AI spots 800 cosmic anomalies hiding in Hubble data

Hubble has been taking pictures of the cosmos for 35 years, but there is a catch: scientists simply do not have the time to look through every single snapshot. Decades of cosmic oddities have been sitting on hard drives, completely unnoticed.

Now, a pair of astronomers at the European Space Agency (ESA) have built an AI tool that cuts through decades of digital dust in just a weekend.

Sifting 100 million cutouts in 60 hours

Astronomers David O'Ryan and Pablo Gómez developed a neural network called AnomalyMatch, designed specifically to flag cosmic oddities that do not fit standard galactic templates.

They turned the model loose on the Hubble Legacy Archive, a treasure trove containing tens of thousands of datasets gathered across Hubble's 35-year mission. ESA pointed out that while human eyes are brilliant at spotting anomalies, the sheer volume of high-resolution data makes manual inspection impossible.

AnomalyMatch scanned nearly 100 million image cutouts in just two and a half days.

It did not completely sideline humans, though. The model spit out a shortlist of candidates, which O'Ryan and Gómez then reviewed by hand. The result was 1,400 confirmed cosmic anomalies, and more than 800 of them had never been documented before.

Cosmic jellyfish and unclassifiable objects

Most of the flagged anomalies turned out to be interacting or merging galaxies, warped into bizarre forms with trailing tails of gas and stars.

Others were gravitational lenses, where the intense gravity of a foreground galaxy warps spacetime and bends background starlight into dramatic rings and arcs. The scan also caught planet-forming disks viewed edge-on, clumpy star-forming galaxies, and so-called jellyfish galaxies.

Intriguingly, several dozen of the discovered objects completely defied existing astronomical classifications, presenting shapes that researchers cannot yet explain.

As Gómez noted, pulling this many brand-new discoveries out of an archive people assumed had already been thoroughly picked over proves the power of the technique.

Why it matters

We see this as a textbook example of AI doing what it actually does best: acting as an exhaustive force multiplier for scientists rather than a synthetic daydream machine.

For astronomers, the real bottleneck today is no longer capturing sky data, but finding the needles in planetary-scale haystacks. AnomalyMatch is clever because it does not try to classify or theorize. It simply narrows down 100 million cutouts into 1,400 curated suspects, leaving the final scientific judgment to expert eyes.

This approach will inevitably reshape astronomy workflows. Hubble's archive is tiny compared to the torrents of data heading our way from next-generation observatories and sky surveys. If an AI can dig up 800 brand-new anomalies from a 35-year-old dataset in a single weekend, applying similar tools to incoming survey streams will become mandatory standard practice.

The only real risk here is algorithmic bias. An anomaly-detection model is only as good as the baseline it is trained against, and there is always a chance it overlooks truly revolutionary physics that mimics instrumental noise. Still, as a way to squeeze massive scientific yield out of existing public investments, it is an absolute home run.