Machine learning just found two new superconductors, and the search method matters more than the materials
Humanity has found roughly 7,000 superconducting materials since the phenomenon was first observed in 1911. Almost all of them were discovered by accident — a researcher testing a compound for an unrelated property who happens to notice it loses electrical resistance at low temperature. Fewer than 20 have ever been predicted by theory before being confirmed in a lab. That ratio is the real story behind a July 2026 announcement from the SuperC consortium, an international research collaboration led by Aalto University's Päivi Törmä: two new superconductors, YRu3B2 and LuRu3B2, found not by accident but by a machine-learning search method built to make prediction the norm instead of the exception.
Why superconductor discovery has stayed slow
Superconductivity is a quantum phenomenon — below a critical temperature, a material's electrical resistance drops to exactly zero, and it can carry current indefinitely without losing energy to heat. That property is valuable enough that finding a material that superconducts near room temperature, rather than near absolute zero, is considered one of physics' most consequential open problems. It would mean lossless power transmission, dramatically more efficient data centers, and computing hardware that doesn't need cryogenic cooling.
The obstacle has never been a lack of ambition. It's the size of the search space. There are billions of plausible material combinations, and calculating whether any given one superconducts requires computationally expensive quantum mechanical simulations. Running that calculation on every candidate is intractable — which is why most discoveries happened by luck rather than by search.
How the ML pre-screening actually works
SuperC's approach splits the problem into two stages instead of one. First, a machine learning model — trained on the physics of known superconductors — scans a large space of candidate materials and flags the ones structurally likely to superconduct, without running the expensive full quantum simulation on each one. Only the small subset that clears this first filter goes on to detailed quantum calculations. This is the same pattern that has accelerated drug discovery and protein folding: use a cheap, fast model to cut a huge search space down to a manageable shortlist, then spend the expensive compute only where it's likely to pay off.
The two materials the method surfaced, YRu3B2 and LuRu3B2, share a specific structural feature: their superconductivity comes from electrons occupying what physicists call “flat bands” within a kagome lattice — a hexagonal geometric arrangement named for a traditional Japanese basket-weaving pattern. In a flat band, electrons can occupy energy states that barely change across the material, which concentrates their interactions and can enable exactly the kind of pairing behavior that produces superconductivity. Kagome-lattice materials have become one of the more promising structural families in superconductor research over the past several years, and having a search method that can systematically hunt through variations of that structure — rather than stumbling onto them individually — is the actual advance here.
What this changes, and what it doesn't
SuperC was established in 2023 with an explicit, ambitious target: find a room-temperature superconductor by 2033. Two new materials don't get the field there on their own — YRu3B2 and LuRu3B2 still require extremely low temperatures to superconduct, same as almost everything discovered before them. What changes is the discovery rate. If a validated ML pre-screening pipeline can realistically process candidates at a scale no human research group could evaluate by hand, the field moves from a small number of research teams making occasional lucky finds to a systematic search that can, in principle, work through a much larger fraction of the plausible material space.
That's a meaningful distinction for anyone tracking this space. A single new superconductor is a data point. A search method that turns superconductor discovery from serendipity into a repeatable pipeline is infrastructure — the kind of tooling change that compounds over years rather than producing one headline and going quiet. The next things to watch are whether SuperC or competing groups publish results at meaningfully larger scale (dozens or hundreds of new candidates rather than two), and whether the flat-band kagome family keeps producing hits or turns out to be a narrower vein than hoped.
The takeaway
For researchers and engineers working adjacent to materials science, the actionable signal here isn't “watch for room-temperature superconductors” — that's still likely years away even under an optimistic timeline. It's that ML-guided pre-screening is now validated enough to trust as a first-pass filter ahead of expensive quantum simulation, in this domain and probably in adjacent ones. Teams sitting on large unexplored candidate spaces in other areas of materials science should be asking whether the same two-stage pattern — cheap model, expensive simulation only on survivors — applies to their own search problem.