Euclid Data Release 1 (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) C: Lens modelling pipeline overview

arXiv:2611.15888 · astro-ph.GA, astro-ph.CO, astro-ph.IM · Submitted 2026-11-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Euclid Data Release 1 (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) C".

Jocelyn: Here's the gist of the paper. Galaxy-scale strong lenses offer precise measurements of galaxy mass distributions, magnified views of distant galaxies, and serve as a competitive cosmological probe.

Vera: This research details an automated lens modelling pipeline applied to 15,072 strong-lens candidates discovered across the Euclid Data Release 1 (DR1) footprint through the ECLIPSE survey—the first end-to-end strong-lens discovery and characterization effort at this scale.

Jocelyn: The study addresses the scalability of strong lens modelling beyond thousands of systems.

Vera: The analysis involved three distinct fitting procedures using PyAutoLens-JAX, employing multi-Gaussian expansion (MGE), pixelised representations, and Sérsic representations of the source, all adopting an isothermal ellipsoid mass model with external shear.

Paper summary: Vera: So we've been looking at how they actually built this pipeline for strong lenses using these Euclid data sets.

Jocelyn: And now we get to talk about what they actually wrote in that paper, "Euclid Data Release one (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) C: Lens modelling pipeline overview".

Subrahmanyan: It’s a lot of detail on how they automated the process of taking those thousands of lens candidates and modeling them.

Vera: They basically show that you can take this whole massive data set, all those candidates, and run it through three different fitting procedures to get a model.

Jocelyn: The key thing they report is the success rate when people actually check the models—that they needed at least seventy five percent of experts to agree for a model to pass.

Subrahmanyan: And for the six thousand seventy three grade A candidates, it worked about sixty seven point five percent of the time. That’s a solid number for a complex modeling task.

Vera: It also gives us some median numbers, like the effective Einstein radius being zero point eight five arcseconds and magnification being eight point one in IE band.

Jocelyn: But they don't just give you the good news; they talk about where it went wrong too. They show that for grade B systems, the success rate drops because there's a lot more disagreement among inspectors.

Subrahmanyan: And those failures often point to inaccurate lens light subtraction being the biggest issue, which makes sense when you’re trying to separate the source from the lens itself.

Vera: So what does this mean for us, Jocelyn? What is this paper actually telling us about using Euclid data?

Jocelyn: It means we have a systematic way to take all that raw imaging and turn it into a population of lenses that we can actually study on a large scale.

Subrahmanyan: And from my side, it sets up the foundation for testing how matter clumps together across the whole universe, not just looking at one isolated event.

Vera: It moves us past just studying individual pretty pictures to understanding the overall structure formation in space.

Jocelyn: So we have a tool ready to go for these massive datasets, but there are still those lower grade systems where things get messy and uncertainty creeps in.

Conclusion: Vera: So we’ve seen how they built this lens pipeline for strong lenses using Euclid data sets.

Jocelyn: Now we talk about that specific paper, "Euclid Data Release one (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) C: Lens modelling pipeline overview".

Subrahmanyan: It’s a lot of detail on how they automated the process of taking those thousands of lens candidates and modeling them.

Vera: They basically show that you can take this whole massive data set, all those candidates, and run it through three different fitting procedures to get a model.

Jocelyn: The key thing they report is the success rate when people actually check the models—that they needed at least seventy-five percent of experts to agree for a model to pass.

Subrahmanyan: And for the six thousand seventy-three grade A candidates, it worked about sixty seven point five percent of the time. That’s a solid number for a complex modeling task.

Vera: It also gives us some median numbers, like the effective Einstein radius being zero point eight five arcseconds and magnification being eight point one in IE band.

Jocelyn: But they don't just give you the good news; they talk about where it went wrong too. They show that for grade B systems, the success rate drops because there's a lot more disagreement among inspectors.

Subrahmanyan: And those failures often point to inaccurate lens light subtraction being the biggest issue, which makes sense when you’re trying to separate the source from the lens itself.

