For nearly a century, the search for new superconductors has been a story of luck as much as science. Researchers stumble onto a promising material, cool it to within a whisper of absolute zero, and hope its electrical resistance vanishes. Most of the time, nothing happens. This week, an international team led by Aalto University showed a different way forward: they used machine learning to sift through a practically infinite space of possible materials and pulled out two genuine, previously unknown superconductors. The result is small in raw numbers but large in what it implies about how physics might be done from here on.
The two new materials, YRu3B2 and LuRu3B2, are not headline-grabbing on their own terms. They become superconducting only at 0.81 and 0.95 kelvin, fractions of a degree above absolute zero and far from anything useful in daily life. What makes them matter is the path that led to them. Instead of testing candidates one by one, the team let an algorithm do the first pass, narrowing an unmanageable field to a handful of plausible bets that were then confirmed by detailed quantum calculations and, finally, synthesized and measured in a lab. The proof-of-concept paper appeared in Physical Review Research.
It is a modest headline with an immodest subtext. If a screening pipeline can reliably surface real superconductors from theory, then the bottleneck that has defined this field for decades may finally be loosening.
Why a fraction of a degree matters
Superconductors carry electric current with zero resistance, a quantum effect that appears only at extremely low temperatures. That single property underwrites an outsized share of advanced technology. Superconducting magnets steer particles in accelerators, generate the fields inside MRI and neuroimaging machines, confine plasma in experimental fusion reactors, and levitate maglev trains above their tracks. Every one of those systems pays a steep price for the privilege: elaborate cooling equipment to hold the material near absolute zero.
That is why the field is gripped by a single, almost mythical goal. A superconductor that worked at room temperature would remove the cooling penalty entirely and, in doing so, rewrite the economics of energy. "Superconductive materials that can operate at room temperature would forever change the way we consume energy," explains Aalto Professor Päivi Törmä, who leads the consortium behind the work. If such a material replaced ordinary conductors in computers and data centres, she notes, global energy consumption could fall sharply and the heat footprint of the information-technology sector could shrink with it.
The difficulty is that these materials are maddeningly rare. Any combination of elements might, in principle, superconduct. Almost none actually do, and the ones that work still demand deep cooling. Finding a scalable, room-temperature candidate has been less an engineering problem than a search through a haystack the size of the periodic table raised to a very large power.
How the search actually works
The team's method rests on a marriage of two ideas: quantum geometry and machine learning. Both of the new materials draw their superconductivity from electrons arranged in what physicists call flat bands, formed within a kagome lattice. The name comes from a traditional Japanese basket-weaving pattern of interlaced triangles, and the geometry is not decorative. That particular arrangement encourages electrons to behave collectively in ways that favour superconductivity, giving the researchers a physical principle to search for rather than a blind guess.
Machine learning supplies the speed. "Our method uses machine-learning-based pre-screening followed by targeted calculations on the promising candidates," Törmä says. In practice, the model rapidly filters enormous numbers of elemental combinations down to a shortlist. Only then does the team spend expensive computing time on rigorous quantum-mechanical calculations, and only the survivors of that stage move on to synthesis. It is a triage system: cheap, fast judgment first, costly certainty later.
The scale of the problem it addresses is striking. Over the decades, researchers have recognised more than 7,000 superconductors, Törmä notes, but mostly by accident. The calculations required to predict a material's viability from first principles are so heavy that scientists have only ever theoretically predicted about twenty of them. A screening layer that trims the field before the hard math begins does not replace physics; it decides where the physics is worth doing.
A collaboration of people and machines
It would be a mistake to read this as an algorithm working alone. The discovery was a relay. The machine-learning stage narrowed the candidates; quantum-geometry calculations confirmed which ones were theoretically sound; and then human hands finished the job. After the theoretical work, collaborators at Rice University, led by Professor Emilia Morosan, synthesised the actual samples, chemically combining raw elements into new compounds and running the physical tests that confirmed superconductivity. No prediction counts until a material exists on a bench and behaves as promised.
That division of labour is quietly the point. The interesting frontier in scientific machine learning is not automation for its own sake but the reallocation of human attention. Screening models are best understood as instruments that tell researchers where to look, in the same way a telescope does not make discoveries but decides which patch of sky deserves a closer stare. The judgment, the synthesis, and the interpretation remain stubbornly human.
