A machine learning system called AnomalyMatch processed 99.6 million small image clippings from the Hubble Legacy Archive in about two and a half days. This is not an independent declaration of discovery. It ranked the archives, then astronomers considered the most likely candidates, creating a catalog of 1,339 unusual sources.
David O’Ryan and Pablo Gomes of the European Space Agency describe the search in a peer-reviewed paper. astronomy and astrophysics. Their coordinate and literature checks revealed that 811 of the selected sources had no existing references in the scientific literature.
This is one study, not a collective consensus. “Anomalous” means that the visible shape of the objects in this particular dataset is different from the general population. This does not mean that the system discovered 1,339 phenomena outside of known physics, nor does it mean that all proposed classifications were confirmed by follow-up observations.
The archive contained cutouts rather than 100 million individual exposures
The scale is real, but the units are important. Hubble doesn’t take 100 million individual photos. The researchers used 99.6 million clippings, each centered around a detected source within a larger mosaic prepared for science. Most were only a few tens of pixels wide and covered about 7 to 8 arc seconds of the sky.
This clipping was created from observations made with Hubble’s Wide Field Survey Altitude Camera through the F814W filter, which records red light and near-infrared light. This search did not integrate all Hubble instruments, filters, and exposures to create a universal inventory. It examined a very large and consistently prepared subset suitable for comparing visible structures.
AnomalyMatch is designed to focus on its structure. Disturbed galaxies with long tidal features, the curved arcs of gravitational lenses, or the dark alleys of edge-on protoplanetary disks can look different from the much larger population of regular sources. Changes in brightness, spectra, and other astronomical differences were outside the scope of this morphology-based search.
Models create rankings, people create catalogs
AnomalyMatch combines semi-supervised and active learning. You can start with a relatively small number of labeled samples, compare them to a much larger unlabeled collection, and then review the expert-selected results and improve them by providing additional labels.
The project began with just three examples of edge-on protoplanetary disks. During development, the system began assigning higher scores to other unusual shapes, such as galaxy mergers and gravitational lenses. During their work, the researchers expanded the training set to include 1,400 images, of which 375 were labeled abnormal and 1,025 were labeled normal.
The model was applied to the entire dataset and assigned an anomaly score to every cutout. The team saved the 5,000 images with the highest scores for closer inspection. Many are duplicate or separate catalog entries created from the same object, a known issue called source shredding. Cross-matching and deduplication reduced the list to 1,339 unique sources.
In this paper, after removing images deemed nominal across 19 task categories, we treated 1,176 images as anomalies of scientific interest. The wide-ranging number of more than 1,300 people in NASA’s January 2026 report represents the complete set of unique and odd-looking candidates that researchers considered.
This difference is not a reason to reject the results. Here’s how this method works in practice. This model concentrates promising materials into a short list. Expert reviews separate astronomical structures from common sources, uncertain cases, and image artifacts.
Most of the strange ones were unusual versions of known processes
Most of the top-ranked sources were galaxies merging or interacting with neighboring galaxies. Their mutual gravity warped the disk, forming several bright centers that drew stars and gas into elongated streams. Such mergers are not unknown, but finding large samples of different shapes can help study how galaxies change during encounters.
Other candidates included gravitational lenses, where a mass of objects in the foreground bends and magnifies light from more distant sources. The catalog also includes jellyfish galaxies that lose gas as they move through dense environments, galaxies with unusually large star-forming clumps, rings, arcs, jets, and two already known edge-on protoplanetary disks.
Forty-three objects resisted the paper’s morphological categories. It does not establish 43 new objects. Some may be familiar systems, mixed sources, observational artifacts, or unusual views of targets that require data at other wavelengths before they can be understood. Rather than assigning labels that the images could not support, O’Ryan and Gómez published them for other researchers to examine.
What does “not appear in literature” mean?
The researchers checked the source coordinates against SIMBAD, EASky, and their associated publications and catalogs. Based on that, 811 of the 1,339 sources had no literature references.
This doesn’t necessarily mean that no one has ever seen that pixel. Hubble is a targeted observatory, not an all-sky surveying telescope. Astronomers have applied to point it at selected coordinates, so an unusual object could be the intended target or appear in the background. Images may be stored in archives without the unique shape of the object being classified or discussed in a paper.
Also, its absence in SIMBAD does not prove that the reference does not exist anywhere. The catalog name, coordination matching, and target range are different. “No references were found for this search” is the correct interpretation. The authors made source identifiers, locations, images, and preliminary classifications available in machine-readable format for repeated checks and corrections.
Astronomers have used automatic filters before
AnomalyMatch is a major extension of an established idea, and it’s not the first attempt to force software to discover unusual galaxies. in 2021 Royal Astronomical Society Monthly Notices According to the paper, computer scientist Liol Shamir applied an unsupervised method to 176,808 objects from several Hubble fields.
This algorithm reduced the collection to 1,100 top images, making manual review practical. Human inspection rejected approximately 86 percent of these candidates, creating a catalog of 147 outliers. This result captured both the value and limitations of automated anomaly detection. Useful systems can perform impossible tasks at human scale while still missing objects or producing false positives.
AnomalyMatch worked on over 500 times more datasets. Although that final candidate set still contained about 10 percent of the nominal images, the researchers did not require it to be an autonomous astronomer. It was necessary to move unusual-looking structures to the front of the column.
The next bottleneck is attention.
Hubble’s archive contains observations made on thousands of individual programs over more than 30 years. Each program was designed around specific questions. The researcher could not cover everything that appears in the foreground and background of any field, especially when relevant patterns may only become apparent after comparison with millions of other sources of information.
Future research archives will be even larger. The Euclid, Vera C. Rubin Observatory, and Nancy Grace Roman Space Telescope are designed to produce repeated views of large areas of the sky. There, a method that can learn from small labeled examples and present a ranked list of candidates to an expert may be more useful than a method trained only to recognize a fixed set of familiar classes.
So Hubble’s results are less about AI making sense of nearly 100 million photos on its own and more about allocating scarce human attention. The constrained model was quickly searched, astronomers checked what it returned, and hundreds of unusual sources that had been missing from the literature became available for other teams to test.
