A materials-screening study published in Scientific Reports predicts 22 plutonium-based compounds with superconducting transition temperatures above 50 kelvin at ambient pressure, including an estimated 89 K for Pu₂Co₁₇. It also proposes more than 40 hydrides with predicted values reaching about 100 K. The candidates come from existing crystallographic records; the reported transition temperatures are computational predictions awaiting experimental confirmation.[1]
What the new framework reports
| Stage | Reported result | Evidence status |
|---|---|---|
| Training data repair | 53,196 materials assembled as the DataG Non-Sc set | Dataset construction intended to supply negative examplesSource 1Source 2 |
| Model comparison | Ten conventional machine-learning and nine neural-network approaches evaluated | Computational benchmarkingSource 1 |
| Plutonium-rich screen | 22 candidates predicted above 50 K at ambient pressure; Pu₂Co₁₇ about 89 K | Model prediction from existing crystallographic entriesSource 1 |
| Hydride screen | More than 40 candidates predicted, with values up to about 100 K | Model prediction from existing crystallographic entriesSource 1 |
| Experimental confirmation | None reported for the proposed candidates | Still requiredSource 1 |
Why negative examples matter
Superconductor datasets are unusually vulnerable to selection bias. Published databases contain many materials studied because researchers suspected superconductivity, while compounds that show no transition may be less systematically recorded. A model trained mainly on positive or interesting examples can learn the shape of the historical search process rather than the boundary between superconductors and non-superconductors. The authors respond by compiling 53,196 materials treated as non-superconducting examples and using them alongside known superconductors.[1][2]
That move improves the learning problem, but the label itself deserves care. A material classified as non-superconducting may simply not have been measured under every relevant temperature, pressure, purity and structural condition. Negative data in materials science often mean 'no reported transition under the available observations', not proof that a transition is impossible. The new dataset is useful because it supplies contrast at scale; its assumptions still travel into the model.[1]
Extrapolation is harder than interpolation
The researchers tested ten machine-learning methods and nine neural-network designs, focusing on whether models could predict beyond the transition-temperature range represented in training. Functional convolutional neural networks and long short-term memory architectures showed extrapolative capability in their tests. This matters because a screen limited to interpolation could miss precisely the higher-temperature candidates researchers want to investigate.[1]
The framework screened entries from the Inorganic Crystal Structure Database, the Crystallography Open Database and the Materials Project. Its convergence on actinide-rich and hydrogen-rich families gives researchers specific structural starting points. Recovering scientifically plausible families supports the screening approach, although it does not validate any individual transition temperature.[1]
What would change the claim
The next tests are concrete: prepare selected candidates, verify their crystal phases and check whether electrical and magnetic measurements show a reproducible superconducting transition at the predicted temperature. Publishing unsuccessful tests would also strengthen the negative evidence the framework was designed to supply.[1]
Sources & further reading
- Genome-inspired AI for extrapolative discovery of high-temperature superconductors Scientific Reports · 29 Oct 2026 · Accessed 30 Oct 2026
- AI-Driven Discovery of High-Temperature Superconductors via Materials Genome Initiative and High-Throughput Screening arXiv · 5 Nov 2025 · Accessed 30 Oct 2026