For over thirty years, the Hubble Space Telescope has captured images of the Universe, amassing a vast amount of data. Faced with this abundance, scientists have been confronted with a harsh reality: the human impossibility of analyzing all these images, each one potentially containing a discovery.
Confronted with this mountain of information, researchers at the European Space Agency have developed an artificial intelligence model named AnomalyMatch. Designed to sift through Hubble's archives, this tool automatically scans images for unusual elements, mimicking the way our brain processes visual information.
A view captured by the Hubble Space Telescope showing distant galaxies.
Credit: ESA/Hubble & NASA, D. O'Ryan, P. Gómez (European Space Agency), M. Zamani (ESA/Hubble)
The analysis of nearly 100 million images has revealed more than 1300 anomalies, hundreds of which had never been documented before. These objects exhibit singular appearances that fall outside usual categories. Completed in just three days of processing, this flood of discoveries demonstrates the method's efficiency when dealing with an immense volume of data.
Among the findings are merging galaxies with massive star-forming clusters, structures resembling jellyfish with gaseous 'tentacles', or planetary formation disks seen edge-on within our own galaxy.
AnomalyMatch operates by learning normal visual patterns, allowing it to identify deviations. This systematic approach is a first for the Hubble archives, which cover 35 years of observations. It paves the way for deeper exploration without requiring prohibitive human time.
Project leaders indicate that this initiative shows how artificial intelligence can significantly increase scientific output. It also highlights the potential of astronomical archives for future discoveries, particularly within the framework of large observation campaigns.
This breakthrough is documented in a paper published in
Astronomy and Astrophysics in December 2025.
How AI pattern recognition works
The artificial intelligence used here relies on neural networks designed to analyze images. These systems are trained with vast datasets that have already been analyzed, where they learn to distinguish the normal features of astronomical objects. Once trained, they can quickly scan new images to spot what is out of the ordinary.
This method mimics certain processes of the human brain, such as detecting contours or textures, but on a much larger scale and at greater speed. It does not require explicit programming for each type of anomaly, making it adaptable to various fields.
Applications extend beyond astronomy, with possible uses in medicine for imaging analysis or in security for surveillance. In Hubble's case, it allows leveraging old data for new discoveries without having to redo observations.
The value of archived data in science
Scientific archives, like those of Hubble, constitute a precious resource that gains importance over time. They preserve observations made in the past, allowing for comparisons and reanalysis with more powerful tools. This avoids losing information and maximizes initial investments.
In the space domain, archived data covers long periods, revealing slow changes or rare events. For example, tracking the evolution of a galaxy over decades becomes possible thanks to these collections.
Access to these archives is facilitated by digital platforms, encouraging international collaboration and open science. Researchers worldwide can refer to them to test new ideas or validate theories.