High-Redshift Gravitational Lens Discoveries in JWST NIRCam using AnomalyMatch

Julia Dima, David O’Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez · 2026

Read High-Redshift Gravitational Lens Discoveries in JWST NIRCam using AnomalyMatch by Julia Dima, David O’Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03442.txt.

Introduction

Julia Dima, David O'Ryan, Sandor Kruk, Laslo E. Ruhberg, and Pablo Gómez hunt high-redshift gravitational lenses in JWST images with a semi-supervised anomaly finder. Strong lenses probe distant mass and source populations, but they are rare and visually subtle in the flood of space telescope data.

Starting from eleven known lenses, AnomalyMatch with human-in-the-loop training scanned two major surveys and found fifty-eight unique lenses, graded by four experts. Thirty-seven were previously uncatalogued, with photometric redshifts reaching beyond two. The haul shows anomaly methods can mine JWST archives for rare distant objects at scale.

Catalog record

CREATORJulia Dima, David O’Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.03442

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

High-Redshift Gravitational Lens Discoveries in JWST NIRCam using AnomalyMatch Julia Dima, David O’Ryan, Sandor Kruk, Laslo E. Ruhberg, Pablo Gómez Abstract Context. Strong gravitational lenses provide a unique tool to probe cosmology and astrophysics at high redshift, offering constraints on the mass distribution of background source populations. Despite their scientific value, their rarity and subtle visual features make them challenging to identify in the wealth of data delivered by facilities such as the James Webb Space Telescope (JWST), whose unmatched resolution and near-infrared coverage make it particularly well-suited to detecting lensing systems in this regime. Aims. We make use of the specialised open-source software AnomalyMatch , a semi-supervised learning method to trawl the ASTRODEEP and COSMOS-Web surveys for gravitational lenses. Methods. Building on a training dataset of eleven previously identified gravitational lenses, we use AnomalyMatch and its iterative human-in-the-loop method to train a neural network to identify gravitational lenses in JWST Level 3 products using ESA Datalabs. Results. In total we identify 58 unique gravitational lenses. These are graded by four experts into 16 Grade A, 16 Grade B, and 26 Grade C lenses. Of all lenses identified, 37 were previously uncatalogued. We analyse their properties such as photometric redshift measurements and spectroscopic redshift, when the latter is available. The lenses previously identified span spectroscopic redshifts to z spec ≤ 1.39 z_{\text{spec}}\leq 1.39 and photometric redshifts to z phot ≤ 2.21 z_{\text{phot}}\leq 2.21 . The uncatalogued lens system with the highest redshift is at z phot = 2.1 z_{\text{phot}}=2.1 . Conclusions. Overall, we demonstrate the potential of AnomalyMatch for large-scale searches for gravitational lenses and other rare high-redshift objects in JWST archives. 1 Introduction Strong gravitational lensing provides powerful insights into the distribution of mass (both baryonic and dark) in galaxies and galaxy clusters ( 41 ; 7 ; 62 ; 71 ) , the evolution of galaxies ( 6 ; 20 ; 40 ) , and cosmological parameters of the Universe ( 47 ; 17 ) . Of particular interest are galaxy-galaxy strong gravitational lenses. These occur due to the chance alignment of two galaxies in the plane of the sky, where the gravitational potential of the foreground galaxy (the ’lens’) causes the light of the background galaxy to be bent around it and be smeared into an arc. The morphology of these arcs is strongly dependent on the geometry of the lensing system, and lensing magnification additionally allows high-resolution imaging of the background source ( 11 ; 25 ; 10 ; 3 ) . Therefore, identifying lensing sources at high redshifts guarantees that we can study highly magnified systems at much earlier epochs of the Universe’s evolution ( 33 ; 59 ) . However, lensed galaxies are not the only systems of interest. Lensed transients such as supernovae can provide independent constraints on H 0 via time-delay cosmography ( 22 ; 57 ; 12 ; 2 ) , as can multiply-imaged quasars ( 31 ; 52 ; 60 ) . If the lens is a galaxy cluster rather than a single galaxy, background sources are stretched into large arcs - or, in cases of near-perfect alignment, near-complete