Vera: So what does this mean for us, Jocelyn? What is this paper actually telling us about using Euclid data?

Jocelyn: It means we have a systematic way to take all that raw imaging and turn it into a population of lenses that we can actually study on a large scale.

Subrahmanyan: And from my side, it sets up the foundation for testing how matter clumps together across the whole universe, not just looking at one isolated event.

Vera: It moves us past just studying individual pretty pictures to understanding the overall structure formation in space.

Jocelyn: So we have a tool ready to go for these massive datasets, but there are still those lower grade systems where things get messy and uncertainty creeps in.

J. W. Nightingale, S. H. Vincken, L. R. Ecker, A. Nersesian, D. J. Bowden, C. M.’Riordan, N.’E.’P.’Lines, T. E.’Collett

ESO

astro-ph.GA, astro-ph.CO, astro-ph.IM

Submitted: 2026-11-11

Updated: 2026-11-11

Comments: 29 pages, 16 figures. Submitted to A&A. Catalogue: https://euclid-eclipse-lenses.org ; data: https://doi.org/10.5281/zenodo.19654904

Code: https://github.com/PyAutoLabs/PyAutoLens

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Galaxy-scale strong lenses offer precise measurements of galaxy mass distributions, magnified views of distant galaxies, and serve as a competitive cosmological probe.

Key concepts

Strong-lens candidates
These are distant galaxies whose light is significantly distorted by the gravity of a massive foreground object (the lens). They are used as precise tools to measure the mass distribution of galaxies and probe cosmological models.
Lens modelling pipeline
This is an automated, step-by-step computational process designed to fit mathematical models to the observed data. It takes raw image information from Euclid and attempts to reconstruct the physical properties of both the foreground lens and its background source.
Multi-Gaussian Expansion (MGE)
This is one of three distinct fitting procedures used in the study. MGE is a method that models complex light distributions by combining multiple Gaussian shapes. It helps accurately separate and characterize the light coming from different components, such as the lens and the background galaxy.
Einstein radius
The Einstein radius is a critical physical measurement derived from strong lensing observations. It represents the characteristic angular scale of magnification caused by a lens, providing a direct measure of how strongly gravity is bending light in that specific system.

Terminology

Summary

Galaxy-scale strong lenses offer precise measurements of galaxy mass distributions, magnified views of distant galaxies, and serve as a competitive cosmological probe. p. 0

Where the paper says it

Collett, et al. (Full author list and affiliations details can be found after the references) ABSTRACT Galaxy-scale strong lenses provide precise measurements of galaxy mass distributions, magnified views of distant galaxies, and a competitive cosmological probe.

This research details an automated lens modelling pipeline applied to 15,072 strong-lens candidates discovered across the Euclid Data Release 1 (DR1) footprint through the ECLIPSE survey—the first end-to-end strong-lens discovery and characterization effort at this scale. p. 1
Where the paper says it

This paper presents strong lens modelling results for 15 072 grade A and B strong-lens candidates discovered across the Euclid DR1 footprint as part of the ECLIPSE survey (Euclid Collaboration: Lines et al. 2026a), the first end-to-end strong-lens discovery and characterisation effort at this scale.

The study addresses the scalability of strong lens modelling beyond thousands of systems. p. 1

Where the paper says it

2026a), the first end-to-end strong-lens discovery and characterisation effort at this scale. In doing so, we address a long-standing question in strong lensing: can strong lens modelling be scaled beyond thousands of lenses?

The analysis involved three distinct fitting procedures using PyAutoLens-JAX, employing multi-Gaussian expansion (MGE), pixelised representations, and Sérsic representations of the source, all adopting an isothermal ellipsoid mass model with external shear. p. 0
Where the paper says it

We present automated lens modelling of 15 072 strong-lens candidates from Euclid Data Release 1 (DR1), over an order of magnitude larger than any previously modelled sample. Using PyAutoLens-JAX, we perform three fits with multi-Gaussian expansion (MGE), pixelised, and Sérsic representations of the source, adopting an isothermal ellipsoid mass model with external shear.

