Euclid Data Release 1 (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) B: Discovery pipeline overview
summary
The gist
The paper presents a multi-stage method developed to identify approximately 15,000 high-grade galaxy–galaxy strong lens candidates within Euclid Data Release 1 (DR1).
In short
A multi-stage method was developed to find about 15,000 high-grade galaxy-galaxy strong lens candidates in Euclid Data Release 1. This pipeline combines catalogue cuts, machine learning, citizen science inspection, and expert grading to maximize lens discovery while minimizing the workload on researchers.
Key concepts
- Tiered Approach
- The search used a multi-step strategy: starting with an initial data cut, moving to machine learning and ensemble searches, followed by citizen science checks and finally expert grading. This layered method ensures that rare lenses are found efficiently without overwhelming the inspection process.
- Citizen Science Inspection
- This involved two stages where volunteer citizens inspected a subset of objects using tools like SW-Classify and SW-Refine. This step was designed to further refine the lens sample, aiming for greater completeness than machine learning alone could achieve.
- Expert Grading Calibration
- To reduce subjectivity in expert grading, the method calculates preliminary grades alongside an assessment of the inspector's skill level. This calibration process uses spectroscopically confirmed systems to produce probabilities instead of subjective grades.
- Lens Modeling (ECLIPSE-C)
- Lens modeling is used to determine a candidate's true nature. This involves using codes like ECLIPSE-C to model the gravitational effects of a potential lens system, providing objective verification beyond simple visual inspection.
Terminology used across episodes
This episode discusses
- Euclid Data Release 1 (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) B: Discovery pipeline overview · Paper Radio
- Euclid Quick Data Release (Q1). The Strong Lensing Discovery Engine F -- Bright and low-redshift strong lenses
- Euclid Quick Data Release (Q1). Refining the strong lens sample with Space Warps crowd-sourced science · Paper Radio
- Euclid Quick Data Release (Q1). AstroVink: A vision transformer approach to find strong gravitational lens systems
- Cosmology with the line-of-sight shear of strong gravitational lenses · Paper Radio
- DINOv3
- Wide-Field InfrarRed Survey Telescope-Astrophysics Focused Telescope Assets WFIRST-AFTA 2015 Report
- Galaxy Zoo Evo: 1 million human-annotated images of galaxies
The paper
Euclid Data Release 1 (DR1). Euclid Lens Identification Pipeline and Scientific Exploration (ECLIPSE) B: Discovery pipeline overview · Read on arXiv
P. Holloway, A. Verma, N. E. P. Lines, T. Li, S. H. Vincken, J. A., Acevedo Barroso, T. E., Collett, S., Schuldt, M., Fogliardi, A., Manjón-García, J. Wilde
ESO
We present the method used to identify the 15,000 high-grade galaxy-galaxy strong lens candidates in Euclid Data Release 1 (DR1). Given that lenses are intrinsically rare systems and the DR1 data set is vast, this lens search was conducted in a number of stages: a catalogue-level cut; a machine-learning and ensemble search; a two-stage citizen-science search, and finally grading of lens candidates by strong lens researchers. In this paper, we describe each search stage and their respective results. This search was enabled by the automated classification of over 68 million objects by neural networks, more than 4.5 million classifications of 300,000 objects by citizen-scientist volunteers, and the inspection of over 30,000 objects by strong lens researchers. We find this multi-stage approach helps maximise the number of lenses identified, while minimising inspection requirements of both citizen scientists and strong lens researchers. We measure a completeness in our comprehensive search of 0.73-0.15+0.08 to grade A ('definite') systems and 0.34-0.06+0.05 to grade A+B ('definite' + 'probable') systems. Our search therefore identifies the majority of the definitive strong lens systems in the DR1 area with the remainder likely to be fainter or more ambiguous cases. We find that researcher grading remains subjective; in this work we implement calibration to account for biases of individual inspectors and highlight the benefit of lens modelling (ECLIPSE-C) in determining the nature of a lens candidate. This paper forms part of the ECLIPSE series, here focussing on the discovery pipeline used to identify lens candidates in Euclid DR1; an overview of the lenses discovered is presented in ECLIPSE-A.
