These six galaxies were among the almost 1,400 anomalous objects buried in the Hubble Legacy Archive. The discovered objects include a ring-shaped galaxy, a bipolar galaxy, a group of merging galaxies, and three galaxies with warped arcs created by gravitational lensing. Credit: ESA/NASA

For more than three decades, the Hubble Space Telescope has collected targeted images to answer specific scientific questions, from mapping galaxies to studying nearby nebulae. Hubble has gathered so much data that despite their best efforts, astronomers haven’t had the time to analyze it all in detail yet.

Alongside its many targets, the Hubble archive also contains many unexpected objects that were never the focus of the original observations.

Now, two astronomers have revisited that massive archive with a new plan. They deployed an artificial intelligence system designed to notice when something looks “wrong”. In just 60 hours of computing time, the tool flagged over 1,300 anomalies hidden within 100 million Hubble snapshots. Hundreds of them have never appeared in scientific literature.

A Sea of Cosmic Thumbnails Close-up of one of the gravitational lens anomalies captured. Credit: ESA/NASA

The new study, published in Astronomy & Astrophysics, focused on the Hubble Legacy Archive and a dataset built from it: 99.6 million “cutouts,” each a tiny square view of sky centered on an extended source (often a galaxy) drawn from the Hubble Source Catalogue.

The cutouts are small but consistent: 150 by 150 pixels, about 7.5 arcseconds on a side, created from Hubble’s Advanced Camera for Surveys. The images were saved as single-channel grayscale JPEGs—good enough to preserve shapes while keeping storage manageable.

David O’Ryan and Pablo Gómez built a system called AnomalyMatch to wade into this sea of thumbnails. Their initial goal was modest: find “protoplanetary disks,” the rare, edge-on dusty pancakes where planets are born. They trained the AI using only three known examples.

However, as the system learned what “unusual” looked like, it went rogue—in a good way. It started surfacing everything from warped gravitational lenses to galaxies crashing into one another. Running on ESA’s Datalabs platform, the AI did in two and a half days what would have taken a human lifetime.

Five examples of every anomaly sub-class for which the team found at least five objects (except quasars). Credit: ESA/NASA

After it finished its ranked search, AnomalyMatch returned a ranked list. Then, the researchers reviewed the 5,000 most unusual images by eye. After removing duplicates caused by catalog errors (single galaxies were mistakenly split into multiple entries) they identified 1,339 distinct anomalies.

About half were merging or interacting galaxies, warped by gravity into lopsided smears and trailing tidal tails. Many others looked like gravitational lenses—chance alignments where a foreground galaxy bends the light of a more distant one into arcs, partial rings, or bright mirrored knots.

A Gallery of the Bizarre Close-up of a galaxy-merging anomaly. Credit: ESA/NASA

Telescopes have become so productive that discovery depends mostly on triage—deciding what is worth a closer look. AnomalyMatch is one possible response to that problem.

Hubble alone has amassed an archive spanning decades, but it is not a survey telescope. Researchers look for specific things; outside the primary objective, data is often not given sufficient attention. The most shocking find of this study is that about 65% of their final anomaly sample (811 objects) were completely new in the scientific literature.

Some of those “new” objects may be new only in the sense that no one wrote them up; they could be background galaxies caught incidentally, or features that were too easy to miss amid a study’s main target.

The paper’s breakdown reads like a catalog of astrophysical edge cases: 629 merging systems, 140 candidate gravitational lenses, 35 jellyfish galaxies, and 43 objects whose forms “defy classification.” Those unclassifiable sources are just unfamiliar enough that the authors chose not to force labels onto them.

These extremes clarify mechanisms. A strong gravitational lens can reveal how mass—much of it dark matter—sits inside a galaxy. A jellyfish galaxy can show how a galaxy loses gas as it plows through the thin plasma of a cluster.

Close-up of a difficult-to-categorize object. It represents a bi-polar galaxy with a compact swirling core and an open lobe at each side. Credit: ESA/NASA

The study also points to a shift in how astronomy will operate as new observatories come online. ESA’s Euclid mission is already mapping the large-scale universe, and NASA’s Nancy Grace Roman Space Telescope and the Vera C. Rubin Observatory will generate torrents of images with far less human selectivity than Hubble’s targeted approach.

In that future, the role of the astronomer may look more like an editor—reviewing algorithmically surfaced candidates, choosing what deserves follow-up, and deciding which oddities are merely unusual and which are the beginnings of something new.