An international physics collaboration coordinated by Aalto University in Finland has used machine learning to help discover two previously unknown superconducting materials — YRu₃B₂ and LuRu₃B₂ — in a study that demonstrates how AI-accelerated screening can be paired with first-principles physics and lab synthesis to find real, working quantum materials rather than just predicted ones on paper.
The pipeline: screening, then calculating, then building
The discovery didn’t come from a single AI model spitting out an answer. It came from a three-stage pipeline. First, a machine-learning model pre-screened a very large space of possible elemental combinations, ranking them for how plausible they were as superconductors — a job that would be computationally out of reach for exhaustive first-principles calculation on every candidate. Second, the small set of most promising candidates that survived that pre-screening was run through targeted first-principles quantum calculations to check the physics in detail. Third, the compounds that passed both computational filters were actually synthesized and tested in the lab, with sample synthesis led by Emilia Morosan’s group at Rice University.
That last step matters. A predicted superconductor is a hypothesis; a synthesized one that shows a real, measured drop in electrical resistance is a confirmed result. The two materials to make it all the way through — YRu₃B₂ and LuRu₃B₂ — showed bulk superconductivity at critical temperatures of 0.81 K and 0.95 K respectively, according to the published paper. Those are extremely low temperatures, close to absolute zero, and nowhere near practical operating conditions. The significance isn’t the temperature — it’s that the pipeline correctly predicted, ahead of any lab work, that these specific never-before-synthesized compounds would superconduct at all.
Why kagome lattices
Both new compounds get their superconductivity from ruthenium atoms arranged in a kagome lattice — a hexagonal, interlocking pattern named after a traditional Japanese woven-bamboo basket design. Kagome lattices are of particular interest in condensed-matter physics because their geometry can produce “flat” electronic bands, where large numbers of electrons share very similar energy states. That flatness tends to amplify the effects of electron-electron interactions, which is part of why kagome materials have become a active hunting ground for unconventional quantum behavior, including superconductivity, in recent years.
Why the “how” is the actual story
Materials scientists estimate that roughly 7,000 superconducting materials have been identified since the phenomenon was first observed over a century ago — the overwhelming majority of them found through trial-and-error experimentation rather than prediction. Only a small number, on the order of a few dozen, had previously been correctly predicted by theory before being made. That imbalance is the real bottleneck in superconductor research: physicists have a decent qualitative understanding of what favors superconductivity, but turning that understanding into “here is a specific new compound worth attempting to synthesize” has historically been slow and expensive.
This project is a proof of concept that machine-learning pre-screening can narrow an enormous combinatorial space of candidate compositions down to a shortlist small enough for rigorous first-principles theory and, ultimately, real synthesis — compressing a search that would otherwise be computationally or experimentally prohibitive. The two confirmed compounds are the first public result of that pipeline, not necessarily the last.
The consortium behind it
The work comes out of SuperC, an international research consortium coordinated by Aalto University and led by physicist Päivi Törmä, formed in 2023 with the explicit long-term goal of finding a room-temperature superconductor by 2033. Room-temperature superconductivity — carrying electrical current with zero resistance without extreme cooling — remains unrealized and is a substantially harder target than what was demonstrated here; nothing in this result claims to have achieved it. What this result does establish is that the consortium’s AI-guided discovery approach can find genuinely new, previously unsynthesized superconductors, which is the methodological building block the longer-term goal depends on.
Publication
The findings were published in Physical Review Research (volume 8, issue 2, 2026; published 7 July 2026) under the title “Machine Learning-Guided Discovery of Kagome Superconductors YRu₃B₂ and LuRu₃B₂,” with contributing authors from Aalto University, Rice University, and collaborating institutions. The paper’s DOI is 10.1103/lpqj-7hyg. Aalto University’s own account of the discovery is available in its press release.
Why it matters for research infrastructure
For research-administration and research-computing audiences, the notable part of this story isn’t the two specific compounds — at well under 1 Kelvin, neither is close to an applied technology on its own. It’s that a screening pipeline combining machine learning, first-principles computation, and targeted experimental synthesis produced a verified, publishable, peer-reviewed result on the first public attempt. That’s the kind of computational-plus-experimental workflow that funders and research-infrastructure planners are increasingly being asked to resource: not AI replacing physics, but AI narrowing a search space so that expensive theory and lab time gets spent on the candidates most likely to pay off.