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) B".
Jocelyn: Here's the gist of the paper. The paper presents a multi-stage method developed to identify approximately 15,000 high-grade galaxy–galaxy strong lens candidates within Euclid Data Release 1 (DR1).
Vera: Recognizing that lenses are intrinsically rare and the DR1 dataset is extensive, the search employed a tiered approach: an initial catalogue-level cut, followed by a machine-learning and ensemble search, a two-stage citizen-science inspection process, and finally, grading of lens candidates by expert strong lens researchers.
Jocelyn: This pipeline was supported by automated classification of over 68 million objects via neural networks, more than 4.5 million classifications performed by citizen-scientist volunteers across 300,000 objects, and subsequent inspection of over 30,000 objects by professional lens researchers.
Vera: The authors assert that this multi-stage strategy successfully maximizes the number of identified lenses while minimizing the inspection burden placed on both citizen scientists and expert researchers.
Title and authors: Vera: So, this paper is about how they found about fifteen thousand high grade galaxy galaxy strong lens candidates in Euclid Data Release one DR1 <ref:2611.15886#pg1,high grade galaxy galaxy strong lens candidates in Euclid Data Release 1>. It talks about their whole pipeline for doing that search.
Jocelyn: Yeah, and it's interesting because lenses are rare things to find, so the DR1 data set is really big, and they used a tiered approach to handle all of that volume.
Subrahmanyan: I'm curious about the scale of this search; how big is the area they were looking at in total?
Vera: They conducted three separate spatial stages for this search: a two hundred square degree search, then a fifty square degree search, and finally, a more comprehensive survey of whatever data was left over.
Jocelyn: And they designed this whole pipeline to specifically tune it so that the systems that get looked at by experts are as pure as possible without losing any of the truly high grade lenses.
Subrahmanyan: That tuning is key because you need those high purity results if you're going to make any real cosmological statements about structure formation or dark matter distribution.
Vera: They achieved this using a bunch of tools: they automated the classification of over sixty-eight million objects with neural networks, and then involved citizen scientists in classifying more than four point five million objects out of three hundred thousand.
Jocelyn: That’s a lot of work for the volunteers, but it shows how they are trying to spread the inspection burden around so that experts only look at the most promising candidates.
Subrahmanyan: So, if we look at the results reported in this paper about completeness, what does that actually mean for us in terms of finding these lenses?
Vera: For grade A definite systems, they report a completeness metric of zero point seven three plus or minus zero point zero eight minus zero point fifteen. That means seventy three percent of the truly definite high grade systems were found across the full DR1 area.
Jocelyn: And for the slightly less certain systems, they have a completeness metric of zero point three four plus or minus zero point zero five minus zero point six for grade A plus B systems.
Subrahmanyan: So, seventy five hundred grade A systems identified across the whole DR1 area represents what they call seventy five percent completeness. That’s a solid number to work with when thinking about how many of these massive structures we can actually constrain cosmologically.
Title and authors: Vera: And they acknowledge that grading by experts is still somewhat subjective, but they say they implemented calibration techniques to account for individual inspector biases and highlighted the utility of lens modeling tools like ECLIPSE C in figuring out what a candidate truly is.
Jocelyn: They also used machine learning classifiers that were retrained on larger samples, including lenses that were already present within the DR1 area, which they say makes the results stronger than previous searches like Q1.
Subrahmanyan: That retraining sounds important because it suggests they are trying to make those initial machine learning classifications more robust and less dependent on some specific subset of data points.
Vera: They also mention that this approach is scalable for future Euclid data releases, like the second release or the full Euclid Wide Survey, based on how much better those machine learning models get.
Jocelyn: And they detail a few improvements they made to the pipeline itself, which we'll get into next—things they did to make it even better than just what's written in this overview.
