arxiv: 2604.14599 · v1 · submitted 2026-04-16 · 🌌 astro-ph.GA

Recognition: unknown

Disk-like galaxies at 4 < z < 7.7 : JWST/NIRCam morphologies revealed by denoising VAE-GCNN classification

S.S. Mirzoyan , A. Avagyan , V.G. Gurzadyan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 10:47 UTC · model grok-4.3

classification 🌌 astro-ph.GA

keywords disk-like galaxieshigh-redshift galaxiesJWST/NIRCamgalaxy morphologyvariational autoencodergraph convolutional neural networkearly universe structure formation

0 comments

The pith

A VAE-GCNN pipeline on JWST/NIRCam images measures a 0.34 disk-like fraction among 100 galaxies at redshifts 4 to 7.7.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors train a U-Net variational autoencoder to clean JWST galaxy images of noise and contaminants while keeping their shapes intact, then feed the results into a rotation- and reflection-equivariant graph convolutional neural network that labels each galaxy as disk-like or not. They apply this identical pipeline across the full sample to produce a uniform disk fraction of 0.34 over 4 < z < 7.7, with the value also reported as a function of galaxy mass. The measurement addresses whether rotationally supported disks could assemble within the first billion years after the Big Bang, a period when such structures were previously thought to be rare.

Core claim

Applying the same denoising VAE-GCNN classification pipeline to multi-filter JWST/NIRCam cutouts, the authors determine the fraction of disk-like galaxies as 0.34 for a sample of 100 galaxies over the redshift range 4 < z < 7.7, also in dependence on the galaxy mass range.

What carries the argument

The denoising VAE-GCNN pipeline, in which a U-Net variational autoencoder removes astrophysical and instrumental contaminants while preserving intrinsic morphology and an equivariant graph convolutional neural network then classifies the cleaned cutouts as disk-like or non-disk.

If this is right

Disk-like galaxies form a substantial fraction of the population already at redshifts greater than 4.
Angular-momentum-supported structures must have assembled within the first billion years of cosmic time.
The disk fraction shows dependence on galaxy mass within the sampled range.
Homogeneous morphology-based studies at these epochs are required to test models of early disk formation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the fraction holds in larger samples, it would indicate that efficient mechanisms for acquiring angular momentum were already operating at these early epochs.
Repeating the pipeline on deeper or wider JWST fields could reveal whether the disk fraction changes with redshift inside the 4 < z < 7.7 interval.
Direct comparison of the classified sample against predictions from cosmological hydrodynamical simulations would test whether current models reproduce the observed 0.34 value.

Load-bearing premise

The variational autoencoder removes contaminants without altering or fabricating disk-like features in the galaxies, and the graph convolutional classifier trained on the cleaned images applies without bias to the high-redshift JWST observations.

What would settle it

Independent visual classification or an alternative morphological method applied to the identical 100 cutouts that returns a disk fraction differing by more than 0.1 from 0.34 would falsify the reported value.

Figures

Figures reproduced from arXiv: 2604.14599 by A. Avagyan, S.S. Mirzoyan, V.G. Gurzadyan.

**Figure 1.** Figure 1: Redshift distribution of the considered JWST galaxy sample. 2 DATASET Our training strategy follows Papers I (Mirzoyan 2025) and II (Mirzoyan 2026), cf. (Mirzoyan, et al 2019). The EFIGI catalog (Baillard et al. 2011) provides, visually inspected galaxies used to train and validate the U–Net VAE denoiser, while Galaxy10 DECaLS (Willett et al. 2013; Walmsley et al. 2022) supplies ten-class morphological lab… view at source ↗

**Figure 3.** Figure 3: At the lowest redshift bin, the disk fraction is fdisk = 0.296+0.189 −0.137 with N = 27 sources. The fraction rises to 0.407+0.186 −0.162 in the [4.5, 5.0) bin, again with N = 27, before declining at z ∼ 5 to 0.250+0.219 −0.138 for [5.0, 5.5) (with N = 20). A local minimum occurs in the [5.5, 6.0) interval, where only 1 of 11 galaxies (fdisk = 0.091) is classified as disk-like, albeit with large uncertain… view at source ↗

