Ghost Projects

Magnets need to move into the 21st century. The NdFeB incumbent, which dominates over 60% of a $30B magnet market, was discovered in the 80s, and the field has not replaced it since.

Project 14 aims to create novel modeling methods for discovering the next generation of magnetic materials. Curie temperature, thermodynamic stability, magneto-crystalline anisotropy energy, and total magnetic density are already being accurately predicted by published Ouro models. The remaining bottleneck is the structure-prediction side: given a composition and a target symmetry, what crystal does it form?

GPSK-300 is our answer for the L1₀ tetragonal magnet family (FePt, CoPt, FeNi, FePd, MnAl, MnGa, MnAlC) and the hexagonal rare-earth–transition-metal magnets (SmCo₅, YCo₅, CeCo₅). A 304M-parameter multimodal diffusion transformer that generates crystals in reciprocal space, with a fully invertible 3-channel representation that decodes directly into a pymatgen Structure, with no auxiliary lattice regressor, element classifier, or MLIP volume scan.

The full write-up (methods, math, evaluation, failure modes):

→ GPSK-300: A Reciprocal-Space Diffusion Model for L1₀ Magnet Structure Prediction

The property side of the problem is data-limited. Magnetocrystalline anisotropy, the property that makes a permanent magnet hard, is absent from every large public database. We built and calibrated a screening dataset of 2,242 computed anisotropy labels for rare-earth-free uniaxial crystals, with relaxed structures and a quantified per-label error model:

→ A Calibrated Screening Dataset of Magnetocrystalline Anisotropy for Rare-Earth-Free Magnet Discovery