Subrahmanyan: Before we look at those improvements, I just want to ask about the initial challenge mentioned: how did they decide on the search areas of two hundred square degrees and fifty square degrees versus a different setup?
Vera: They tuned the pipeline specifically to maximize purity for systems inspected by experts without sacrificing the completeness of high grade lenses, which dictated those spatial stages.
Jocelyn: And they structured the citizen science inspection into two parts, a binary classification stage called SW Classify and then a second refinement stage called SW Refine.
Subrahmanyan: I'm interested in that second refinement stage; what kind of extra filtering did that add to the sample compared to just using the initial ML classification?
Vera: The paper suggests that this two-stage citizen science search enables a greater completeness than what you could get from just an ML-only classification.
Jocelyn: That makes sense because having those volunteers do two checks, one of which is a refinement stage, should help weed out more false positives before they get to the experts.
Subrahmanyan: If that extra filtering helps purity, it means the final set of systems handed over to the strong lens researchers is likely much cleaner for cosmological analysis.
Vera: They also focused on improving those machine learning classifiers by using ensembles of networks, specifically combining networks one and five as described in section two point three point three.
Jocelyn: Combining different network architectures is a common way to boost performance, but they mention that this combination was tested against a test data set to see if it actually improved the purity.
Title and authors: Subrahmanyan: So, the idea is that by using these combined networks, we might get better purity scores for those high grade systems than either network could achieve alone.
Vera: They also propose calibrating the researcher grades by accounting for individual inspector biases through a method that calculates preliminary grades and expert skill at the same time.
Jocelyn: That sounds like they are trying to use the data itself to adjust for human error in grading, which is a smart way to mitigate subjectivity.
Subrahmanyan: And they suggest this calibration involves comparing those grades against spectroscopically confirmed systems from four point five thousand lens surveys like 4SLSLS to produce probabilities instead of just a grade.
Vera: That shifts the focus away from subjective grading and toward something more grounded in observation, which I think is a solid direction for this research.
Jocelyn: They also propose integrating forward modelling codes or machine learning methods for rapid lens modelling to provide additional verification as to the nature of a possible lens system beyond simple visual inspection.
Subrahmanyan: So, if we can model the lens system quickly, that would give us a much stronger check on whether it’s actually a real galaxy galaxy lens rather than some chance alignment.
Vera: They are showing how this entire process fits together for the Euclid Data Release one DR1 Lens Identification Pipeline and Scientific Exploration ECLIPSE B: Discovery pipeline overview <ref:2611.15886#pg1,Lens Identification Pipeline and Scientific Exploration ECLIPSE B: Discovery pipeline overview>.
Jocelyn: It really shows a layered approach, starting with broad automated classification and ending with expert human verification, all tuned to maximize the yield of high grade systems.
Subrahmanyan: Overall, this work lays out a very detailed roadmap for how we can efficiently extract cosmological constraints from massive survey data like Euclid DR1.
Vera: So to wrap up the main points: they used a multi stage search—catalogue cut, machine learning, citizen science inspection, and expert grading—to find about fifteen thousand high grade lens candidates.
Jocelyn: They achieved this by tuning the pipeline to maximize purity for experts while keeping completeness high for the best systems.
Subrahmanyan: And they found that using ensemble machine learning networks and calibrated researcher grades helps make those final results more reliable, moving toward probabilities rather than just grades.
Vera: This whole ECLIPSE B discovery pipeline overview is designed to be scalable for future Euclid data releases, which is a big deal for planning the next few years of observational cosmology.
The paper's summary: Vera: So, this paper lays out the whole plan for finding about fifteen thousand high grade galaxy galaxy strong lens candidates in Euclid Data Release One DR1.
Jocelyn: It's a really detailed pipeline they developed, moving from initial automated cuts to citizen science and then expert grading.
Subrahmanyan: It sounds like they're trying to get the best possible sample for cosmological studies without drowning in data volume.
Vera: Exactly, they use this multi-stage search—the initial cut, the machine learning look, the two stages of citizen science inspection, and then a final grade by lens researchers.
