For over thirty years, the Hubble Space Telescope has been capturing targeted images to explore various astronomical phenomena, from mapping distant galaxies to examining nearby nebulae. With an immense volume of data collected, astronomers have struggled to thoroughly analyze each piece of information.
Among the vast amount of data, there exist numerous unexpected objects that were not the focus of initial observations. Recently, two astronomers took on the challenge of revisiting this extensive archive with an innovative approach. They employed an artificial intelligence system designed to identify anomalies that appear "off." Within just 60 hours of processing, the AI flagged more than 1,300 unusual objects hidden in 100 million Hubble images, many of which had never been documented in scientific literature.
A Sea of Cosmic Thumbnails
The research, published in Astronomy & Astrophysics, concentrated on the Hubble Legacy Archive and a dataset derived from it, consisting of 99.6 million "cutouts," each representing a small square of the sky centered on an extended source, typically a galaxy.
These cutouts are uniform in size, measuring 150 by 150 pixels, and were generated from Hubble's Advanced Camera for Surveys. They were saved as single-channel grayscale JPEGs, effectively preserving shapes while optimizing storage.
David O'Ryan and Pablo Gómez created a system named AnomalyMatch to navigate through this extensive collection of thumbnails. Initially, their goal was simple: to locate "protoplanetary disks," the rare, edge-on dusty structures where planets form. The AI was trained using only three known examples.
However, as the system learned to recognize what constituted "unusual," it began to uncover a broader range of anomalies, including warped gravitational lenses and colliding galaxies. Running on ESA's Datalabs platform, the AI accomplished in two and a half days what would take a human a lifetime to achieve.
Once the AI completed its search, it generated a ranked list of the most unusual images. The researchers then manually reviewed the top 5,000 images, eliminating duplicates caused by catalog errors, ultimately identifying 1,339 unique anomalies.
About half of these anomalies were merging or interacting galaxies, distorted by gravitational forces into irregular shapes with trailing tidal tails. Many others resembled gravitational lenses--alignments where a foreground galaxy bends the light of a more distant galaxy into arcs or rings.
A Gallery of the Bizarre
The productivity of telescopes has reached a point where the discovery process relies heavily on prioritization--determining which phenomena warrant further investigation. AnomalyMatch presents a potential solution to this challenge.
Hubble's archive spans decades, but it is not designed as a survey telescope. Researchers typically seek specific targets, often neglecting the wealth of data beyond these primary objectives. Remarkably, around 65% of the anomalies identified in this study (811 objects) were entirely new to scientific literature.
Some of these "new" objects may simply have gone unreported; they could be background galaxies observed incidentally or features overlooked due to the focus on main targets.
The findings of this study reveal a fascinating array of astrophysical phenomena: 629 merging systems, 140 candidate gravitational lenses, 35 jellyfish galaxies, and 43 objects that defy classification. The latter group is so unique that the researchers opted not to assign labels.
These extremes provide valuable insights into cosmic mechanisms. A strong gravitational lens can unveil the distribution of mass, including dark matter, within a galaxy. Meanwhile, a jellyfish galaxy illustrates how a galaxy loses gas as it traverses the sparse plasma of a cluster.
This research also indicates a shift in the future of astronomy as new observatories become operational. ESA's Euclid mission is already mapping the universe on a large scale, while NASA's upcoming Nancy Grace Roman Space Telescope and the Vera C. Rubin Observatory are expected to generate vast amounts of imagery with less selective human intervention than Hubble's targeted approach.
In this evolving landscape, the role of astronomers may transform into that of an editor--analyzing algorithmically identified candidates, determining which phenomena merit further exploration, and discerning between mere curiosities and groundbreaking discoveries.