Executive Summary

Since superconductivity was discovered in 1911, more than 7,000 superconductors have been confirmed by experiment. Yet fewer than 20 of them were predicted theoretically before the material was ever made. The rest were mostly products of trial and error and luck. That is why, more than a century after superconductivity appeared, discovery remained a matter of searching blindly.

The SuperC consortium, which includes Aalto University and Rice University, reversed that order. Machine learning first narrows the promising combinations of elements, calculations then filter them, and only afterward are the materials actually synthesized and confirmed in the lab. The materials predicted and verified this way are the kagome-structured YRu₃B₂ and LuRu₃B₂. Their measured critical temperatures, however, came in at 0.81 K and 0.95 K — far below the theoretical predictions of 3.37 K and 1.88 K.

This article looks less at the two materials than at how they were found. The result is not two superconductors but a pipeline that turned discovery from serendipity into search.

Under 20

Superconductors predicted before synthesis

The other ~7,000 came from trial and error and chance

0.81·0.95 K

Measured critical temperatures

Cryogenic, far from practical — the result is the method, not the temperature

Billions

Candidates that can be pre-screened

The scale of chemical space machine learning makes tractable

2033

SuperC's room-temperature target

This result validates the method on that roadmap

1

115 Years, 7,000 Materials, Fewer Than 20

Superconductivity is the phenomenon in which a material's electrical resistance vanishes completely below a certain temperature. Since it was first observed in 1911, more than 7,000 superconductors have been confirmed experimentally. But the way that list was built was mostly a slow, luck-dependent process: pick a candidate by hand, synthesize it one at a time, and measure it at low temperature.

Of those 7,000, fewer than 20 were predicted theoretically before the material was made. The reason is that the mechanism of superconductivity is itself hard to predict. In systems where electrons interact strongly with one another, calculating from first principles when and under what conditions superconductivity will appear has never been easy, and the experimental variables are complex too. So discovery has looked less like prediction and more like excavation.

The trouble is that the combinations of materials are effectively infinite. Depending on which elements you place in what proportions and in what structure, the number of candidates grows astronomically. Trying the most promising-looking ones one by one, guided by intuition and experience, only ever covers a tiny fraction of that vast space.

2

Predict First, Synthesize Second

The SuperC consortium, led by Professor Päivi Törmä of Aalto University, is an international collaboration launched in 2023 that includes Rice, Princeton, Ruhr University Bochum, and the Donostia International Physics Center. This result was published in the journal Physical Review Research in June 2026. The heart of their method is the ordering: where the field had long built a material first and then measured whether it superconducted, they let computation nominate the candidates first and synthesized only those.

The pipeline runs in four stages.

  • Machine-learning pre-screening — a machine-learning model sweeps a vast set of element combinations and narrows them to candidates with a high likelihood of superconductivity. Instead of running expensive precise calculations on every candidate, the key is to place a cheap predictor upstream to compress the space.
  • First-principles calculation — the narrowed candidates undergo precise quantum-mechanical calculations to confirm theoretically whether superconductivity will actually appear and what the critical temperature will be.
  • Synthesis — materials that pass the calculations are actually made by the group of Professor Emilia Morosan at Rice University.
  • Experimental verification — the magnetic susceptibility, specific heat, and electrical resistance of the samples are measured to confirm that superconductivity really occurs.
1 2 3 4 ML Pre-Screening First-Principles Calculation Synthesis Experimental Verification
▲ SuperC's four-stage pipeline — machine-learning pre-screening compresses the candidate pool before calculation, synthesis, and experimental verification. Original Pebblous diagram (reinterpretation)

Törmä described the approach as one that "pre-screens with machine learning and then runs targeted calculations only on promising candidates," and said it will greatly accelerate the discovery of superconductors going forward. She added that this approach could raise the number of materials that can be handled into the billions — a search on a scale orders of magnitude beyond what people used to survey by hand.