**Figure 4.** Figure 4: Disk fraction versus galaxy mass bin with Wilson 95% confidence intervals [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

read the original abstract

Understanding the prevalence of disk-like galaxies at very high redshifts is crucial for constraining the early formation of angular momentum-supported structures. The advent of JWST now permits rest-frame UV and optical morphological studies deep into cosmic epochs where disks have traditionally been considered uncommon. We apply an identical denoising VAE-GCNN classification pipeline to multi-filter JWST/NIRCam cutouts in order to obtain homogeneous, morphology-based disk fractions across the sample. Our approach comprises two steps: (i) a U-Net Variational Autoencoder (VAE) is trained to remove astrophysical and instrumental contaminants while preserving intrinsic morphology, and (ii) a rotation - and reflection - equivariant GCNN classifier is applied to the denoised cutouts to distinguish disk-like galaxies from non-disks. We determine the fraction of disk-like galaxies as 0.34 for a sample of JWST 100 galaxies over the redshift range 4 < z < 7.7, also in dependence on the galaxy mass range. Our GCNN-based morphological analysis indicates that disk-like systems constitute a significant fraction of the considered high-redshift population and underscore the importance of such studies for the models of disk formation in the first billion years.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper runs an existing VAE denoising plus equivariant GCNN pipeline on 100 JWST galaxies at 4<z<7.7 and reports a 0.34 disk fraction, but supplies no validation or uncertainty estimates for the classifier at these redshifts.

read the letter

The main takeaway is that Mirzoyan et al. applied a U-Net VAE for contaminant removal followed by a rotation- and reflection-equivariant GCNN to classify JWST/NIRCam cutouts, yielding a 0.34 fraction of disk-like galaxies across their sample of 100 objects, with some dependence on mass. They emphasize the homogeneous pipeline across filters and redshifts as a way to get consistent morphology statistics where disks were once thought rare. That uniform treatment is the clearest practical step forward here; it sidesteps the usual mix of different classifiers and selection effects that plague high-z morphology work. The choice of an equivariant network also makes sense for avoiding orientation biases without extra augmentation. The soft spot is exactly what the stress test flags: the abstract states the 0.34 number but gives no accuracy figures, confusion matrix, recovery tests on injected disks, or comparison against raw versus denoised classifications. There is also no reported check on whether the VAE step preserves or suppresses the flatness and arm features that define disks, and no error bars or redshift-dependent bias tests. Without those, the fraction cannot be read as a firm constraint on early disk formation. The citation pattern is standard and pulls in the relevant JWST and ML-astro references without obvious gaps or loops. This work is aimed at groups tracking the emergence of rotationally supported systems and at people testing ML pipelines on new JWST data. A reader looking for a quick preliminary number from the first deep JWST fields could extract the 0.34 value, but anyone needing a result they can fold into models would want the validation details first. I would send it for peer review; the data are new and the pipeline is straightforward, so referees can request the missing tests without it being a waste of time.

Referee Report

3 major / 2 minor

Summary. The manuscript describes a two-step machine-learning pipeline applied to JWST/NIRCam multi-filter cutouts of 100 galaxies at 4 < z < 7.7: (i) a U-Net variational autoencoder (VAE) trained to denoise images by removing astrophysical and instrumental contaminants while preserving intrinsic morphology, and (ii) a rotation- and reflection-equivariant graph convolutional neural network (GCNN) classifier that labels the denoised cutouts as disk-like or non-disk. The pipeline yields a disk-like fraction of 0.34 across the full sample, reported also as a function of galaxy stellar mass.

