Galaxy Zoo: Cosmic Dawn (GZCD) Survey

• Galaxy Zoo: Cosmic Dawn (GZCD) is a deep morphological survey that combines ultra-deep HSC imaging, citizen science, and machine learning to classify galaxies.
• It employs an active decision tree and the Zoobot model to accelerate the classification of 47,347 subjects, yielding efficient triage and high-confidence results.
• The project uncovers rare objects like strong gravitational lenses, providing a valuable training dataset for future deep-learning applications in cosmic-dawn research.

Galaxy Zoo: Cosmic Dawn (GZCD) is a Galaxy Zoo morphological-classification program built on ultra-deep Hyper Suprime-Cam imaging in the Euclid Deep Field North (EDFN) from the Hawaii Twenty Square Degree (H20) survey, itself part of the wider Cosmic DAWN survey. In its published form, GZCD provides morphological classifications of over 41,000 galaxies out to $z_{\rm phot}\sim2.5$ across six square degrees of the EDFN, with tens of thousands of volunteers and the deep learning foundation model Zoobot jointly classifying 47,347 subjects down to $m_{\rm HSC-i}=21.5$. The project combines a Galaxy Zoo-style decision tree, active learning, and public data release, and it has also yielded 51 newly discovered gravitational lenses. Within the Euclid/DAWN ecosystem, GZCD functions both as a scientific morphology catalog and as a truth set for machine-learning applications in deep, seeing-limited survey data (Pearson et al., 26 Sep 2025, Collaboration et al., 2024).

1. Survey setting, field selection, and sample construction

GZCD is embedded in a larger survey architecture. The Cosmic DAWN survey complements Euclid deep data with matched-depth multiwavelength imaging and spectroscopy in the UV--IR to provide consistently processed Euclid-selected photometric catalogs, accurate photometric redshifts, and measurements of galaxy properties to a redshift of $z\sim10$. DAWN covers the three Euclid Deep Fields and six Euclid Auxiliary Fields over $\sim59$ deg$^2$, and it was designed to supply the external data backbone needed for high-redshift galaxy evolution studies in the Euclid era (Collaboration et al., 2024).

Within that infrastructure, GZCD uses the first six square degrees of the EDFN observed at full depth in H20. The H20 Hyper Suprime-Cam imaging reaches 5$\sigma$ depths of about 27 AB mag, has 0.168 arcsec pixels, and median seeing of 0.67 arcsec. The cutouts are drawn from ultra-deep $gri$ HSC images, and the project restricts the sample to galaxies with effective radii $r_{\rm eff} \geq 0.6$ arcsec so that the sources are large enough to classify morphologically at the HSC seeing scale (Pearson et al., 26 Sep 2025).

The source-selection sequence is explicit. The initial extracted source set contained about 350,000 objects down to $m_{\rm HSC-i}=25$. The volunteer-facing sample was then restricted to $m_{\rm HSC-i}<21.5$, a cut chosen to maximize engagement and ensure that the decision-tree questions remained meaningful. The final Galaxy Zoo subject set contained 47,347 subjects. For analysis and release, the focus is on 45,742 subjects with either $m_{\rm HSC-i}=21.5$0 or failed photometric-redshift estimates, corresponding to over 41,000 cutouts with central galaxies (Pearson et al., 26 Sep 2025).

A recurrent point of clarification is that the current GZCD release is not primarily a very-high-redshift morphology census. The photometric-redshift distribution has a median of $m_{\rm HSC-i}=21.5$1, with 95.4 per cent of sources at $m_{\rm HSC-i}=21.5$2, although the sample has a tail reaching $m_{\rm HSC-i}=21.5$3. The “Cosmic Dawn” designation therefore reflects the survey framework and science trajectory more than the redshift distribution of the present release (Pearson et al., 26 Sep 2025).

2. Morphological decision tree and annotation schema

The GZCD classification workflow adapts the Galaxy Zoo decision-tree paradigm to deep HSC imaging. Volunteers classify one subject at a time, beginning with a top-level assessment of whether the target is smooth, featured/disk-like, or a star/artifact/problem. Subsequent branches address disk inclination, bulges, bars, spiral arms, arm winding, merging or disturbance, and clumps, while a dedicated rare-features branch includes lenses, rings, and irregulars (Pearson et al., 26 Sep 2025).

Several modifications distinguish the GZCD tree from earlier Galaxy Zoo implementations. Because the depth of the imaging sometimes caused the H20 pipeline to merge nearby sources into poor cutouts, the older “Star or Artifact” response was replaced by “Star, Artifact, or Bad Zoom,” followed by a discrimination between star, non-star artifact, and bad zoom. The artifact branch was expanded to include bleed trails, diffraction spikes, satellite trails, cosmic rays, scattered light, ghosts, and other non-galaxy features. A new clump question, “Are there any obvious bright clumps?”, was introduced to better capture clumpy systems. Because strong lenses were a science priority, “Lens or arc” was moved to the top of the rare-features list (Pearson et al., 26 Sep 2025).

