🔭 astrophysics

High-Redshift Gravitational Lens Discoveries in JWST NIRCam Using AnomalyMatch

This paper demonstrates the effectiveness of the semi-supervised machine learning tool AnomalyMatch in identifying 58 high-redshift gravitational lenses, including 37 previously uncatalogued systems, within JWST NIRCam data from the ASTRODEEP and COSMOS-Web surveys.

Original authors: Julia Dima, David O'Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Julia Dima, David O'Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Needle in a Cosmic Haystack

Imagine you are looking at a massive, high-resolution photograph of a crowded city square taken from a helicopter. You are looking for a very specific, rare event: a person wearing a bright red hat standing next to a person wearing a bright blue hat, where the red hat is slightly distorted because of a weird glass lens in front of the camera.

In the universe, this "red hat" is a distant galaxy, and the "glass lens" is a massive galaxy sitting in front of it. The gravity of the front galaxy bends the light of the back galaxy, stretching it into a curved arc or a ring. This is called gravitational lensing.

The problem? These events are incredibly rare. In the vast archives of the James Webb Space Telescope (JWST), which has taken millions of pictures of the deep universe, finding these specific "bent light" patterns is like finding that needle in a haystack the size of a continent.

The Tool: AnomalyMatch (The "Smart Detective")

The researchers didn't want to look at every single picture one by one (which would take a human lifetime). Instead, they used a computer program called AnomalyMatch.

Think of AnomalyMatch as a smart detective that has been trained to spot "weirdness."

The Training: The detective was shown only 11 examples of these rare lensing events (the "needles") and 400 examples of normal galaxies (the "hay").
The Learning: Instead of needing thousands of examples to learn, this detective uses a special "semi-supervised" trick. It looks at the 11 examples, then looks at millions of unknown pictures, guesses which ones look a bit like the examples, and asks a human expert, "Is this one a lens?"
The Loop: If the human says "Yes," the detective learns. If the human says "No," the detective learns what not to look for. This happens in a loop, getting smarter with every question.

The Search: Scanning the Archives

The team used this detective to scan two massive collections of JWST data:

ASTRODEEP: A compilation of deep space surveys.
COSMOS-Web: A specific, wide-area survey of a patch of sky.

They processed 614,015 individual galaxy images. To make the images easier for the computer to read, they turned them into colorful "cutouts" (like taking a small square out of a larger photo) and adjusted the colors to make faint details pop, similar to how you might boost the contrast on an old photo to see a hidden face.

The Results: 58 New Discoveries

After the detective did its work, the team found 58 potential gravitational lenses. They didn't just take the computer's word for it; they had four human experts look at the top candidates and grade them like a school report:

Grade A (The "A+" Students): 16 lenses. These are the clear winners. They have perfect, bright arcs or rings that are clearly different colors from the galaxy in front of them.
Grade B (The "B" Students): 16 lenses. These look very promising but might be missing a small detail (like a second image) or the colors aren't quite distinct enough to be 100% sure.
Grade C (The "C" Students): 26 lenses. These are the "maybe" pile. They have some weird features that could be a lens, but they might also just be a galaxy merger or an optical illusion.

The Best Part:
Out of these 58 discoveries, 37 of them were completely new to science. No one had ever cataloged them before.

The highest redshift (meaning the most distant and oldest) new discovery is at z = 2.1.
To put that in perspective: This is a galaxy we are seeing as it existed when the universe was much younger, closer to the "Cosmic Noon" era.

Why This Matters

It's Fast: The entire search of over 600,000 images was done on a single graphics card in a matter of hours. This proves that we can scan the entire JWST archive quickly without needing armies of human volunteers.
It's Efficient: The team proved you don't need a massive database of labeled examples to find rare things. You can start with just a handful (11 lenses) and let the AI learn the rest.
It's Deep: Because JWST can see very far back in time, this method helps us find lenses that older telescopes (like Hubble) simply couldn't see because they were too faint or too far away.

The Caveats

The authors are honest about the limitations. Because they trained the AI on only the "spectacular" (very obvious) lenses, the system might miss the subtle, faint ones. Also, the "Grade C" candidates are just possibilities; they need more study to confirm they are real lenses and not just cosmic coincidences.

