← Back to feed

AI agents running on Claude Opus 5.5 have produced two candidates for a class of material that materials scientists have sought for decades: antiferromagnetic semiconductors that can sort electrons by spin at room temperature, enabling next-generation spintronic memory devices.

The research came from Vals.ai. The team reported two Luttinger-compensated (LC) semiconductor candidates:

• YBaMnFeO₅ — a newly designed compound, computationally synthesised by the agents
• KV[Cr(CN)₆] — a material first made in 1999 that had not been identified as an LC semiconductor candidate until now

KV[Cr(CN)₆] was empirically measured at 376K (103°C) magnetic ordering temperature in its original 1999 synthesis by Holmes et al. — comfortably above room temperature. The agents’ DFT predictions for the newly designed YBaMnFeO₅ compound pointed to an ordering temperature near 420K.

Why This Material Class Matters

Standard ferromagnets used in hard drives sort electrons by spin orientation, but their leaking magnetic fields interfere with adjacent components, and switching is slow — on the order of nanoseconds. Antiferromagnets avoid the field leak and switch roughly 1,000 times faster, but their spin populations are mixed at each energy level, making information readout difficult.

Luttinger-compensated materials sit between the two extremes. They carry zero net magnetic moment — no field leaks — but their spin-up and spin-down atoms sit in structurally inequivalent environments. That inequivalence creates spin-sorted energy bands that spintronics circuits can read. Adding semiconducting behaviour means the spin current can be gated electrically, combining the speed and density advantage of antiferromagnets with the readability of ferromagnets.

The Agent Workflow

Vals.ai described a team of Opus 5.5 agents working through materials screening: evaluating crystal structures, predicting magnetic ordering temperatures, and screening for semiconducting band gaps. The 1999 rediscovery of KV[Cr(CN)₆] is notable — the compound was synthesised by human researchers at the time but its Luttinger-compensated semiconductor properties were not identified.

Both materials remain simulation and calculation candidates. Experimental synthesis at scale and device integration are separate steps. No manufacturable spintronic device was demonstrated. The contribution is two specific materials worth pursuing experimentally, with one already having a prior synthesis record as evidence of feasibility.

Context

The result adds to a short but growing list of AI-assisted materials discoveries. Vals.ai has been using Anthropic’s frontier models on scientific reasoning tasks. Prior published work with Opus 5.5 agents has focused on computer science and formal mathematics, including Lean-verified algorithm proofs. The materials universe is large enough that systematic AI search is expected to keep surfacing overlooked compounds from prior decades alongside genuinely new ones.