Researchers have used machine-learning screening to identify and then experimentally confirm two new superconductors, YRu3B2 and LuRu3B2, both built on a kagome lattice structure. The compounds were flagged by an ML-guided workflow that combined computational predictions with first-principles calculations, then synthesized in the lab and tested with magnetization and specific-heat measurements. The result is a concrete demonstration that artificial intelligence can compress the years-long cycle of finding materials that conduct electricity with zero resistance, a property that could eventually reshape power grids and high-field magnets.
Why kagome superconductors found by machine learning matter right now
Superconductors lose no energy to electrical resistance, but discovering new ones has historically been slow and expensive. Researchers typically rely on chemical intuition and trial-and-error synthesis, a process that can consume years before a single promising compound reaches the lab bench. The two newly confirmed kagome-lattice compounds show that ML-accelerated screening can cut that timeline dramatically by narrowing thousands of candidate structures down to a short list worth synthesizing.
The broader question is whether this acceleration is repeatable and scalable. A closed-loop ML framework, in which experimental results feed back into the model to sharpen future predictions, has already produced at least one prior success: a zirconium-based alloy in the Zr–In–Ni system that was predicted, synthesized, and validated using the same iterative approach. That earlier result established the principle. The kagome pair extends it to a different crystal family and suggests the method is not locked to a single chemistry.
At the same time, the field is exploring complementary AI routes. Generative models that propose crystal structures, followed by density-functional-theory checks, have produced large libraries of hypothetical superconductors with calculated transition temperatures above target thresholds. The gap between computed candidates and experimentally confirmed superconductors remains wide, but the sheer volume of AI-generated proposals creates a pipeline that did not exist a decade ago. If closed-loop systems that return lab-measured critical temperatures back into their generative models can raise the fraction of candidates that survive first-round synthesis, the field will have moved from proof of concept to a reliable discovery engine.
How ML screening led to YRu3B2 and LuRu3B2
The peer-reviewed study in Physical Review Research details a workflow in which machine-learning models first screened a large space of ternary boride compounds, then ranked candidates by predicted likelihood of superconductivity. First-principles electronic-structure calculations refined the short list, weeding out unstable or electronically unfavorable structures. YRu3B2 and LuRu3B2 emerged as top targets and were subsequently synthesized as polycrystalline samples.
Experimental confirmation came through two standard probes. Magnetization measurements detected the diamagnetic signal characteristic of a superconducting transition, and specific-heat data showed the jump in electronic heat capacity that accompanies the onset of Cooper pairing. These are the same diagnostics used to validate any claimed superconductor, and the fact that both compounds passed both tests strengthens the case that the ML predictions were not false positives. Within the reported uncertainty, the critical temperatures and thermodynamic signatures are consistent with conventional, phonon-mediated superconductivity rather than an exotic pairing mechanism.
The kagome lattice itself is significant for physics beyond engineering applications. Kagome structures feature a network of corner-sharing triangles that can produce unusual electronic states, including flat bands and geometric frustration. These ingredients are associated with correlated phases such as charge-density waves, unconventional superconductivity, and quantum spin liquids. Finding superconductivity in this geometry gives theorists new data points for understanding how lattice topology interacts with electron pairing and whether flat bands can enhance the superconducting transition temperature.
A companion effort described in a methodology preprint laid the groundwork for the closed-loop concept. That earlier work showed how feeding experimental critical-temperature values back into the ML model improved prediction accuracy on the next iteration, effectively teaching the algorithm from its own successes and failures. The kagome discovery builds on that feedback architecture, though the two projects involve overlapping but distinct research teams and candidate chemistries. Together, they outline a general recipe: propose, compute, synthesize, measure, and retrain.
Open questions after two AI-predicted superconductors reach the lab
No independent laboratory has yet published replication measurements for YRu3B2 or LuRu3B2. The experimental data, including critical-temperature values and magnetization curves, come solely from the originating research group. Independent confirmation is a standard expectation in superconductor science, especially after high-profile replication failures in other materials. Until outside teams reproduce the results and verify sample quality, the findings carry a normal degree of provisional status.
The exact training data and model hyperparameters used in the ML screening step have not been released as open data. Methodological details appear in the associated workflow description, but without full transparency on the input dataset, other groups cannot directly benchmark their own models against the same baseline. That limits the ability to test whether the closed-loop feedback mechanism genuinely raises the survival rate of AI-proposed candidates or whether the improvement is specific to a particular dataset and model architecture. More open benchmarks and shared candidate lists would help separate generalizable advances from one-off successes.
The practical distance between these discoveries and loss-free power transmission is also large. Both compounds are conventional low-temperature superconductors, meaning they require cooling well below room temperature to function. The engineering goal of superconducting power lines, MRI magnets, or fusion devices is typically to operate as close to ambient conditions as possible, where cooling costs and infrastructure demands are lower. From that perspective, YRu3B2 and LuRu3B2 are better viewed as test cases for the discovery pipeline than as immediate technology platforms.
Yet even low-temperature materials can have impact. Superconductors used in particle accelerators, quantum-computing circuits, and sensitive magnetometers do not need to operate at room temperature; they need predictable properties, reproducible fabrication, and compatibility with existing cryogenic systems. If AI-guided searches can routinely deliver new compounds with slightly higher critical currents, better mechanical robustness, or improved tolerance to magnetic fields, they could still shift the performance frontier of these specialized applications.
Another open question is how broadly the current ML models can generalize across chemical space. The kagome borides sit in a relatively narrow compositional window with transition metals and light p-block elements. Extending the same workflow to cuprates, nickelates, or hydrides under extreme pressure may require different descriptors, training data, and physical constraints. There is also the risk that models trained on known superconductors will overfit to familiar motifs, missing genuinely novel mechanisms that do not resemble the existing catalog.
For now, the discovery of YRu3B2 and LuRu3B2 underscores a more modest but important point: machine learning is becoming an active participant in experimental condensed-matter physics rather than a purely theoretical tool. By closing the loop between prediction and measurement, researchers can iterate faster, explore riskier ideas, and quantify where their models succeed or fail. The next steps-independent replication, broader data sharing, and application to more diverse material families-will determine whether this approach remains a niche success or matures into a standard route for finding the superconductors of the future.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
