Zero Net Magnetism, Sorted Spins: Claude Opus 5.5 Agents Find Two Room-Temperature Spintronic Candidates

The hardest problems in computer memory physics are rarely solved by writing prompts. But on October 4, 2026, AI-evaluation firm Vals AI published something unusual: two candidate materials for next-generation spintronic memory, one designed from scratch and one rescued from a 1999 chemistry paper, both found by a team of Claude Opus 5.5 agents running quantum-mechanical simulations. The work landed on the Hacker News front page within hours, where it drew more than a hundred upvotes and ninety comments — largely because the authors published everything: input files, raw outputs, analysis code, a one-command checker, independent re-runs, and a frank list of known caveats.

Why magnets are a memory problem

The physics primer in the Vals post is worth the 90 seconds it asks for. Every electron carries a quantum property called spin, which gives it a magnetic moment and can be modeled as pointing up or down. Spintronics — the technology behind hard-drive read heads and MRAM — stores information in these spin orientations. To do that usefully, you need a material that sorts electrons by spin at the edge of its band gap: a “spin window” where every available electron state has the same spin. The larger that window is relative to the ~26 meV of thermal jiggling at room temperature, the more reliably the electrons stay sorted.

The two familiar kinds of magnets each fail this test in opposite ways. Ferromagnets (fridge magnets) sort spins by energy naturally, but their leaking macroscopic magnetic field interferes with neighboring devices, and switching them is slow and power-hungry. Antiferromagnets have no stray field, can be packed far denser, and switch roughly a thousand times faster — but their up and down atoms sit in equivalent environments, so spins stay mixed at every energy level and there is nothing for spintronic read-out to grab onto.

The material the field actually wants is a unicorn: a semiconductor with a band gap, zero net magnetism, and spin sorting that survives at room temperature. Physicists call this class Luttinger-compensated (LC) magnets, after Luttinger’s theorem, which locks the net spin moment of an insulator’s repeating crystal unit to a whole number — once zero, always zero. In an LC material the spin-up and spin-down atoms sit in inequivalent environments (different elements, or the same element in two different sites), which breaks the symmetry and lets spins be sorted by energy exactly as ferromagnets do, while still cancelling macroscopically. A 2025 study that predicted two other LC semiconductors found that both lose their magnetic order below room temperature, and explicitly named a room-temperature LC semiconductor as the open goal.

How the agents worked

The workflow Vals describes is closer to a research lab than a chatbot session. The agents ran density functional theory (DFT) simulations — the standard quantum-mechanical method for crystal electronic structure — at two levels of approximation: fast PBE+U and the slower, usually more accurate HSE06. All band gaps and spin windows reported come from the higher-accuracy HSE06 calculations.

Candidate 1 is a designed material: YBaMnFeO₅. A new compound of five elements (yttrium, barium, manganese, iron, oxygen) that, as far as the team could find, has never been synthesized or proposed as this kind of magnet. The simulations predict it is a semiconductor with a 2.35 eV band gap and spin sorting on both sides: a 1.0 eV window for holes and 1.4 eV for electrons — tens of times larger than room-temperature thermal noise. The agent also predicted magnetic order persisting to roughly 420 K raw, or about 490 K after calibrating the simulation against a known magnet. The catch is synthesis: the design requires Mn and Fe atoms to sit in a perfect checkerboard, and when the agents simulated atomic arrangement at different temperatures, that checkerboard collapses into a random mix around 950 K. Since oxides of this type are typically made at 900–1300 °C and atoms barely move at lower temperatures, standard synthesis would likely scramble the crystal — destroying the spin sorting. A promising design, but probably hard to actually make.

Candidate 2 is the rediscovery: KV[Cr(CN)₆]. Here the agent found something hiding in plain sight. The compound, a member of the same Prussian-blue pigment family chemistry as the 300-year-old paint, was first synthesized in 1999. Its zero net magnetism was deliberate — the original chemists designed the two metals’ magnetism to cancel. Even its spin sorting existed on paper: a 2008 study using hybrid functionals plotted its electron states spin by spin, and both band edges carry the same spin, but that paper was about magnetic coupling under pressure and never remarked on it. Nobody had identified the material as a Luttinger-compensated semiconductor, quantified its spin-sorted windows, or tested their robustness — until now.

The predictions for KV[Cr(CN)₆] are strong: a band gap of about 2.1 eV, with spin windows of 2.6 eV for holes and 1.6 eV for electrons, and — critically — magnetic order in the 1999 sample measured up to 376 K (103 °C), above room temperature (365 K after the sample had been heated). Most importantly, its structure locks each metal into its own site: chromium bonds to the carbon end of each cyanide ligand, vanadium to the nitrogen end. That is exactly the site-locking YBaMnFeO₅ lacks.

The caveats are stated plainly. The predictions apply to perfect, dry crystals. The only existing sample, from 1999, is a powder with water in its pores, and it showed a small residual magnetic moment of 0.125 Bohr magnetons per formula unit where a perfect crystal would show zero. The two simulation methods disagree on how much the water weakens the effect: HSE06 says the spin sorting survives; PBE+U says the hole window shrinks by more than half. Neither the band gap nor the spin sorting has ever been measured. The next step is to resynthesize KV[Cr(CN)₆] and measure its spin sorting directly.

Why this matters beyond one blog post

Two things make this more than a curiosity. First, the “hiding in plain sight” finding is a genuine pattern-shift result: the argument advanced in the Vals post is that materials for the next step in spintronics may already exist in the literature, waiting to be recognized. A compound synthesized 27 years ago, characterized honestly by its makers, plotted spin-by-spin in a 2008 paper with nobody connecting the dots — that is exactly the class of cross-literature synthesis that long-context agents with simulation tools are starting to do well. It is the same shape as AI systems that found faster shortest-path algorithms or new proof strategies: not replacing scientists, but compressing the search over everything already known.

Second, the reproducibility standard is the real headline for AI-for-science methodology. Every computation behind the predictions — input files, raw DFT outputs, analysis code, a one-command checker, independent re-runs, and a written list of known caveats — is public in a GitHub repository. In a field where AI “discoveries” often arrive as press releases with appendices, a fully auditable ledger including the design calculations that produced the candidates is closer to how the discipline should work. The Hacker News thread’s engagement reflects that: commenters could actually pull the repo and check.

The honest framing matters too. These are predictions, not discoveries — DFT-based candidates whose band gaps and spin windows have never been touched by an instrument. One of the two is probably unmakeable in useful form; the other needs a 27-year-old powder re-synthesized and measured. But the pattern is the point: an agent team that designs a novel compound, predicts its failure mode, then re-finds a forgotten candidate that dodges it, all in a transparent public ledger, is a credible preview of how a meaningful slice of materials discovery will run in 2027.