An AI searched nearly 100 million old Hubble images in two and a half days and found 1,300 cosmic oddities — more than 800 of them had never appeared in the scientific literature

A machine-learning system called AnomalyMatch processed 99.6 million small image cutouts derived from the Hubble Legacy Archive in roughly two and a half days. It did not declare discoveries on its own. It ranked the archive, after which astronomers examined the strongest candidates and assembled a catalogue of 1,339 unusual sources.

David O’Ryan and Pablo Gómez of the European Space Agency describe the search in a peer-reviewed paper in Astronomy & Astrophysics. Their coordinate and literature checks found 811 of the selected sources had no existing reference in the scientific literature.

This is one study, not settled consensus. “Anomaly” means that an object’s visible shape stood apart from the general population in this particular dataset. It does not mean the system found 1,339 phenomena outside known physics, nor that every proposed classification has been confirmed with follow-up observations.

The archive contained cutouts, not 100 million separate exposures

The scale is real, but the unit matters. Hubble has not taken 100 million separate photographs. The researchers used 99.6 million cutouts, each centred on a source detected within larger science-ready mosaics. Most were only a few dozen pixels across and covered about seven to eight arcseconds of sky.

The cutouts were produced from observations made with Hubble’s Advanced Camera for Surveys Wide Field Channel through the F814W filter, which records red and near-infrared light. The search did not combine every Hubble instrument, filter and exposure into a universal inventory. It examined a very large, consistently prepared subset suited to comparing visible structure.

That structure is what AnomalyMatch was designed to notice. A disturbed galaxy with long tidal features, a curved gravitationally lensed arc or the dark lane of an edge-on protoplanetary disc can look unlike the much larger population of ordinary sources. Brightness changes, spectra and other forms of astronomical difference were outside this morphology-based search.

A model produced the ranking and people made the catalogue

AnomalyMatch combines semi-supervised learning with active learning. It can begin with relatively few labelled examples, compare them with a much larger unlabelled collection, and then improve as a specialist reviews selected results and supplies additional labels.

The project began with only three examples of edge-on protoplanetary discs. During development, the system began assigning high scores to other unusual shapes, including galaxy mergers and gravitational lenses. The researchers expanded the training set as they worked until it contained 1,400 images, 375 labelled as anomalous and 1,025 as nominal.

Applied to the full dataset, the model assigned an anomaly score to every cutout. The team retained the 5,000 highest-scoring images for closer inspection. Many were duplicates or separate catalogue entries created from the same object, a known problem called source shredding. Cross-matching and de-duplication reduced the list to 1,339 unique sources.

The paper treated 1,176 as scientifically interesting anomalies after removing images judged nominal, spanning 19 working categories. The wider figure of more than 1,300 in NASA’s January 2026 account describes the full set of unique, odd-looking candidates reviewed by the researchers.

This distinction is not a reason to dismiss the result. It shows how the method actually works. The model concentrates promising material into a short list; expert review separates astronomical structure from ordinary sources, uncertain cases and imaging artefacts.

Most oddities were rare versions of known processes

Most of the high-ranked sources were galaxies merging or interacting with neighbours. Their mutual gravity had distorted discs, created several bright centres and pulled stars and gas into elongated streams. Such mergers are not unknown, but finding large samples with varied shapes is useful for studying how galaxies change during encounters.

Other candidates included gravitational lenses, in which the mass of a foreground object bends and magnifies light from a more distant source. The catalogue also contains jellyfish galaxies losing gas as they move through dense environments, galaxies with unusually large star-forming clumps, rings, arcs, jets and two already known edge-on protoplanetary discs.

Forty-three objects resisted the paper’s morphological categories. That does not establish 43 new kinds of object. Some may be unusual views of familiar systems, blended sources, observational artefacts or targets that need data at other wavelengths before they can be understood. O’Ryan and Gómez released them for other researchers to examine rather than assigning labels the images could not support.

What “never appeared in the literature” means

The researchers checked source coordinates against SIMBAD, ESASky and the publications and catalogues associated with them. On that basis, 811 of the 1,339 sources had no literature reference.

This does not necessarily mean no person had ever seen the pixels. Hubble is a targeted observatory rather than an all-sky survey telescope. Astronomers apply to point it at selected coordinates, so an unusual object may have been the intended target or may have appeared in the background. An image can sit in an archive without the object’s peculiar shape being classified or discussed in a paper.

Nor does absence from SIMBAD prove that no reference exists anywhere. Catalogue names, coordinate matching and coverage differ. “No literature reference found through this search” is the precise reading. The authors made the source identifiers, positions, images and provisional classifications available in machine-readable form so the checks can be repeated and amended.

Astronomers have used automated filters before

AnomalyMatch is a large-scale extension of an established idea, not the first attempt to let software find peculiar galaxies. In a 2021 Monthly Notices of the Royal Astronomical Society paper, computer scientist Lior Shamir applied an unsupervised method to 176,808 objects from several Hubble fields.

That algorithm reduced the collection to 1,100 high-ranked images, making manual review practical. Human inspection rejected about 86 per cent of those candidates and produced a catalogue of 147 outliers. The result captured both the value and the limit of automated anomaly detection: a useful system can miss objects and raise false positives while still cutting an impossible task to human size.

AnomalyMatch operated on a dataset more than 500 times larger. Its final candidate set still included roughly ten per cent nominal images, but the researchers did not need it to be an autonomous astronomer. They needed it to move rare-looking structures towards the front of the queue.

The next bottleneck is attention

Hubble’s archive spans observations made for thousands of separate programmes over more than three decades. Each programme was designed around particular questions. Its investigators could not exhaust everything visible in the foreground and background of every field, especially when the relevant pattern might become apparent only after comparison with millions of other sources.

Future survey archives will be larger still. Euclid, the Vera C. Rubin Observatory and the Nancy Grace Roman Space Telescope are designed to produce broad and repeated views of the sky. A method that can learn from a small labelled sample and present specialists with a ranked shortlist may be more useful there than one trained only to recognise a fixed set of familiar classes.

The Hubble result is therefore less about an AI independently understanding nearly 100 million pictures than about allocating scarce human attention. A constrained model searched quickly, astronomers checked what it returned, and hundreds of unusual sources that had escaped the literature became available for other teams to test.