Claude agents flag two magnetic semiconductor candidates

A Vals AI team used 90 Claude Opus 5.5 agents to search for room-temperature magnetic semiconductors. After three days of calculations and literature review, the system produced two leads: a newly designed oxide and a compound first synthesized in 1999 whose relevant electronic properties had gone unrecognized.

Vals AI released the findings in a project report and published a reproducible public ledger with simulation inputs, outputs, analysis scripts, and a one-command checker. Both candidates remain computational hypotheses. Neither has received the experimental tests needed to confirm its electronic structure.

Zero net moment, readable spins

The search targeted Luttinger-compensated magnets, a class of antiferromagnetic material with equal and opposite local magnetic moments. Those moments cancel overall, eliminating the net magnetization and stray fields associated with a ferromagnet. The atoms carrying opposite spins occupy inequivalent chemical environments, however, allowing the electronic bands for spin-up and spin-down electrons to separate in energy.

Magnetic random-access memory stores information in magnetic states and reads it electrically. Compensated magnets could support much faster switching and denser packing than conventional ferromagnets, while spin-split bands could provide a readable electrical signal. No semiconductor has yet demonstrated that full combination in a working room-temperature device.

A designed oxide meets a synthesis wall

The agents assembled the first candidate, YBaMnFeO₅, from five elements. Density functional theory predicted a 2.35 eV band gap, with the valence-band and conduction-band edges dominated by the same spin channel. The calculated spin-polarized windows were 1.0 eV for holes and 1.4 eV for electrons, far above the approximately 26 meV thermal-energy scale at room temperature.

The predicted magnetic-ordering temperature was approximately 420 K before the team’s calibration and 490 K afterward. Density functional theory approximates how electrons behave in a crystal, making it useful for estimating band structures and magnetic states. It does not prove that the required crystal can be synthesized or remain stable.

YBaMnFeO₅ requires manganese and iron atoms to occupy an ordered checkerboard. Simulations predicted that this arrangement would become a random mixture near 950 K, or 677°C. Comparable oxides are commonly fired between 900°C and 1,300°C, so a conventional synthesis would probably destroy the ordering responsible for the effect. The agents’ adversarial review therefore classified the material as a design study rather than a realizable discovery.

A 1999 compound gets a new label

A separate literature search surfaced KV[Cr(CN)₆], a compound related to Prussian blue. Its original chemists designed the chromium and vanadium moments to cancel. A 2008 study using a hybrid functional, a more computationally expensive treatment of electronic exchange, plotted electronic states consistent with spin-polarized band edges but did not identify the compound as a Luttinger-compensated semiconductor.

The new analysis connects that earlier evidence with several useful measurements and predictions:

• The calculated band gap is approximately 2.1 eV, with same-spin band edges and spin-polarized windows of 2.6 eV for holes and 1.6 eV for electrons.
• The sample reported in 1999 remained magnetically ordered up to 376 K, or 103°C.
• The crystal chemistry assigns each metal a fixed site: chromium bonds to the carbon end of cyanide, while vanadium bonds to the nitrogen end. This arrangement resists the site mixing predicted for YBaMnFeO₅.

The existing evidence has important limits. The only reported sample was a hydrated powder with a residual moment of 0.125 Bohr magnetons per formula unit, rather than the zero moment expected for an ideal crystal. The two computational methods used by the team also disagree about how strongly the water changes the predicted behavior.

Ninety agents divide the search

Vals AI’s 90-agent system split the project into parallel search lanes, literature review, simulation, and adversarial checking. Agents recorded pass-or-fail criteria before decisive calculations, fixing acceptance thresholds before seeing the results. Referee agents then attempted to invalidate each surviving claim through additional calculations and prior-art searches.

The calculations ran on Modal cloud CPUs with Quantum ESPRESSO 7.5, an open-source package for electronic-structure modeling. The Luttinger-compensated search lane alone submitted about 750 jobs.

The ledger maps all 61 reported claims to one of four verification paths:

The ledger also records five corrections made during compilation. Examples include an incorrect distance from the thermodynamic stability hull, an HSE06 calculation that stopped before convergence, a truncated output file, and the initially missed 2008 paper. Publishing those failures lets researchers distinguish the surviving evidence from superseded calculations.

What the bench must settle

Experimental work on KV[Cr(CN)₆] would require a fresh sample with measured hydration, composition, site occupancy, and magnetic ordering. Spin-resolved photoemission could then test whether the valence-band and conduction-band edges carry the predicted spin character. That technique measures electron energy and momentum together with spin, directly addressing the central computational claim.

Testing YBaMnFeO₅ would first require a synthesis route that preserves the manganese-iron checkerboard, followed by structural measurements to verify the ordering. Electronic and magnetic measurements would become relevant only after that structural hurdle is cleared.

Three days of agent-directed screening produced testable hypotheses with traceable computational provenance. The public inputs, outputs, scripts, corrections, and prior-art trail allow other researchers to rerun the work and challenge individual claims. The rediscovery of KV[Cr(CN)₆] also shows how automated literature review can connect older calculations with newer material classifications, provided every conclusion remains tied to auditable evidence.