Einstein rings - enabling detailed reconstruction of the cluster’s dark matter distribution ( 54 ) . Because gravitational lensing depends on chance alignments along the line of sight, such systems are intrinsically rare. Existing catalogues typically contain on the order of a few 10s to \sim 1000 confirmed lensing systems ( 34 , e.g.) . For identification of further examples, we must search large observational datasets spanning either large areas of the sky or to high depths to probe a large redshift distribution. The James Webb Space Telescope ( JWST ) offers unprecedented opportunities for lens discovery due to its resolution, depth, and near-infrared coverage, which is particularly advantageous for detecting lenses at higher redshifts ( 59 ) . Launched in 2021, this space observatory has built a large archive of data hosted on platforms such as the Barbara A. Mikulski Archive for Space Telescopes (MAST) 1 1 1 https://mast.stsci.edu/search/ui/#/jwst . We focus on observations made with the Near-Infrared Camera (NIRCam) through two parallel modules each containing short ( F115W , F150W used in this work) and long ( F277W , F444W used in this work) wavelength channels. NIRCam’s sensitivity to faint, high-redshift galaxies makes it well-suited for identifying strong lensing systems, while its wide set of near-infrared filters enables colour composite images in which the colour separation enhances the visual detectability of lensing features and aids automated algorithms in distinguishing the lens from the lensed source. Although JWST is primarily a ‘point-and-shoot’ telescope - limiting the discovery of unknown lenses in targeted observations - several wide-area surveys such as COSMOS-Web ( 19 ) , JADES ( 26 ) , and PRIMER ( 23 ) provide ideal datasets for searching for gravitational lenses across many hundreds of thousands of sources and a wide redshift range. Traditional methods for gravitational lens detection - including visual inspection ( 45 ) , morphological/arc-finding algorithms ( 65 ; 15 ) , and spectroscopic dual-redshift identification ( 14 ; 69 ; 70 ) - are limited in their scalability and reliability. Manual inspection is hampered both by the sheer data volumes of modern surveys and by the difficulty of distinguishing true lenses from interlopers such as mergers, edge-on galaxies, and low-surface-brightness structures. Machine learning offers a compelling alternative, and since lenses are extremely rare among galaxies, their detection can be framed as an anomaly detection problem - a task extensively studied across supervised ( 39 ) , semi-supervised ( 63 ) , and unsupervised ( 24 ; 68 ) paradigms, with dedicated astronomical tools such as Astronomaly ( 49 ) , AnomalyMatch ( 38 ) and LAISS ( 1 ) . Large supervised models fine-tuned specifically for lens detection have also been applied across observatories, from space-based missions such as Euclid ( 56 ; 48 ; 28 ) to ground-based surveys such as Pan-STARRS ( 18 ) . However, such supervised approaches typically require large labelled training sets, which are difficult to assemble for intrinsically rare objects such as gravitational lenses. In this work, we leverage the newly developed open-source software AnomalyMatch ( 38 ; 61 ) , a semi-supervised active learning framework specialised in large-scale searches for efficient identification of rare objects. Through an additional human-in-the-loop learning step ( 73 ) and the FixMatch approach ( 67 ) , AnomalyMatch utilises both labelled and unlabelled data to train a neural network to present likely candidate objects of interest. Through an intuitive user interface, the user can then refine the model during training by providing additional labels on highly scored objects, iteratively improving performance. Thus, the approach facilitates rapid identification of the desired anomaly with very few initial examples. The main contributions of this work are as follows: • We present the first application of AnomalyMatch to gravitational lens discovery, demonstrating its effectiveness for identifying rare morphological features in JWST imaging. • We search 614,015 individual NIRCam cutouts from the ASTRODEEP and COSMOS-Web surveys, identifying 58 lens candidates classified into Grades A, B, and C based on confidence level: 16 Grade A (high-confidence), 16 Grade B (probable), and 26 Grade C (possible) candidates. • Of these,

Read the complete text