AI Identifies Two New Superconductors, Paving Way to Screening Millions of Materials

Researchers from Finland's Aalto University and Rice University in the United States have announced the discovery of two new superconductors, YRu3B2 and LuRu3B2, first flagged by a machine learning algorithm and only later confirmed in the lab. It is the first case in which a method combining artificial intelligence with quantum calculations has led to the real, experimentally verified synthesis of a new superconducting material.

The method works in two stages. First, a machine learning model scans a vast number of theoretically possible chemical compounds and flags those with the highest chance of exhibiting superconductivity. Only the selected candidates then go through far more time-consuming first-principles quantum calculations, and finally chemical synthesis and lab testing.

How the machine learning filter worked

Classic searches for new superconductors involve laboriously calculating the properties of one compound after another using methods that are accurate but extremely costly in computing power. The Aalto team reversed the order, having a machine learning model first roughly evaluate hundreds of thousands of candidates, then directing limited computing resources only toward those that actually showed promise.

Our method uses preliminary screening through machine learning, followed by targeted calculations for promising candidates - Päivi Törmä, professor at Aalto University, leader of the SuperC consortium

The compounds YRu3B2 and LuRu3B2 identified by the algorithm did not previously exist as characterized materials. Only after the model flagged them did a team led by professor Emilia Morosan of Rice University chemically synthesize samples from their constituent elements and experimentally confirm that both compounds do indeed become superconducting.

A kagome lattice as the common thread

Both new materials share a crystal structure known as a kagome lattice, in which atoms form a repeating pattern of triangles and hexagons resembling traditional Japanese basket weaving. This geometry causes electrons moving through the material to experience a phenomenon called kinetic frustration, which has made kagome compounds one of the hottest topics in condensed matter physics for several years.

Previously studied kagome superconductors, such as CsV3Sb5, became superconducting at higher temperatures, reaching around 2.5 kelvin. The newly discovered YRu3B2 and LuRu3B2 superconduct at lower temperatures, below 1 kelvin, but their significance lies not in a record critical temperature but in the fact that they were found through a predictive model rather than chance or manual search.

The room-temperature superconductor goal

The discovery fits into the goals of the SuperC consortium, launched in 2023 as the first coordinated global scientific collaboration aimed at finding a room-temperature superconductor by 2033. Superconductors today require cooling to temperatures close to absolute zero or, in the case of high-temperature superconductors, to temperatures still far from room conditions, which limits their use to specialized installations such as particle accelerators or MRI scanners.

If a material capable of conducting electricity without resistance at room temperature and normal pressure were found, it would open the door to power grids without transmission losses, far more efficient quantum computers and cheaper magnetic levitation. The problem is that the number of theoretically possible chemical compounds to test runs into the billions, and classical computational methods simply cannot process that many candidates in a reasonable time.

Implications for further research

What matters most about this discovery is that the method was verified end to end for the first time, from the algorithm's initial pick through theoretical calculations to physical synthesis and laboratory measurement. The researchers stress that the algorithm itself did not replace physicists or turn labs into automatic discovery factories, but it did point to a path that makes the search for new materials less like a lottery.

Commentators on the study also point to the risk: poorly used machine learning models can produce false leads just as quickly as accurate ones, and materials science is a rigorous enough field that wrong predictions often only surface at the stage of costly laboratory synthesis. That is why the Aalto and Rice teams treat machine learning as a filter that speeds up the work, not as an oracle that replaces classical experimental verification.

For industry and investors following the race for superconductors, the result matters mainly as proof that the method works, not as a ready commercial product. The SuperC consortium's next step is to scale the approach to a much larger number of candidates, with the ambition of processing hundreds of thousands, and eventually billions, of potential compounds in the search for a material that would change the rules of the game in energy and electronics.