at 20 atoms. MatterGen then generated crystal structures across valid supercell multiples.
Those generated structures went through a hierarchy of checks. A machine-learning interatomic potential provided the prescreening, and structures that passed were relaxed with spin-polarized VASP PAW-PBE density functional theory, or DFT, a calculation used here to validate their energies.
The scale of the funnel was large at the front and narrow at the end: 239,770 structures were generated, 7,178 passed the machine-learning screen, and 7,162 received DFT calculations. After the remaining filters, 344 were labeled final candidates and five were on the DFT hull.
The machine-learning potential tracked DFT very closely for absolute energy. The correlation was above 0.999, with a mean absolute error, or average difference in magnitude, of 0.013 to 0.043 eV per atom. But for energy above the hull, the correlation was weaker, at 0.40 to 0.65. All five true DFT-hull structures had passed prescreening.
Novelty was common, but not universal
The screen did not simply recover familiar crystal arrangements. Of 110 binary candidates across the Y and Sm systems, 78, about 71 percent, were classified as structurally novel, while 32 matched known binary prototype structures.
The ternary results were more mixed. The study reported 234 candidates: 139, about 59 percent, were novel; 91, about 39 percent, were derived from binary prototypes; and four, about 2 percent, directly matched known ternary prototypes.
The authors also reported strong symmetry relationships between a majority of newly identified ternary candidates and binary parent compounds. The report does not clarify how that qualitative majority relationship maps onto the separate prototype classifications.
Iron-rich examples led the magnetic results
The highest modeled values appeared in iron-rich compositions. Across the 344 candidates, SmFe12 reached 1.837 T, YFe12 reached 1.797 T and YFe18 Ti reached 1.799 T. Co-rich systems were around 1.257 T and Ni-rich candidates around 0.583 T. Among the calculated stable ternary phases, Y2 Co16 Mn reached 1.276 T.
Selected Y-Fe examples showed the same direction. Transition-metal fraction rose from 0.75 in YFe3, which had 1.10 T magnetization, to 0.89 in Y2 Fe17, which had 1.83 T; Y2 Fe7 was intermediate at 0.78 and 1.18 T. The paper presents these as selected modeled examples, without an uncertainty analysis.
The study also compared two calculated Sm3 Fe14 derivatives. The Mn case had a positive site-average moment of 1.96 μB and magnetization of 1.270 T, close to the parent's 1.303 T. The Cr case had a negative moment of -2.29 μB and magnetization of 1.177 T; neighboring Fe moments were 2.23 μB with Cr versus 2.07 μB in the parent.
The shortlist still needs experimental testing
The reported stability and magnetic values are calculated outputs. Experimental stability was not assessed, and the magnetic figures are modeled saturation magnetizations rather than measurements.
The search was bounded by its 20-atom unit-cell cap, leaving higher iron ratios outside the present search space. The authors propose searches in cells larger than 20 atoms to look for stable phases exceeding 1.8 T.
The underlying materials are available
The authors report that source code, calculation scripts, instructions and an interactive magnet-structure resource are available through the listed GitHub and UNC Charlotte URLs.
Paper data and sources
Original title: High-throughput Discovery of Magnetic Rare Earth Transition Metal Alloys
Authors: Shuo Tao, Osman Goni Ridwan, Liqin Ke, Qiang Zhu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text