This schema reflects the specific observational regime of the project. The data are multiband, deep, and seeing-limited rather than space-based, which changes the failure modes, confusion patterns, and visible structure compared with earlier HST-centered morphology projects. The resulting labels are therefore tuned to the actual appearance of galaxies in HSC cutouts rather than to an idealized intrinsic-morphology taxonomy (Pearson et al., 26 Sep 2025).

3. Human--machine active learning with Zoobot

A central methodological feature of GZCD is the integration of Zoobot into an active-learning cycle rather than using it only as a post hoc classifier. The first 16,671 subjects were fully classified by at least 40 volunteers each without machine learning. Those volunteer labels were then used to fine-tune Zoobot, which in this phase used an EfficientNetB0 architecture trained from prior Galaxy Zoo data. The model predicted the fractions of volunteers expected to choose each response as Dirichlet distributions, returning a mean predicted vote fraction together with 90 per cent confidence intervals (Pearson et al., 26 Sep 2025).

The second subset contained 30,676 JPEG subjects. For these, Zoobot acted as a gatekeeper: a subject was retired if the model predicted with 90 per cent confidence that fewer than 20 per cent of volunteers would choose “Features or Disk” on the first question. Subjects passing that criterion were shown to volunteers, and the accumulated volunteer labels were used each week to retrain Zoobot before the next pass through the queue (Pearson et al., 26 Sep 2025).

The operational outcome was substantial triage of easy cases. Zoobot retired 77.8 per cent of galaxies early, including 528 subjects that were retired before any volunteer saw them. Among the subjects retired after fewer than five volunteer votes, 62 per cent were smooth galaxies by volunteer consensus; for the 528 Zoobot-only subjects, the corresponding figure was 95 per cent. The project reports several speed-up estimates: the second subset was retired in 83 days compared to 109 days for the first subset, implying an average classification-rate increase of about a factor of 2.4; estimates based on the mean number of volunteer classifications before retirement yield a factor of about 2.9; and for the most strongly retired objects, where the mean was only 6 volunteer classifications, the speed-up was roughly a factor of 7 (Pearson et al., 26 Sep 2025).

The scientific and procedural significance lies in the division of labor. Zoobot handled the smooth, high-confidence cases, while volunteers were preferentially presented with the more difficult, unusual, or visually interesting systems. The project explicitly treats this as both a performance improvement and an improvement in volunteer experience (Pearson et al., 26 Sep 2025).

4. Catalog outputs, vote fractions, and public data products

The morphology outputs show the broad class balance expected for deep, seeing-limited imaging. For the 45,742-subject analysis sample, majority-vote classifications among sources with photometric redshifts identify 23,533 smooth galaxies, 2,920 featured-or-disk galaxies, and 2,346 problem objects. Including objects without photometric redshifts, the corresponding totals are 28,605 smooth, 3,201 featured-or-disk, and 4,123 problem objects, with 9,297 less certain. The paper summarizes this as implying at least 41,000 subjects with central galaxies (Pearson et al., 26 Sep 2025).

Two standard Galaxy Zoo post-processing steps were explicitly not applied. The project defines the classic volunteer-weighting scheme as

$m_{\rm HSC-i}=21.5$4

where $m_{\rm HSC-i}=21.5$5 is the mean consistency of a volunteer’s votes relative to others, but weighting was not used because only 0.8 per cent of logged-in volunteers would have had weights below 1, and because the revised “Problem” branch removed the earlier incentive for low-quality “star or artifact” clicking. Redshift debiasing was also not applied, because the sample spans a much wider redshift range than most prior Galaxy Zoo datasets and because photometric redshifts are missing for about one-fifth of the sources. As released, the vote fractions therefore represent visual appearance in the HSC images rather than an attempt at intrinsic morphology (Pearson et al., 26 Sep 2025).

The public release is extensive. It includes aggregated volunteer classifications, Zoobot predictions, the image cutouts used in the project, and associated metadata for each subject. The metadata include the H20 ID, object name, effective radius or major-axis proxy, $m_{\rm HSC-i}=21.5$6, $m_{\rm HSC-i}=21.5$7, and $m_{\rm HSC-i}=21.5$8 magnitudes, photometric redshift, sky coordinates, Talk-board tags, and API links to external databases and the H20 Image Viewer. The release is divided into 15,745 PNG subjects fully classified before active learning and 29,997 JPEG subjects classified with active learning. The volunteer catalog provides counts and vote fractions for each decision-tree response together with the number of volunteers per subject; the Zoobot catalog provides predicted vote fractions and 90 per cent upper and lower confidence bounds, but not for the rare-features question, which Zoobot was not trained to predict. “Leaf-fraction” versions are also provided, masking downstream branches unless earlier branching probabilities make them relevant (Pearson et al., 26 Sep 2025).

The paper also gives approximate selection cuts for constructing science samples:

Sample	Approximate cut
Smooth galaxies	$m_{\rm HSC-i}=21.5$9
Featured/disk galaxies	$z\sim10$0
Edge-on disks	$z\sim10$1 and $z\sim10$2
Spirals	$z\sim10$3, $z\sim10$4, and $z\sim10$5
Clumpy galaxies	$z\sim10$6 and $z\sim10$7
Lenses	$z\sim10$8 and $z\sim10$9

These cuts are presented as approximate rather than definitive; the authors stress that purity and completeness depend on the science goal (Pearson et al., 26 Sep 2025).

