Vals AI Uses Claude Agents to Find Magnetic Semiconductors

Vals AI recently deployed a swarm of more than 90 Claude Opus 5.5 agents to search for room-temperature magnetic semiconductors, specifically targeting Luttinger-compensated magnets. In just three days, the multi-agent system identified two candidate materials after running approximately 750 density functional theory calculations. The agents utilized Quantum ESPRESSO 7.5 on Modal cloud CPUs and employed an adversarial referee system to cross-examine and validate their own findings.

The first candidate, a newly designed five-element oxide named YBaMnFeO5, showed a predicted band gap of 2.35 eV. Its calculated spin-polarized windows reached 1.0 eV for holes and 1.4 eV for electrons, with an estimated magnetic-ordering temperature of 420 K before calibration and 490 K after. However, the agents' adversarial review flagged a major synthesis hurdle: the material requires an ordered manganese-iron checkerboard that simulations predict will scramble at 950 K (677°C). Because comparable oxides are typically fired between 900°C and 1,300°C, conventional synthesis would likely destroy the necessary structure.

The second candidate, KV[Cr(CN)6], was originally synthesized in 1999 but its potential as a room-temperature spin-sorted semiconductor had gone unnoticed. The agent system calculated a band gap of roughly 2.1 eV, with spin-polarized windows of 2.6 eV for holes and 1.6 eV for electrons. This material remains magnetically ordered up to 376 K (103°C) and resists the site-mixing issues of the oxide. Although the historical sample exhibited a residual moment of 0.125 Bohr magnetons per formula unit instead of the ideal zero, its crystal structure makes it a highly promising candidate for physical re-synthesis.

For materials scientists and AI practitioners, this deployment establishes a highly transparent, auditable workflow for autonomous discovery. Vals AI published a fully reproducible public ledger on GitHub mapping 61 claims to four verification paths, alongside five recorded corrections and a one-command checker. By automating the pipeline of literature review, simulation, and adversarial checking, this approach proves that LLM agents can rapidly surface and filter complex physical hypotheses, saving human researchers months of manual screening.