3

Two Superconductors from a Kagome Lattice

The materials predicted and verified through this pipeline are YRu₃B₂ and LuRu₃B₂. Both have a hexagonal CeCo₃B₂-type structure in which ruthenium (Ru) atoms form a kagome lattice on a plane. Kagome — named after the weave pattern of traditional Japanese baskets — is a distinctive arrangement of interlocking triangles and hexagons. In this lattice, electrons form a low-mobility flat band, and that condition induced superconductivity.

▲ Kagome lattice — ruthenium (Ru) atoms (orange dots) form a planar arrangement of interlocking triangles and hexagons, and this structure gives electrons a flat band. Original Pebblous diagram (arXiv:2512.16945 reinterpretation)

The critical temperatures confirmed in experiment were 0.81 K for YRu₃B₂ and 0.95 K for LuRu₃B₂. In both materials the superconducting volume fraction reached nearly 100 percent, confirming bulk superconductivity — the entire sample enters the superconducting state, not just the surface or a portion of it.

3.1Where Theory and Measurement Diverged

The predictions were not exactly right. The theoretically calculated critical temperatures were 3.37 K for YRu₃B₂ and 1.88 K for LuRu₃B₂ — well above the measured values. The team explained the gap as a metastable phase transition arising from excessive phonon softening. In plain terms, the model captured the material's atomic-vibration characteristics more simply than reality, and as a result overestimated the critical temperature.

Theoretical Prediction vs Measured Critical Temperature (Tc)

YRu₃B₂ Predicted 3.37 K · Measured 0.81 K
LuRu₃B₂ Predicted 1.88 K · Measured 0.95 K

Bar length = measured value as a share of the theoretical prediction · Orange = measured, light gray = theoretical prediction (overestimated due to phonon softening). Original Pebblous diagram

What matters is that this discrepancy is not a failure of the pipeline. The model got the presence of superconductivity right and missed on temperature, a quantitative value. It was even explained physically where and why it was off. Even when the prediction is not perfect, the loop of predicting, making, and confirming itself worked.

4

The Result Is the Method, Not the Material

A critical temperature below 1 K is far from practical. A superconductor that works only at cryogenic temperatures near absolute zero has nothing yet to do with the long-standing goal of carrying electricity without resistance at room temperature. And yet this work is still called a result — because of the order, not the temperature.

YRu₃B₂ and LuRu₃B₂ were predicted before they were synthesized, and confirmed to be superconductors as predicted (with some error). If most of the discoveries of the past 115 years were chance and excavation, this time calculation pointed to the candidates first, and experiment followed to confirm them. The mode of discovery shifted from serendipity to search.

The SuperC consortium has set a goal of finding a superconductor that works at room temperature by 2033. These two materials are not the achievement of that goal but the first proof that the method for reaching it actually works. Holding a search tool that can sweep billions of candidates is a result that will outlast the two cryogenic superconductors themselves.

5

The Pebblous View: Data Sets the Resolution of the Search

Look at the predict-first, synthesize-later structure from a data perspective and the real bottleneck is not the accuracy of the screening model but what sits in front of it. Whether you can narrow chemical space down to billions of candidates depends on which features you use to represent a material and how coherent the experimental data you trained that representation on is. If the representation is coarse, even a good model searches in the wrong places; if the training data is shaky, the prediction cannot get even a single temperature right.

The point where theory and measurement diverged in this study was, in the end, a representation problem. The model captured phonon characteristics more simply than reality, and so overestimated the temperature. Which physical quantities you capture in the data, and at what precision, becomes the resolution of the search. Search speed is decided first by the representation and quality of the data, not by the amount of computation.

Predict-first pipelines are not a superconductor story alone. In R&D domains where experiments are costly and the candidate space is vast — drug discovery, battery materials, catalyst design — the same pattern is already taking hold. It is a shift away from treating experiments as a byproduct of data, toward having data and models point to the next experiment.

The force that turns discovery into search is not faster computation but a data structure that represents candidates well and feeds experimental results back coherently. The autonomy of R&D arrives first in the organizations that have built that representation and that quality.

R

References

Primary Source & Academic Paper

Industry & News Coverage