Astronomical archives contain more observations than any human team can fully review. The Hubble Legacy Archive used by the AnomalyMatch researchers held 99.6 million source cutouts. The NEOWISE Single-Exposure Database held roughly 200 billion infrared apparitions collected over 10.5 years. The data was already there; the challenge was searching it at scale.
Two recent projects show how machine learning can systematically scan these archives and surface objects that earlier searches missed. One identified 811 previously undocumented anomalies in Hubble data. The other produced a catalog of 1.9 million variable-object candidates from NEOWISE data, including roughly 1.5 million that did not appear in existing catalogs. Neither project discovered new physics. Both extracted more information from observations astronomers had already collected.
AnomalyMatch: 100 Million Hubble Cutouts in Two and a Half Days
David O’Ryan and Pablo Gómez, both at the European Space Agency, built a neural network called AnomalyMatch and pointed it at the Hubble Legacy Archive. The tool processed 99.6 million 150-by-150-pixel cutouts, each 7.5 arcseconds square, in roughly two and a half days.
The system identified 1,339 objects with unusual appearances. Of those, 811 had not previously been documented in scientific literature. The results were published in Astronomy & Astrophysics in December 2025.
AnomalyMatch learned visual patterns that distinguish unusual sources from the archive’s more common galaxies, stars, and other objects. The researchers then visually inspected and classified the highest-ranked candidates, rejecting some as ordinary sources or imaging artifacts.
The inspected objects included:
- 629 merging or interacting galaxies, including 417 not previously reported
- 140 gravitational lenses, including 138 candidates, with 86 of those candidates newly reported
- 35 jellyfish galaxies, including 18 new examples
- 12 ring galaxies, a rare type formed when one galaxy passes through the center of another
- 11 galaxies with massive star-forming clumps
- Two edge-on protoplanetary disks seen in profile
- 43 objects placed in an unknown category because they did not fit the paper’s existing classifications
That last category is worth paying attention to. Follow-up observations will be needed to understand what those sources are.
“This is a powerful demonstration of how AI can enhance the scientific return of archival datasets,” said Pablo Gómez. O’Ryan’s main research focus is the relationship between galaxy evolution and galaxy morphology, and the candidate list gives researchers hundreds of additional unusual systems to examine.
The gravitational-lens candidates are particularly useful. Gravitational lenses act as natural magnifying glasses, allowing astronomers to study distant objects that would otherwise be too faint to observe. Finding more candidates in archival data can support new analyses without requiring new observations.
VARnet: A High School Student Finds 1.5 Million Candidates
The second project came from Pasadena High School student Matteo Paz, who built an AI model called VARnet while participating in Caltech’s Summer Research Connection program in 2023. Working with Caltech senior scientist Davy Kirkpatrick, Paz designed the system to identify variable astronomical objects - sources that change brightness over time - in NASA’s NEOWISE infrared data.
VARnet works in three stages. First, wavelet decomposition reduces measurement noise in the raw data. A Finite-Embedding Fourier Transform, or FEFT, then extracts periodic patterns from irregularly sampled light curves. Finally, convolutional neural networks sort each source into four broad categories: non-variable, transient, intrinsic pulsator, or eclipsing binary.
The system processed each source in under 53 microseconds on a GPU with roughly 22 GB of memory and achieved an F1 score of 0.91 on validation data.
Applied to the roughly 200-billion-apparition NEOWISE Single-Exposure Database, VARnet produced a catalog of 1.9 million infrared variable-object candidates. Roughly 1.5 million had no prior record in existing catalogs. The candidates include potential quasars, young stellar objects, variable stars, and transient events such as supernovae.
Paz wrote the work as a single-author paper that was accepted by The Astronomical Journal, and the project earned first place at the 2025 Regeneron Science Talent Search.
Caltech reported that Paz planned to release the full VarWISE catalog with updated classifications and cross-matches against other surveys.
What This Pattern Means
These two projects share a common structure: take an existing archive that is too large for manual review, build a system to scan it, and rank objects that earlier searches missed. Neither AnomalyMatch nor VARnet required new observations. The candidates were already present in the data.
The AnomalyMatch paper discusses Euclid, the Vera C. Rubin Observatory, and the Nancy Grace Roman Space Telescope as possible future applications. Its authors argue that this kind of anomaly search could help researchers examine larger surveys, but its ability to scale to those datasets still has to be demonstrated.
Both projects also depend on human verification. AnomalyMatch researchers visually inspected the candidates and rejected some as ordinary sources or artifacts. VARnet produced 1.9 million candidates that require further work to confirm and classify. The systems handle the initial search across millions or billions of records, while people make the final scientific judgments.
The workflow changes the task. Instead of searching an archive directly, astronomers can review a ranked list of candidates. That makes it practical to investigate unusual sources that might otherwise remain buried in existing observations.