These methods are used to derive foreground lens-light subtraction, adaptive Delaunay source reconstructions, and deblended photometry. p. 0
Where the paper says it

Using PyAutoLens-JAX, we perform three fits with multi-Gaussian expansion (MGE), pixelised, and Sérsic representations of the source, adopting an isothermal ellipsoid mass model with external shear. These fits provide foreground lens-light subtraction, adaptive Delaunay source reconstructions, and deblended photometry.

The modelling success was assessed by over 20 expert inspectors for each resulting model; a classification was assigned based on agreement among at least 75% of the inspectors. p. 0

Where the paper says it

These fits provide foreground lens-light subtraction, adaptive Delaunay source reconstructions, and deblended photometry. More than 20 expert inspectors visually assess the resulting models, and each model is assigned the classification on which at least 75% of its inspectors agree.

The results showed that for the 6073 grade A (secure) lens candidates, 4098 (67. p. 0
Where the paper says it

More than 20 expert inspectors visually assess the resulting models, and each model is assigned the classification on which at least 75% of its inspectors agree. Of the 6073 grade A (secure) lens candidates, inspectors classify 4098 (67.

5%) were successfully modelled, with a median effective Einstein radius of 0''.85 and a median total IE-band source magnification of 8. p. 0
Where the paper says it

Successfully modelled grade A and B systems have a median effective Einstein radius of 0. ′′85 and a median total IE-band source magnification of 8.

1.

Key findings include:

  • Modelling Success Rates: At 75% expert agreement, 4098 of 6073 grade A candidates (67.5%) and 3882 of 8999 grade B candidates (43. p. 7

    Where the paper says it

    2), 4098 of 6073 grade A candidates (67. 5%) are classified as successfully modelled (SC), compared with 3882 of 8999 grade B candidates (43.

    1%) were successfully modelled, totaling 7,980 successful models. p. 1
    Where the paper says it

    This enables us to quantify the success rate of automated lens modelling, provide a catalogue of validated lens models for future scientific study,1 and systematically identify the dominant failure modes of automated lens modelling. For successfully modelled systems, we measure the principal lens and source properties to characterise the sample and provide a foundation for subsequent scientific studies.

  • Grade B Uncertainty: The lower success rate for Grade B systems is primarily attributed to greater disagreement among inspectors (37.6% without consensus), rather than a comparable increase in agreed failures, with much of this disagreement stemming from uncertainty regarding the genuine nature of Grade B candidates as lenses. p. 0

    Where the paper says it

    6%), rather than a comparable increase in agreed failures (12. We attribute this greater disagreement to uncertainty over whether these candidates are genuine lenses.

  • Modelling Failures: For a subset of 2990 candidates, 210 consensus modelling failures were analyzed. The most frequent causes were inaccurate lens-light subtraction (31. p. 0

    Where the paper says it

    6%) to preprocessing and segmentation errors, 66 (31. 4%) to inaccurate MGE lens-light subtraction, and 12 (5.

    4%), followed by segmentation errors (18. p. 0
    Where the paper says it

    Of the 210 failures, they attribute 39 (18. 6%) to preprocessing and segmentation errors, 66 (31.

    6%) and source-reconstruction or mass-model failures (5. p. 0
    Where the paper says it

    4%) to inaccurate MGE lens-light subtraction, and 12 (5. 7%) to source-reconstruction or mass-model failures, while 93 (44.

    7%).

  • Lens Properties: Successfully modelled systems exhibit a median Einstein radius of 0''.85, with the 16th–84th percentile range spanning 0''.61 to 1''.29.

The ECLIPSE catalogue provides the successful lens models and their inferred properties, facilitating population-level studies and detailed investigations of individual systems. p. 0

Where the paper says it

′′85 and a median total IE-band source magnification of 8. We publicly release the successful lens models, their inferred properties, and magnifiedsource reconstructions through the ECLIPSE catalogue, demonstrating automated modelling at Euclid sample sizes and enabling population-level strong-lensing science.

