For the past year, most of my materials work on Ouro has pointed at one question: can we find a strong permanent magnet made only from cheap, abundant elements, with no rare earths? The pipeline generates candidate crystals, screens them with ML, and sends the survivors through DFT for magnetization, anisotropy, and Curie temperature.
That pipeline is good at producing candidates. It is not good at telling you which of them are real. So I'm now working with Jeff Snyder's group at Northwestern, pairing Ouro's computational side with people who know how these materials behave in a furnace.
Computation is outrunning synthesis
The first thing that came up when we sat down: generating samples on a computer is cheap, and making them in a lab is not. Autonomous labs help with volume, but they tend to produce many near-identical samples. Novel compositions and the process tweaks that make them work still mostly come from a grad student or postdoc at a bench.
There are also plenty of materials hiding in plain sight. Gadolinium nitride was identified as a phase fifty years ago and overlooked as a magnet, partly because you can't make it by arc melting. Some compounds get synthesized and then sit for years before anyone checks their magnetic properties.
Screen in the right order
Jeff's suggested priority for screening candidates, which I'm adopting for the Ouro pipeline:
- Thermodynamic stability. On or very near the convex hull.
- Ferromagnetic vs. antiferromagnetic. Confirm it's actually a ferromagnet.
- Saturation magnetization. Bohr magnetons per volume.
- Magnetocrystalline anisotropy. Last, and only if the rest holds up.
I had been leaning on anisotropy earlier than that, because it's the property that makes a magnet permanent. But it's also the most expensive to compute, and it doesn't matter if the phase can't exist. The other point that stuck with me: essentially all binaries and most simple ternaries have already been made, so the interesting space is systems with three or more elements.
A first target: Fe-W-B
We spent a while on the Fe-W-B phase diagram. I brought a tetragonal Fe₁₂W₄B candidate, and Jeff flagged that it had too many bonds to boron to be convincing. The compound that got more interesting was Fe₂₂B₆W (mp-1224794): iron-rich, with a single tungsten substituting into an iron boride structure. That's likely much easier to make than a lot of what generative models suggest.
Materials Project puts it about 148 meV above the hull. My ML-relaxed hull puts it lower. The gap is probably the difference between a machine-learned potential's energy surface and DFT's, and it's exactly the kind of discrepancy you want an experimentalist around for. Garbage in, garbage out: if the hull is wrong, everything built on it is wrong too.
What's next
- Recompute the Fe-W-B phase diagram with an updated hull and reconcile it with Materials Project.
- Check whether Fe₂₂B₆W and its neighboring ternary borides have ever been made and properly characterized, and prioritize them as first experimental targets.
- Advise Sebastian in the group on DFT and ML workflows as he builds out a higher-throughput synthesis effort.
- Open Ouro's magnet leaderboard beyond rare-earth-free candidates to all permanent magnets.
I'm excited about this one. Computational discovery with no lab attached tends to drift into producing nice-looking structures nobody can make. This keeps it honest.