An AI Picked Two Superconductors. Then a Lab in Texas Actually Made Them.

In December 2025, a machine-learning screen pointed at two compounds that nobody had ever bothered to cool down and check: YRu3B2 and LuRu3B2. Ruthenium, boron, and a rare-earth metal, arranged in a lattice pattern borrowed from Japanese basket weaving. Both are now on the list of confirmed superconductors — not because a model predicted it, but because a group at Rice University fused the constituent elements into crystals, chilled them to within a degree of absolute zero, and watched the electrical resistance go to nothing.

The paper, Machine-learning-guided discovery of kagome superconductors YRu3B2 and LuRu3B2, was published in Physical Review Research on June 17, 2026 (DOI 10.1103/lpqj-7hyg), after first appearing as arXiv preprint 2512.16945 on December 16, 2025. The author list runs twelve deep and spans Aalto University in Finland, Rice University, Princeton, Ruhr University Bochum, and the Donostia International Physics Center in Spain — first author Rose Albu Mustaf, with Päivi Törmä, Miguel A. L. Marques, B. Andrei Bernevig, and Emilia Morosan among the senior authors. It is a product of SuperC, an international consortium founded in 2023 with the stated goal of finding a room-temperature superconductor by 2033.

Here are the numbers that matter. YRu3B2 superconducts below 0.81 kelvin. LuRu3B2 superconducts below 0.95 kelvin. Both were confirmed by three independent experimental probes — magnetization, specific heat, and electrical transport — and both showed nearly 100 percent superconducting volume fractions, meaning essentially the whole sample went superconducting, not a stray impurity phase faking the signal. That last detail is the difference between a measurement and a press release.

The funnel: model first, physics second, furnace third

The method is less glamorous than the phrase AI discovers superconductor implies, and considerably more credible for it.

Predicting whether a material superconducts from first principles means computing how its electrons couple to its lattice vibrations. That calculation is brutally expensive — hours to days of supercomputer time per compound — so it has almost never been run at scale. The team inverted the order of operations. A machine-learning model, trained on existing calculated materials data, does a cheap first pass over an enormous space of candidate element combinations, ranking them by how likely they are to be stable and superconducting. Only the survivors get the expensive quantum-mechanical treatment. Only the survivors of that get sent to a furnace.

Törmä, who leads SuperC, framed the scale of the problem bluntly in the Aalto University announcement: “Over the decades researchers have recognized over 7,000 superconductors, but mostly serendipitously. The process of identifying possible materials is so computationally heavy that, in fact, researchers have only been able to theoretically predict the viability of about 20 of these.”

On the method itself, she said: “Our method uses machine-learning-based pre-screening followed by targeted calculations on the promising candidates. This approach will greatly speed up superconductor discovery in the future. With machine learning, we may be able to push the number of materials we can process into the billions.”

Both compounds crystallize in a hexagonal CeCo3B2-type structure, in which the ruthenium atoms form flat kagome networks — the interlaced triangle-and-hexagon pattern of a traditional Japanese basket. Kagome lattices interest physicists because they can produce nearly flat electronic bands, where electrons pile up and interact strongly. The paper is candid about the outcome: compared with the related compound LaRu3Si2, the new materials have a more dispersive quasi-flat band and a stiffer phonon spectrum, both of which push the electron-phonon coupling down. Translation: the geometry helped find them, but the geometry did not deliver a high critical temperature. Superfluid weight calculations show conventional physics dominating over the exotic quantum-geometric effects the field has been chasing. The paper's own conclusion is modest — it demonstrates, the authors write, the effectiveness of integrating machine-learning screening, first-principles theory, and experimental synthesis.

Where AI-for-materials actually stands

Set against the field's recent credibility problems, this looks better than it sounds.

In November 2023, Google DeepMind announced GNoME, claiming 2.2 million new crystal structures and 380,000 stable candidates. In April 2024, Anthony Cheetham and Ram Seshadri published a perspective in Chemistry of Materials reporting scant evidence of compounds fulfilling the trifecta of novelty, credibility, and utility; at that point only a few hundred of the millions of predictions had been experimentally realized. The companion result — Berkeley's A-Lab, which claimed 41 new compounds from 58 targets in 17 days of autonomous operation — drew immediate objections from UCL materials chemist Robert Palgrave, who argued the X-ray diffraction data showed substitutional variants of already-known compounds rather than genuinely new ones. Nature issued a correction to that paper; as Chemical and Engineering News reported in January 2026, the original critics still do not consider the matter closed.

The kagome result sidesteps most of those failure modes by being small. It is a prospective prediction, posted publicly before synthesis, of two specific named compounds, followed by three orthogonal measurements and a full thermodynamic characterization in a physics journal. Two materials fully verified beats 380,000 materials unverified — and the honest accounting includes the fact that both land below one kelvin, colder than niobium, colder than plain lead. This is a proof of pipeline, not a proof of room-temperature anything.

Worth flagging: a version of this story circulating on aggregator sites recast it as a Materials Project consortium predicting and synthesising three high-temperature candidates with room-temperature potential. Wrong consortium, wrong count, and off by roughly 300 degrees. The real result is smaller, and considerably more interesting.

What to watch is whether the pre-screening hit rate survives contact with scale. Sifting a manageable family of kagome borides is one thing; Törmä's stated ambition of processing billions of candidates is another, and hit rates have a way of collapsing as search spaces expand. Watch also for the next batch of SuperC targets, and for whether independent groups can reproduce YRu3B2 and LuRu3B2. The consortium's self-imposed 2033 deadline is seven years out, and it has just added two entries to a list of roughly 7,000 — the first two, arguably, that anyone saw coming.