How 90 Claude Opus 5.5 agents found a 27-year-old compound in 3 days to solve computing's heat problem

Over three days, 90 Opus 5.5 software agents screened simulations to find two room-temperature magnetic semiconductor candidates: YBaMnFeO₅ and KV[Cr(CN)₆].

Chemists synthesized KV[Cr(CN)₆] back in 1999. Its ability to filter electrons by spin while canceling external magnetism sat unnoticed for 27 years. Because its synthesis recipe already exists, researchers can immediately test it to build denser memory chips without magnetic cross-talk.

90 AI agents, 3 days, one 27-year-old compound

A recent computational search has now produced two candidates that appear to combine both properties. One is a newly designed compound, YBaMnFeO₅. The other is KV[Cr(CN)₆], a material first synthesized in 1999.


The work was carried out using more than 90 Claude Opus 5.5 agents over three days.

The unusual property being targeted is known as Luttinger compensation.

In a conventional antiferromagnet, neighboring magnetic atoms point in opposite directions and cancel each other's magnetic moments. But their electronic environments can also make the available states mixed between the two spin directions.

A Luttinger-compensated material takes a different route. The opposing magnetic atoms have equal and opposite moments, so the overall magnetism can cancel, but the atoms occupy inequivalent environments.

The result is a material that can have essentially zero net magnetic moment while still presenting electrons with spin-selective energy states. That combination is attractive for spintronics because it could provide spin information without the large stray magnetic field associated with ordinary ferromagnets.

The important quantity is the spin window — the energy range near the semiconductor's band edges in which the available states are dominated by one spin direction.

At room temperature, thermal energy is only about 26 meV. The calculations reported by Vals AI produce spin windows measured in fractions of an electron volt, much larger than that thermal scale.

The first candidate, YBaMnFeO₅, is a new design. The researchers say they could not find evidence that it had previously been made or proposed as a Luttinger-compensated magnet.

The calculations give it a predicted band gap of 2.35 eV. They also predict spin-sorted windows of about 1.0 eV for holes and 1.4 eV for electrons. Its magnetic ordering temperature was estimated at roughly 420 K in the raw calculation and about 490 K after calibration against a known magnet.

On paper, those numbers put the material comfortably above room temperature.

There is a major problem. The useful electronic structure depends on manganese and iron occupying an ordered checkerboard arrangement.

The simulations suggest that this arrangement breaks down at around 950 K. That matters because producing this class of oxide can require temperatures around 900–1300°C.

KV[Cr(CN)₆] changes the story because researchers don't have to start by inventing it.

The compound was first synthesized in 1999. It belongs to the family of Prussian-blue-type materials and was already known for unusually strong magnetic ordering. Independent literature describes a Curie or ordering temperature of about 376 K, or roughly 103°C.

The new analysis suggests that the material may also have the electronic structure required for Luttinger compensation.

The calculations put its band gap at about 2.1 eV. More strikingly, both sides of that gap are predicted to favor the same spin direction, with calculated spin windows of about 2.6 eV for holes and 1.6 eV for electrons.

The researchers also found something unusual in the history of the material.

A 2008 calculation had already plotted its electronic states using a hybrid functional. According to the new analysis, those plots contained evidence of the same-spin band edges, but the earlier work was focused on magnetic coupling under pressure and did not identify the compound as a Luttinger-compensated semiconductor.

Calling KV[Cr(CN)₆] a confirmed zero-magnetism semiconductor would go too far.

The existing sample was a hydrated powder rather than the perfect dry crystal used in the new calculations. It also showed a small residual magnetic moment of about 0.125 Bohr magnetons per formula unit.

The researchers' two computational approaches disagree on how much the water affects the spin-sorting behavior. The more sophisticated HSE06 calculation predicts that the effect survives, while the faster PBE+U calculation predicts a substantial reduction in the hole spin window.

Most importantly, neither the predicted 2.1 eV band gap nor the predicted spin sorting has yet been directly measured. That leaves a fairly clear experimental test.