Summary

This paper is a success story of human-AI teamwork. By teaching a computer to spot the "weird" shapes of gravitational lenses using just a few examples, the team unlocked 37 new cosmic mysteries hidden in the JWST archives, proving that we can efficiently explore the deep universe without getting lost in the data.

Technical Summary: High-Redshift Gravitational Lens Discoveries in JWST NIRCam using AnomalyMatch

Problem Statement
Strong gravitational lenses are rare astrophysical phenomena that provide critical constraints on mass distributions, galaxy evolution, and cosmological parameters. However, their intrinsic rarity and subtle visual features make them difficult to identify within the vast datasets produced by modern observatories. Traditional detection methods, including visual inspection and morphological algorithms, struggle to scale with the volume of data from facilities like the James Webb Space Telescope (JWST). Furthermore, supervised machine learning approaches typically require large, labeled training sets, which are difficult to assemble for intrinsically rare objects like gravitational lenses.

Methodology
The authors employ a semi-supervised active learning framework called AnomalyMatch to search for gravitational lenses in JWST NIRCam data. The workflow proceeds as follows:

Data Selection: The study utilizes the ASTRODEEP photometric catalogues (covering Abell 2744, CEERS, JADES, NGDEEP, and PRIMER fields) and the COSMOS-Web survey. After cross-matching with Level 3 calibration data on the ESA Datalabs platform and applying a size cutoff to remove unresolved high-redshift objects, the final search pool consists of 614,015 sources across approximately 0.74 square degrees.
Image Processing: Three-channel RGB cutouts (224 × 224 pixels) are generated from four NIRCam filters (F115W, F150W, F277W, F444W). To enhance the visibility of low-surface-brightness lensing features near bright galaxy cores, the authors apply a mid-tones transfer function (MTF) to normalize dynamic range. The four bands are combined into RGB images using empirically determined weights to balance the galaxy center, lensing features, and sky background.
Training Strategy: AnomalyMatch utilizes the FixMatch approach, combining supervised cross-entropy loss on labeled data with unsupervised consistency regularization on pseudo-labeled augmented unlabelled data. The initial training set is extremely small, consisting of only 11 "spectacular" lenses from the COSMOS-Web Lens Survey (COWLS) and 400 nominal (non-lens) sources.
Active Learning Loop: The model undergoes three training cycles. In each cycle, the top 1,000 ranked anomalies are visually inspected by experts. Confirmed lenses are added to the labeled training set to refine the model. This human-in-the-loop process allows the system to iteratively improve performance with minimal initial labels.
Classification and Grading: Identified candidates are graded by four experts via majority voting on the Zooniverse platform into three categories:
- Grade A: High confidence (clear arcs with distinct colors and counter-images).
- Grade B: Probable (plausible geometry but lacking definitive features like color separation).
- Grade C: Possible (weak/ambiguous features, potentially contaminated by mergers or artifacts).

Key Results

Discovery Count: The search identified 58 unique gravitational lens systems.
Novelty: Of the 58 systems, 37 (63.8%) were previously uncatalogued in the literature. This includes 9 Grade A, 10 Grade B, and 18 Grade C systems.
Redshift Reach: The identified lenses span a wide redshift range. The highest redshift newly discovered candidate is at $z_{phot} \approx 2.1$ (Grade C). The highest spectroscopic redshift identified is $z_{spec} = 1.39$ (Grade B). The full sample extends from $z \approx 0.09$ to $z_{phot} = 2.21$ .
Efficiency: The entire search and training process for these surveys was completed on the order of hours using a single GPU, demonstrating the scalability of the approach.

Significance and Claims
The paper claims to demonstrate the potential of AnomalyMatch as a label-efficient tool for large-scale searches for rare objects in JWST archives. By leveraging semi-supervised learning and active learning, the authors show that high-performance lens detection is achievable with very few initial examples (11 lenses), overcoming the bottleneck of assembling large labeled datasets.

The work highlights the synergy between ESA Datalabs and advanced machine learning, proving that terabytes of archival data can be scanned efficiently to uncover systems missed by previous surveys. The authors assert that this approach extends the redshift reach of lens discovery beyond what was possible with earlier observatories and previous JWST visual searches (like COWLS), pushing detection capabilities toward "Cosmic Noon" and the early Universe. The resulting graded catalogue serves as a resource for future machine learning training and scientific study, though the authors caution that lower-grade candidates (B and C) may contain contamination and should be treated with appropriate skepticism.