5. Rare objects and strong gravitational lenses

Rare-object detection is a major scientific output of GZCD. Lens candidates were flagged through three routes: the main workflow, Talk-board tags, and a separate Galaxy Zoo Mobile lens-search workflow that targeted subjects retired early by Zoobot. The project reports 122 unique lens candidates before expert vetting, with candidates identified via 32 main-workflow cases, 56 Talk-board tags plus 5 ring-tags, and 63 Galaxy Zoo Mobile cases (Pearson et al., 26 Sep 2025).

Expert inspection using HSC and Euclid imaging reduced these to 57 confirmed lenses: 15 new grade-A lenses, 9 grade-B, 27 grade-C, and 6 lenses already known to Euclid. Of these, 51 are newly discovered lenses in GZCD. The paper emphasizes that the multiplicity of discovery channels was essential: Talk-board tagging was more complete, while the mobile workflow found the largest number of candidates missed by other methods (Pearson et al., 26 Sep 2025).

Beyond lenses, the catalog includes edge-on disks, major disturbances, clumpy galaxies, rings, and other structured systems. This breadth matters because GZCD was designed not only to recover common morphologies but also to surface the rare and ambiguous systems that are especially valuable for follow-up and for training automated classifiers under realistic survey conditions (Pearson et al., 26 Sep 2025).

6. Position within cosmic-dawn research and survey-era morphology

The immediate value of GZCD is morphological, but its significance is shaped by the broader cosmic-dawn survey landscape. DAWN was designed to combine Euclid deep imaging with matched UV--IR coverage, spectroscopy, accurate photometric redshifts, and galaxy-property measurements out to $\sim59$0, while also supporting science cases that include UV luminosity functions, the origins of large-scale structure, the first quenched galaxies, and clustering-based links between galaxies and dark matter halos (Collaboration et al., 2024). In that setting, GZCD provides a morphology layer over one of the best-instrumented Euclid deep environments.

The project’s machine-learning role is equally explicit. The GZCD paper argues that the dataset is especially valuable as a truth set for future deep-learning work, particularly for ground-based surveys such as Rubin/LSST, because it spans deep, seeing-limited multiband HSC imaging that is closer in character to Rubin than to space-based morphology sets. This makes the catalog relevant both for direct scientific selection and for supervised training in a regime that includes common galaxies, stars and artifacts for rejection, and rare classes such as mergers, rings, clumps, and strong lenses (Pearson et al., 26 Sep 2025).

The surrounding literature clarifies why such a morphology resource matters. Clustering analyses of JWST Lyman Break Galaxies indicate rapid evolution in the galaxy--halo relation during cosmic dawn, with typical host halo masses declining from $\sim59$1 at $\sim59$2 to $\sim59$3 at $\sim59$4, alongside increasing effective bias (Dalmasso et al., 5 Jan 2026). FIRE-2 simulations show that ordinary, feedback-driven bursty star formation can reproduce the observed abundance of UV-bright galaxies at $\sim59$5 without invoking a non-standard cosmology, a top-heavy IMF, or strongly enhanced star-formation efficiency (Sun et al., 2023). The Phoebos simulation interprets early galaxies as compact, clumpy, rapidly growing systems in a weak-feedback regime and presents itself as an interpretive framework for JWST and GZCD morphology work (Donkelaar et al., 7 Jul 2025). In parallel, studies of dust at $\sim59$6 show that the ultraviolet extinction bump in JADES-GS-z6-0 is not explained by small graphite grains, implying that early ISM chemistry cannot simply be mapped from local Milky Way dust models (Li et al., 12 Feb 2025). Bayesian 21-cm and UV-luminosity-function inference likewise shows that combining observables is necessary to break degeneracies in the astrophysics of reionization and cosmic dawn (Park et al., 2018).

Taken together, these results suggest a division of labor across datasets and methods. Wide-field morphology catalogs such as GZCD do not themselves solve the galaxy--halo, feedback, dust, or reionization problems, but they provide structured labels, rare-object identification, and training data that can be linked to photometric-redshift catalogs, spectroscopic follow-up, Euclid imaging, and eventually higher-redshift samples. A plausible implication is that GZCD’s long-term value will be largest when its classifications are joined to the DAWN photometric infrastructure and to physical models that already predict compactness, clumpiness, high specific star-formation rates, and environmentally biased growth in the early Universe.

A final point follows from the released vote fractions. Because GZCD does not apply redshift debiasing and because most of the current catalog lies at $\sim59$7, its labels should be interpreted first as measurements of appearance in ultra-deep HSC data. That is not a weakness but a definition of scope: the catalog is simultaneously a morphology dataset, an active-learning demonstration, a strong-lens discovery engine, and a survey-era training resource for future work that will extend more directly into the redshift regime conventionally denoted as cosmic dawn (Pearson et al., 26 Sep 2025).