The methodology involved a complex pipeline where MGE fitting utilized a joint linear solution framework (H24) with non-negative least squares solvers (JaxNNLS), adaptive split regularisation based on Delaunay triangulation, and specific statistical objectives (chi 2) tailored for both analytic fits and pixelised stages. This demonstrates that automated strong-lens modelling can successfully scale to the thousands of systems discovered by Euclid, providing a robust foundation for studying galaxy mass distributions, evolution, and cosmology. p. 0
Where the paper says it

(Full author list and affiliations details can be found after the references) ABSTRACT Galaxy-scale strong lenses provide precise measurements of galaxy mass distributions, magnified views of distant galaxies, and a competitive cosmological probe. Historically, such studies have been limited to hundreds of lenses, with automated modelling methods proving difficult to scale.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated preprocessing can be improved by developing machine learning frameworks that segment foreground lens light, lensed source emission, and nearby galaxies to define fitting masks and candidate conjugate points automatically using a framework like Euclid Collaboration: Vincken et al. (2026a). This automation makes the masking procedure reproducible for thousands of lenses.

  2. The AI system can be trained to identify double-source-plane lenses (DSPL) and multi-galaxy lenses (MGL) by assessing model diagnostics, as inspectors are instructed to identify DSPL and MGL systems as standalone categories in the initial classification stage. This would improve the accuracy of classifying systems that violate the single-source-plane assumption.

  3. The system can be enhanced by incorporating agentic AI to assess imaging and model diagnostics together, allowing an AI agent to identify reliable models, diagnose failures, and recommend or perform targeted remodelling. This addresses the bottleneck of human inspection for scaling to over 100 000 lenses.

  4. The system can utilize machine-learning predictions as priors for PyAutoLens by incorporating neural-network predictions and uncertainties from methods like LeMoN (Euclid Collaboration: Busillo et al. 2026). This could improve convergence and recover failed models by informing the fitting process with rapid estimates of lens mass and light parameters.

  5. The system can be refined by developing complementary machine-learning classifiers to specifically identify double-source-plane lenses, multi-galaxy lenses, and other systems requiring specialized modelling. These classifiers would assist human inspectors in focusing on ambiguous cases while automating the identification of complex lens types.

Abstract

Galaxy-scale strong lenses provide precise measurements of galaxy mass distributions, magnified views of distant galaxies, and a competitive cosmological probe. Historically, such studies have been limited to hundreds of lenses, with automated modelling methods proving difficult to scale. We present automated lens modelling of 15 072 strong-lens candidates from Euclid Data Release 1 (DR1), over an order of magnitude larger than any previously modelled sample. Using PyAutoLens-JAX, we perform three fits with multi-Gaussian expansion (MGE), pixelised, and Sérsic representations of the source, adopting an isothermal ellipsoid mass model with external shear. These fits provide foreground lens-light subtraction, adaptive Delaunay source reconstructions, and deblended photometry. More than 20 expert inspectors visually assess the resulting models, and each model is assigned the classification on which at least 75% of its inspectors agree. Of the 6073 grade A (secure) lens candidates, inspectors classify 4098 (67.5%) as successfully modelled, 580 (9.6%) as failures, 317 (5.2%) as multi-galaxy lenses, 15 (0.2%) as double-source-plane lenses, and 7 (0.1%) as unsure if lens, while 1032 (17.0%) reach no consensus and 24 (0.4%) have incomplete models. For the 8999 grade B (probable) candidates, the modelling success rate falls to 43.1%, primarily reflecting an increase in cases without consensus (37.6%), rather than a comparable increase in agreed failures (12.9%). Successfully modelled grade A and B systems have a median effective Einstein radius of 0.85 arcsec and a median total I E-band source magnification of 8.1. We publicly release the successful lens models, their inferred properties, and magnified-source reconstructions through the ECLIPSE catalogue, demonstrating automated modelling at Euclid sample sizes and enabling population-level strong-lensing science.

Sources

Related papers