Significance. If the VAE-GCNN pipeline can be shown to introduce no net bias in disk-feature recovery at these redshifts, the reported 0.34 fraction would constitute a useful empirical constraint on the prevalence of angular-momentum-supported structures in the first billion years, directly informing models of early disk formation that have traditionally predicted lower disk fractions.

major comments (3)

[Abstract] Abstract: the central numerical claim (disk-like fraction = 0.34) is presented without any accompanying classifier performance metrics (accuracy, precision, recall, F1, or confusion matrix), training-set composition details, or uncertainty estimates, rendering the result impossible to interpret as a physical measurement.
[Abstract and Methods] Abstract and Methods: the claim that the U-Net VAE 'preserves intrinsic morphology' is stated but unsupported by any quantitative test (e.g., morphology-injection recovery curves on simulated high-z galaxies, comparison of denoised vs. raw classifications on overlapping HST data, or ablation of training-set redshift distribution).
[Results] Results: the reported mass dependence of the disk fraction is given without binning details, sample sizes per bin, or error bars, so it is impossible to assess whether the trend is statistically significant or driven by small-number statistics.

minor comments (2)

[Abstract] Abstract, sentence 3: 'JWST 100 galaxies' should read '100 JWST galaxies' for grammatical clarity.
[Abstract] Abstract, final sentence: 'in dependence on the galaxy mass range' is awkward; 'and its dependence on galaxy stellar mass' would be clearer.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for their thorough review and valuable suggestions, which have helped us improve the clarity and robustness of our work. Below, we provide point-by-point responses to the major comments and indicate the revisions made to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central numerical claim (disk-like fraction = 0.34) is presented without any accompanying classifier performance metrics (accuracy, precision, recall, F1, or confusion matrix), training-set composition details, or uncertainty estimates, rendering the result impossible to interpret as a physical measurement.

Authors: We agree that the abstract should include key performance metrics to allow proper interpretation of the result. In the revised manuscript, we have added the classifier accuracy, precision, recall, and F1 scores obtained from cross-validation, a brief description of the training-set composition, and bootstrap-derived uncertainty estimates on the disk fraction. The full confusion matrix and additional training details remain in the Methods section. revision: yes
Referee: [Abstract and Methods] Abstract and Methods: the claim that the U-Net VAE 'preserves intrinsic morphology' is stated but unsupported by any quantitative test (e.g., morphology-injection recovery curves on simulated high-z galaxies, comparison of denoised vs. raw classifications on overlapping HST data, or ablation of training-set redshift distribution).

Authors: We acknowledge that quantitative validation strengthens the claim. We have added a dedicated subsection in the Methods describing morphology-injection recovery tests performed on simulated high-redshift galaxies, a direct comparison of denoised versus raw GCNN classifications for the subset of sources with overlapping HST coverage, and an ablation study varying the redshift distribution of the training set. These tests are now summarized in the abstract as well. revision: yes
Referee: [Results] Results: the reported mass dependence of the disk fraction is given without binning details, sample sizes per bin, or error bars, so it is impossible to assess whether the trend is statistically significant or driven by small-number statistics.

Authors: We have revised the Results section to specify the stellar-mass bin edges, the number of galaxies per bin, and binomial (or bootstrap) error bars on the disk fractions. The updated figure and text now allow readers to evaluate the statistical significance of the mass trend directly. revision: yes

Circularity Check

0 steps flagged

No circularity: disk fraction is direct empirical classifier output

full rationale

The paper's central result is the measured fraction 0.34 of disk-like galaxies obtained by applying a fixed VAE-GCNN pipeline to 100 JWST cutouts. No equations, fitted parameters, or derived quantities are presented that reduce to the input data or to the result itself by construction. The pipeline steps (denoising then classification) are described as applied identically but are not shown to be self-referential or justified solely via author self-citation chains. The reported fraction is therefore a straightforward count from the classifier labels and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is populated at the level of standard assumptions implicit in any supervised image-classification pipeline; no explicit free parameters, ad-hoc axioms, or new physical entities are stated.

axioms (1)

domain assumption The VAE reconstruction loss and GCNN training objective produce morphology-preserving outputs on high-redshift data
Implicit in the claim that denoised cutouts can be directly fed to the classifier for reliable disk identification

pith-pipeline@v0.9.0 · 5537 in / 1277 out tokens · 38586 ms · 2026-05-10T10:47:10.308734+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 6 canonical work pages

[1]

L., Teyssier, R., & Dekel, A

Andalman, Z. L., Teyssier, R., & Dekel, A. 2025, MNRAS, 540, 3350

2025
[2]