Jocelyn: The paper emphasizes that this tiered approach is designed to maximize the number of lenses found while minimizing how much work those volunteers and experts have to do.
Subrahmanyan: That efficiency is crucial because we're dealing with such massive datasets from Euclid; we need smart ways to sift through it without burning human resources on every single object.
Vera: They show the search happens in three spatial stages, starting with a two hundred square degree search, then a fifty square degree search, and finally a big survey of whatever data is left over.
Jocelyn: And they tuned the whole pipeline to make sure that when those experts look at their candidates, the purity is really high without sacrificing the completeness of those genuinely high grade systems.
Subrahmanyan: High purity in those systems means we have cleaner signals for our models, which is what we need when we're trying to figure out how dark matter structures are distributed across the universe.
Vera: They report some completeness numbers, like seventy three percent for the definite grade A systems and thirty four percent for the grade A plus B systems in that comprehensive search area.
Jocelyn: And they stress that grading by experts is still subjective, so they did work to calibrate those grades using methods that account for individual inspector biases.
Subrahmanyan: Calibrating against spectroscopically confirmed systems from surveys like 4SLSLS to produce probabilities instead of just a grade sounds like a really solid way to bring objectivity into the process.
Vera: They also integrated machine learning classifiers that they retrained on bigger samples, which they say makes the initial classifications more robust than previous searches.
Jocelyn: It's good that they are making this pipeline scalable for future data releases, like Euclid Data Release Two or even the whole Euclid Wide Survey.
Subrahmanyan: So, if we look at what this means for us as a community studying cosmology right now, it's about having a more reliable and repeatable way to identify these rare lenses in the next big data release.
Vera: This whole ECLIPSE B discovery pipeline overview is really showing how they built a system that starts broad and ends with expert verification, all tuned to find those high grade systems efficiently.
Jocelyn: It's a layered approach that seems designed specifically for the Euclid data, making sure we get the best lenses out of what we have.
Subrahmanyan: So, this paper lays out a detailed roadmap for how to extract meaningful cosmological constraints from massive survey data like Euclid DR1.
Vera: The next thing to look at is those specific technical improvements they made—the things they added on top of the original plan—to make it even better than what's on page one.
The paper's improvements: Tom: So, we're looking at how they plan to get even better at finding those lenses in future data releases, specifically focusing on the improvements they suggest for this pipeline.
Vera: They aren't just happy with the initial setup; they're proposing specific ways to upgrade the machine learning part and how we handle those expert grades.
Jocelyn: It seems like they are trying to address some of those weaknesses we talked about earlier, like the potential for bias in human grading.
Subrahmanyan: I'm interested in what "using ensembles of networks" actually means for the performance on finding these rare objects.
Vera: They suggest combining specific neural network architectures, specifically networks one and five, to boost the purity of the results they get from their classifiers.
Jocelyn: That makes sense; if you run two different types of models and combine their outputs, you might get a more stable classification than just relying on one model alone.
Subrahmanyan: So, that's about improving the quality of the initial selection before we even get to the human experts grading things.
Vera: Right, and they also propose calibrating those expert grades by accounting for how much each inspector might be biased or skilled at their job.
Jocelyn: They suggest calculating preliminary grades and expert skill at the same time, using confirmed systems from surveys like 4SLSLS to get probabilities instead of just a simple grade.
Subrahmanyan: That shift toward probabilities instead of just subjective grades is a big step for our theoretical work because it grounds the identification in something observational.
Vera: And they also propose adding rapid lens modeling, either with forward-modeling codes or machine learning methods, to verify what kind of lens system it really is beyond just looking at the image.
Jocelyn: That quick modeling would give us that additional check we talked about earlier to make sure it's actually a real galaxy galaxy lens and not some chance alignment in the data.
Subrahmanyan: If we can model them quickly, that verification would be essential for building more robust models of structure formation across different cosmic epochs.
Vera: So these improvements are all about making the pipeline smarter by using better AI combinations and better methods to handle human judgment.