What comes next
The work is part of the SuperC consortium, a coordinated international collaboration that Törmä and a group of physicists formed in 2023 with an unusually concrete ambition: to find a room-temperature superconductor by 2033. This latest result is best read as a proof of concept for the method rather than a destination in itself. The two materials are the demonstration; the pipeline is the product.
What the team hopes to scale is throughput. "With machine learning, we may be able to push the number of materials we can process into the billions," Törmä says, framing that reach as a critical step toward the room-temperature goal. Whether the 2033 target proves prescient or optimistic, the shift in approach is real. A search that once depended on serendipity now has a systematic front end, and a field that could theoretically vet only a few dozen candidates may soon be able to reason about vastly more. The consortium's work will feature in Aalto University's Designs for a Cooler Planet exhibition this autumn, a fitting venue for research explicitly aimed at the climate stakes of energy efficiency.
Closing thoughts
There is a temptation to measure this discovery by its temperatures and find it wanting. Under one kelvin, two obscure compounds, no immediate application. But that framing misses what changed. The scarce resource in materials science has never been raw computing power alone; it has been knowing where to point it. By letting a model handle the first, wasteful pass of elimination, the researchers converted an intractable search into a directed one.
If the method holds, the more interesting superconductors are the ones nobody has found yet, hidden in combinations no human would have thought to test. The prize at the end of that search, a material that carries current without loss at ordinary temperatures, would touch nearly everything that runs on electricity, from the power grid to the data centres now straining under the weight of computation itself. For now, the achievement is narrower and more honest: a demonstration that the haystack can be searched, and that machines and physicists together can find the needle faster than either could alone.
한글 요약
알토대학교가 이끄는 국제 연구진이 머신러닝을 활용해 두 개의 새로운 초전도체(YRu3B2, LuRu3B2)를 발견했습니다. 두 물질은 각각 절대영도에 가까운 0.81K, 0.95K에서만 초전도성을 보여 실용성 자체는 낮지만, 발견에 이르는 '방법'이 핵심입니다. 연구팀은 사실상 무한한 물질 조합을 머신러닝으로 먼저 걸러낸 뒤, 유망한 후보에 대해서만 정밀한 양자역학 계산을 수행하고, 최종적으로 라이스대학교(에밀리아 모로산 교수 연구실)에서 실제 시료를 합성해 초전도성을 확인했습니다. 두 물질 모두 일본 전통 바구니 문양인 '카고메 격자' 구조에서 전자가 형성하는 '평평한 띠(flat band)'로부터 초전도성을 얻습니다.
이 성과가 중요한 이유는 초전도체 탐색의 오랜 병목을 겨냥하기 때문입니다. 지금까지 알려진 초전도체는 7,000종이 넘지만 대부분 우연히 발견되었고, 이론적으로 예측된 것은 약 20종에 불과합니다. 물질 하나의 초전도 가능성을 계산하는 데 막대한 연산이 들기 때문입니다. 머신러닝 기반 사전 선별은 값비싼 계산을 시작하기 전에 후보군을 좁혀, '어디에 물리학을 집중할지'를 결정해 줍니다. 알토대의 파이비 퇴르마 교수는 이 방식으로 처리 가능한 물질 수를 향후 수십억 개 규모까지 끌어올릴 수 있다고 전망했습니다.
이번 연구는 상온 초전도체를 2033년까지 찾겠다는 목표로 2023년 출범한 국제 컨소시엄 SuperC의 일환이며, 개념 증명 논문은 학술지 Physical Review Research에 게재되었습니다. 상온에서 저항 없이 전류를 흘리는 물질이 실현된다면 컴퓨터와 데이터센터의 전력 소비를 크게 줄여 정보기술 분야의 에너지·발열 부담을 낮출 수 있습니다. 이번 발견은 그 자체로 혁명은 아니지만, 우연에 의존하던 탐색을 체계적인 방향 탐색으로 바꿔 놓았다는 점에서 의미가 있습니다.
참고: Aalto University 보도자료, Physical Review Research(lpqj-7hyg), SuperC consortium.