Porch Herald

World edition

Science

Euclid’s search for cosmic lenses still needs human eyes

Five AI models, nearly 7,000 volunteers and 67 specialists helped sift a vast sky survey into 15,072 high-confidence lens candidates.

Nine processed Euclid views show blue arcs and rings around brighter foreground galaxies, with multiple images of some distant objects.
Processed Euclid images of nine gravitational-lens systems, released Nov. 12, 2026. The collage shows separate views; foreground mass bends light from more distant objects. Original image title: “Euclid reveals the largest-ever catalogue of strong cosmic lenses.” Image: ESA/Euclid/Euclid Consortium/NASA, image processing by N. Lines and T. Li, CC BY-SA 3.0 IGO.

Euclid’s new catalogue of cosmic lenses began with 68 million possible targets. AI made that search manageable, but thousands of volunteers and a final review by astronomers helped determine which objects made the list.

The result, released with the space telescope’s first major data batch on Nov. 12, contains 15,072 high-confidence candidates. Each is a possible natural magnifying glass: foreground mass bends the light of something farther away into arcs, rings or repeated images.

Those distortions let researchers study the mass of foreground galaxies, including dark matter, and see distant objects that would otherwise be too faint. The catalogue’s scale gives astronomers many examples of configurations that were previously rare finds.

AI narrowed the search; people examined the shapes

The Euclid collaboration’s catalogue paper describes a search combining five machine-learning models with classifications from 6,975 volunteers and 67 strong-lensing specialists. The 68 million targets had first passed selection rules that excluded many stars and objects too faint for this search.

On the Space Warps platform, volunteers first judged whether an image looked like a lens. A second stage asked them to grade promising candidates. Some objects with especially strong machine or volunteer scores went directly to expert inspection.

The process did more than count votes. Test images, including simulated lenses, helped researchers assess volunteers’ classifications and combine their judgments. Expert grades were also calibrated to account for differences in how readily individual astronomers called something a lens.

The final limit was expert attention

The companion methods paper identifies a practical constraint: specialists could inspect only so many objects. Thresholds determined which candidates reached them. Loosening those thresholds would recover additional lenses while increasing the number of images needing expert review.

Volunteer screening let the team send a broader range of promising objects toward that final assessment. The authors also examined which known candidates the selection stages missed, rather than treating the finished catalogue as a complete census.

Of the 15,072 high-confidence candidates, 6,073 received the strongest grade and 8,999 were classed as probable lenses. The team separately lists 9,010 possible lenses. Those categories preserve the distinction between a convincing image and the further observations needed to establish a system’s nature.

The release supplies lens models and information about the foreground galaxies as well as the candidate list, giving other researchers a starting point for detailed studies of mass, distant galaxies and unusual lens systems.

Sources