AI agents using Claude Opus find two candidate materials for computer memory

AI agents built on Anthropic’s Claude Opus 5.5 helped identify two candidate materials for next-generation computer memory, according to a post by a researcher published on vals.ai. One compound was designed from scratch. The other was a crystal first synthesized in 1999 that the agents recognized as having a magnetic property nobody had previously described.

The search targets a gap between two familiar kinds of magnet. Ferromagnets, like a fridge magnet, have atomic spins (the up-or-down orientation that gives electrons their magnetism) that all point the same way, which lets devices sort electrons by spin for storage but also creates a stray magnetic field that limits how densely bits can be packed. Antiferromagnets have spins that point opposite ways and cancel out, allowing denser, faster-switching storage, but their electrons can’t be sorted by spin, which makes them hard to read. The researcher was hunting for a middle category called Luttinger compensated, where spins still cancel to zero overall but sit in different atomic environments, allowing spin sorting without a stray field.

The first candidate, YBaMnFeO5, is a compound of yttrium, barium, manganese, iron and oxygen that the post says has apparently never been made or proposed before. Simulations predict a 2.35 eV band gap (the energy range where no electrons can sit, a hallmark of a semiconductor), with spin-sorted windows of 1.0 eV and 1.4 eV on either side, far larger than the roughly 26 meV of thermal jitter at room temperature. The design predicts magnetism surviving up to about 420–490 K. But it depends on manganese and iron atoms holding a precise checkerboard pattern, and simulations found that pattern collapsing into a random mix around 950 K, within the 900–1300°C range typically used to make such oxides.

A compound rediscovered

The second candidate, KV[Cr(CN)6], belongs to the same chemical family as Prussian blue, the centuries-old pigment. It was synthesized in 1999 by chemists who designed its two metals’ magnetism to cancel out, and a 2008 study plotted its spin-separated electron states without remarking on the significance. The new analysis says nobody had previously identified it as a Luttinger-compensated semiconductor or measured its spin-sorting windows. Simulations predict a band gap around 2.1 eV with spin windows of 2.6 eV and 1.6 eV. The 1999 sample itself stayed magnetically ordered up to 376 K (103°C), above room temperature, as measured by the chemists who made it.

The predictions carry caveats. The only existing sample is a water-containing powder that showed a small leftover magnetic moment rather than the perfect zero predicted for a dry crystal, and the two simulation methods used disagree on how much that water weakens the spin-sorting effect. Neither material’s band gap nor its spin sorting has been directly measured in a lab. The researcher published the simulation inputs, outputs and analysis code on GitHub, along with a checker script and known caveats, inviting others to verify the work.

The next step, according to the post, is to resynthesize KV[Cr(CN)6] and measure its spin sorting directly, since the compound already exists and only its newly identified magnetic class awaits experimental confirmation.