2011, A&A, 532, A74

Baillard, A., Bertin, E., De Lapparent, V., et al. 2011, A&A, 532, A74

2011
[3]

et al, 2025, ApJ, 994, 126 Capozziello S., Di Valentino E., Gurzadyan V.G., 2025, Eur

Borbolato, L. et al, 2025, ApJ, 994, 126 Capozziello S., Di Valentino E., Gurzadyan V.G., 2025, Eur. Phys. J. Plus, 140, 1011 MNRAS000, 000–000 (2026) Disk galaxies at4≤z <7.7 5

2025
[4]

E., Danhaive, A

Carreira, C., Robertson, B. E., Danhaive, A. L., et al. 2026, arXiv:2601.15957

work page arXiv 2026
[5]

Group Equivariant Convolutional Networks

Cohen, T. S., & Welling, M., 2016,Group Equivariant Convolu- tional Networks, Proc. 33rd International Conference on Ma- chine Learning (ICML), arXiv:1602.07576

work page Pith review arXiv 2016
[6]

Conselice, C. J. 2014, ARA&A, 52, 291

2014
[7]

J., Basham, J

Conselice, C. J., Basham, J. T. F., Bettaney, D. O., et al. 2024, MNRAS, 531, 4857

2024
[8]

et al, 2025, JHEA, 48, 100405

Dainotti, M.G. et al, 2025, JHEA, 48, 100405

2025
[9]

et al, 2026, arXiv:2603.00497 Di Valentino E., Levi Said J., Saridakis E.N., arXiv:2509.25288

Dainotti, M.G. et al, 2026, arXiv:2603.00497 Di Valentino E., Levi Said J., Saridakis E.N., arXiv:2509.25288

work page arXiv 2026
[10]

2026, AJ, 171, 59

Fang, G., Zhu, S., Xu, J., et al. 2026, AJ, 171, 59

2026
[11]

Genin, A., Shuntov, M., Brammer, G., et al. 2025, arXiv:2505.21622 Gurzadyan V.G., Ozernoi L.M., 1981, A&A, 95, 39 Gurzadyan V.G., Stepanian A., 2021, A&A, 653, A145 Gurzadyan V.G., Fimin N.N., Chechetkin V.M., 2023, A&A, 677, A161 Gurzadyan V.G., Fimin N.N., Chechetkin V.M., 2025, A&A, 694, A252 Gurzadyan V.G., Fimin N.N., Chechetkin V.M., 2026, A&A, 705, A70

work page arXiv 2025
[12]

2024, MNRAS, 534, 918

Kocjan, Z., Cadiou, C., Agertz, O., & Pontzen, A. 2024, MNRAS, 534, 918

2024
[13]

2023, arXiv:2312.12389

Kuhn, V., Guo, Y., Martin, A., et al. 2023, arXiv:2312.12389

work page arXiv 2023
[14]

et al, 2025, ApJ, 978, 14L

Lee, L.L. et al, 2025, ApJ, 978, 14L

2025
[15]

A., Conselice, C

Mager, V. A., Conselice, C. J., Seibert, M., et al. 2018, ApJ, 864, 123

2018
[16]

Mirzoyan, S.S., Khachatryan H., Yegorian G., Gurzadyan V.G., 2019, MNRAS, 489, L32

2019
[17]

Mirzoyan, S. S. 2025, Res. Astron. Astrophys., 25, 095006

2025
[18]

Mirzoyan, S. S. 2026, Res. Astron. Astrophys., 26, 025010

2026
[19]

Pandya, S., Patel, P., Blazek, J., 2023, arXiv:2311.01500 Rees M.J., 1988, Nature, 333, 523

work page arXiv 2023
[20]

2022, MNRAS, 509, 3966

Walmsley, M., Lintott, C., G´ eron, T., et al. 2022, MNRAS, 509, 3966

2022
[21]

W., Lintott, C

Willett, K. W., Lintott, C. J., Bamford, S. P., et al. 2013, MNRAS, 435, 2835

2013
[22]

2024, Res

Yao, Y., Zhang, J., Du, P., & Dong, S. 2024, Res. Astron. Astro- phys., 24, 035015

2024
[23]

Xu, Y., et al, 2024, ApJ, 976, 142 MNRAS000, 000–000 (2026)

2024