For more than three decades, the Hubble Space Telescope has been observing the universe with unparalleled precision. From distant galaxies and nebulae to supernovae and gravitational lenses, its instruments have amassed a staggering amount of data. Over 35 years of observation, Hubble has generated hundreds of millions of images—a volume so vast that no human team could reasonably analyze it all. For a long time, part of this data remained unusable, not due to a lack of scientific interest, but because of cognitive limitations. This is precisely where artificial intelligence came into play, revealing 1,400 cosmic anomalies, the majority of which had never been observed before.
\n\n\n\nModern astronomy faces a paradox. Instruments are becoming increasingly powerful, but human capacity to analyze data remains limited. By way of comparison, the James Webb Space Telescope alone generates about 57 GB of data per day, a continuous stream that adds to Hubble’s historical archives. Faced with this cosmic data overload, two astronomers from the European Space Agency, David O’Ryan and Pablo Gómez, made a radical decision: to entrust the comprehensive analysis of the archives to artificial intelligence.
\n\n\n\nThe model developed, called AnomalyMatch, does not work like the traditional classification algorithms used in astronomy. Typically, computer vision models require supervised learning, in which humans manually label millions of images to teach the algorithm what a spiral galaxy, a star, or a quasar is. This approach is accurate, but extremely slow and biased by known categories.
\n\n\n\nAnomalyMatch takes the opposite approach. It relies on a distribution-based anomaly detection method, a field of unsupervised machine learning. The model first learns to define what is statistically “normal” in astronomical images. Once this normality has been modeled, it scans the entire dataset to identify anything that deviates significantly from it. In other words, it does not look for what it already knows, but rather for what does not fit into any existing category.
\n\n\n\nThe model’s efficiency is remarkable. In three days, AnomalyMatch analyzed approximately 100 million images from the Hubble archives using a single graphics processing unit (GPU). From this massive dataset, the AI identified 1,400 potential candidates deemed statistically anomalous. After validation by human astronomers, 1,300 of these objects were confirmed as genuine astrophysical anomalies, including approximately 800 that were completely new.
\n\n\n\nThe results were published in December 2025 in volume 704 of the journal *Astronomy & Astrophysics*, marking one of the most significant discoveries to emerge from an algorithmic reanalysis of astronomical data.
\n\n\n\nAmong the most significant discoveries are 86 potential new gravitational lenses. These phenomena, predicted by Albert Einstein, occur when a massive object bends spacetime enough to deflect light from a celestial object located behind it. They are key tools for mapping dark matter and measuring the expansion of the Universe.
\n\n\n\nAnomalyMatch has also identified 417 galaxies in the process of merging, providing a natural laboratory for studying large-scale gravitational dynamics. Another remarkable discovery is 35 new jellyfish galaxies, recognizable by their long trails of stars and gas, sculpted by the pressure of the intergalactic medium. These rare objects allow us to explore the extreme effects of interactions between galaxies and the cosmic environment.
\n\n\n\nThis study does not mean that AI is replacing astronomers. It acts as a cognitive filter, capable of sorting through, prioritizing, and revealing structures invisible to the human eye within massive amounts of data. Each anomaly detected then requires validation, interpretation, and scientific contextualization by experts.
\n\n\n\nThis approach reflects a profound shift in scientific research. Whereas researchers once formulated hypotheses before exploring the data, AI now allows unexpected phenomena to emerge, paving the way for unforeseen discoveries.
\n\n\n\nWhile AI significantly speeds up the discovery process, it also raises methodological questions. The anomalies detected depend on the model of normality learned by the algorithm. A bias in this modeling could obscure certain phenomena or exaggerate others. Furthermore, a statistical anomaly is not necessarily a new astrophysical object, hence the importance of human oversight.
\n\n\n\nDespite these limitations, this collaboration between astronomers and artificial intelligence marks a key milestone in the exploration of the cosmos. As observatories generate ever-increasing amounts of data, AI is becoming an indispensable tool for pushing the boundaries of knowledge.
\n\n\n\nThe contribution of artificial intelligence to astronomy illustrates, more broadly, how AI is transforming scientific disciplines faced with unprecedented volumes of data. On a related topic, check out our article “MLE-STAR: Google’s Approach to Effectively Structuring Machine Learning Engineering”, which analyzes methods for ensuring the reliability and scalability of AI models used in demanding research contexts.
\n\n\n\n1. O’Ryan, D., Gómez, P. (2025). Large-scale anomaly detection in Hubble archives. Astronomy & Astrophysics, Vol. 704.
https://www.aanda.org
2. ESA Hubble Science Archive. Hubble data legacy and archives.
https://www.cosmos.esa.int/web/hubble
3. Euclid Consortium. (2023). Unsupervised anomaly detection in astrophysical surveys.
https://arxiv.org/abs/2306.01234
4. NASA. James Webb Space Telescope data rates and operations.
https://www.nasa.gov/webb
For more than three decades, the Hubble Space Telescope has been observing the universe with unparalleled precision. From distant galaxies and nebulae to supernovae and gravitational lenses, its instruments have amassed a staggering amount of data. In 35 years of observation, Hubble…
\n","protected":false},"author":2,"featured_media":526200,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","footnotes":""},"categories":[129],"tags":[],"class_list":{"0":"post-526199","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-ia-science"},"yoast_head":"\n