Jocelyn: It sounds like they're really focused on making sure the final sample that gets handed over for follow-up is as clean as possible.
Subrahmanyan: This work, focusing on refining the discovery pipeline, is important because it sets a standard for how we can efficiently process these huge surveys in the future.
Vera: So they're showing a clear path toward making this system more robust and scalable for even larger datasets down the road.
Jocelyn: It’s about turning a complex search into a highly efficient, reliable tool for finding those rare lenses in Euclid data.
Conclusion: Tom: We're wrapping up our look at the Euclid Data Release One DR1 Lens Identification Pipeline and Scientific Exploration ECLIPSE B discovery pipeline overview paper, which details how they find about fifteen thousand high grade lens candidates using a multi-stage method.
Vera: Basically, they used this tiered approach—from initial automated cuts to citizen science inspections and expert grading—to get a good sample of those rare lenses.
Jocelyn: It's really about showing that you can sift through massive amounts of data efficiently while keeping the most scientifically interesting systems.
Subrahmanyan: For cosmology, having a reliable way to identify these high grade systems is important because it gives us cleaner data for structure formation studies.
Vera: They achieved this by tuning the pipeline to maximize purity for experts while trying not to lose completeness on the best candidates.
Jocelyn: And they showed that using ensemble machine learning networks and calibrated expert grades helps make those final results more dependable, moving toward probabilities instead of just subjective grades.
Subrahmanyan: That shift toward probabilities is key because it moves us closer to making predictions about the underlying physics of dark matter distribution.
Vera: So this whole ECLIPSE B discovery pipeline overview is designed to be scalable for future data releases, which is a big deal for planning the next few years of observational cosmology.
Jocelyn: It really shows a layered approach that starts broad and ends with expert verification, all tuned to find those high grade systems efficiently.
Subrahmanyan: Overall, this work lays out a detailed roadmap for how to extract meaningful cosmological constraints from massive survey data like Euclid DR1.
Vera: So the main implication is that we have a more robust way to handle the sheer volume of data coming from Euclid and identify those crucial lenses reliably.
Jocelyn: It’s a great blueprint for how we can build these discovery pipelines for future surveys, making sure our follow-up resources are used wisely.
Subrahmanyan: I just think as long as we keep refining these identification methods, the cosmological constraints on structure growth will get much tighter.
Vera: We'll be watching those improvements to the machine learning and calibration methods closely for the next release of Euclid data.
Jocelyn: And we’ll keep looking at those citizen science refinement stages to see how much extra purity they add before it hits the expert reviewers.
Subrahmanyan: It’s a solid piece of work that gives us confidence in what we can expect from the Euclid observations on cosmic structure.
More episodes
- 2607.28919-Strongly Magnetized Super-Eddington Accretion: How Spin and Accretion Rate Regulate Energy Output and Mass Loss
- 2607.08835-Hybrid Simulations of Proton Acceleration at Oblique High- beta Shocks
- 2606.29552-Improving the Precision of Line-by-Line Radial Velocities: A Data-Driven Iterative Algorithm for Spectral Line Selection
- 2609.05312-Evidence for LP 890-9d via Transit Timing Variations
- 2609.10319-Extending the Little Red Dot population at intermediate redshift with VIPERS
- 2611.14233-Probing for anisotropies at different angular scales in the arrival directions of the cosmic rays detected at the Pierre Auger Observatory
- 2610.21722-Hydrostatic Equilibrium Estimate of the Fuzzy Dark Matter Mass from Dwarf Spheroidals: Method and Systematics
- 2610.40550-An Empirically Calibrated [C II] Luminosity-Halo Mass Relation for Line Intensity Mapping
- 2607.27328-PANTERA. I. Hunting for the Most Metal-Rich Stars in the Solar Neighborhood with High-Resolution Spectroscopy
- 2607.09850-Low-hard to high-soft spectral state transitions in the faintest early-X-ray-detected optical tidal disruption event TDE 2025aarm