Enhancing Galaxy Classification with U-Net Variational Autoencoders. III. Disk-like Galaxy Identification in JWST Samples of up to redshift 4

S. S. Mirzoyan, A. Avagyan · 2026

Read Enhancing Galaxy Classification with U-Net Variational Autoencoders. III. Disk-like Galaxy Identification in JWST Samples of up to redshift 4 by S. S. Mirzoyan, A. Avagyan (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-2605-31420.txt.

Introduction

Mirzoyan and Avagyan extend a U-Net VAE galaxy sorting framework to a larger JWST sample spanning redshift 0.5 to 4. They focus on massive systems with stellar masses above 10 billion solar masses. Their cleaning and sorting pipeline analyses 1,380 galaxies to identify disk-like shapes across cosmic time. Their sorter finds 382 disk-like galaxies; some show fuzzy features in line with current deep-learning limits. Disk-like forms persist over many redshifts and stellar masses, hinting massive disks were more common in the early universe than once assumed.

Catalog record

CREATORS. S. Mirzoyan, A. Avagyan
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40849
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.31420

Browse the full library · About the Tower of Babel

Full text

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

Enhancing Galaxy Classification with U-Net Variational Autoencoders. III. Disk-like Galaxy Identification in JWST Samples of up to redshift 4 S. S. Mirzoyan, A. Avagyan Abstract In this third study of the series, we extend our U-Net Variational Autoencoder-based galaxy classification framework to a significantly larger JWST sample spanning the redshift range 0.5 < z < 4 0.5<z<4 . Focusing on massive systems with stellar masses exceeding 10 10 ​ M ⊙ 10^{10}\,M_{\odot} , we analyze 1,380 galaxies that satisfy these criteria and apply our previously developed denoising and classification pipeline to identify disk-like morphologies across cosmic time. Within this population, our classifier detects 382 disk-like galaxies, with a subset showing uncertain features consistent with the expected performance limits of current deep-learning models. This expanded dataset allows us to examine the distribution of disk-like systems in a statistically meaningful high-redshift regime, including epochs where well-ordered disks are traditionally expected to be rare. The results demonstrate that disk-like structures persist across a broad range of redshifts and stellar masses, suggesting that massive disks may be more common in the early universe than previously assumed. These findings emphasize the value of combining advanced denoising techniques with machine-learning-based morphological analysis for characterizing galaxy populations in large JWST surveys. \volnopage Vol.0 (20xx) No.0, 000-000 keywords galaxies: classification: morphology - techniques: convolutional neural networks: variational autoencoder: u-net 1 Introduction Understanding when and how ordered disks assemble and survive in massive galaxies requires a uniform, redshift-spanning morphological census anchored to astrophysically motivated selection criteria. Classical frameworks emphasize the diversity of morphologies within broad Hubble classes ( Buta 2019 ; Mo et al. 2010 ; Sparke & Gallagher 2000 ; Sandage 2005 ) , while comprehensive reviews highlight the physical drivers of morphological transformation across cosmic time ( Bell et al. 2026 ; Conselice 2014 ; Naab & Ostriker 2017 ; Genel et al. 2018 ) . Robust inferences at high redshift are sensitive to surface-brightness dimming, PSF and bandpass effects, and morphology- k -corrections ( Conselice et al. 2024 ; Conselice et al. 2023 ; Conselice 2020 ; Lotz et al. 2004 ; Mager et al. 2018 ; van der Wel et al. 2014 ) . A mass-limited strategy mitigates some of these biases and enables direct comparison across epochs. JWST (James Webb Space Telescope) ( Gardner et al. 2006 ; Rigby et al. 2023 ) imaging extends quantitative morphology well into the early universe, with large public catalogs facilitating population-level analyses ( Genin et al. 2025 ; Carreira et al. 2026 ; Shuntov et al. 2025 ; Sun et al. 2024 ) . Spectroscopic surveys further contextualize structural evolution during reionization and Cosmic Noon ( Meyer et al. 2025 ; Belli et al. 2025 ; Rieke et al. 2023 ; van der Wel et al. 2014 ) . These data require morphological pipelines that can handle heterogeneous signal-to-noise conditions, diverse structural features, and domain-shift effects across filters and redshifts. Recent JWST -based observational studies have provided important context for the present work. In particular, visual classifications from the CEERS (Cosmic Evolution Early Release Science) survey have revealed a surprisingly high fraction of spiral-like galaxies at intermediate redshifts, with up to nearly half of systems at z\sim 0.5 exhibiting discernible spiral structure, declining to \sim 8\% by z\sim 4 ( Kuhn et al. 2024 ) . These fractions exceed earlier estimates based on HST imaging and suggest that dynamically cold disks and spiral structure may have emerged at much earlier cosmic times than traditionally assumed. Artificial redshifting tests further indicate that intrinsic spiral fractions below \sim 10 – 20\% are statistically disfavored at z\sim 2 – 3 , motivating mass-limited and resolution-aware morphology analyzes when probing disk formation in the early universe. Independent structural analyses based on JWST /NIRCam surface-brightness decompositions have further reported that flattened, disk-like components are already common at high redshift, with only a modest decline in disk-dominated systems toward earlier epochs for massive galaxies ( Tsukui et al. 2026 ) . Broader morphology surveys spanning multiple public JWST fields have likewise reported high disk fractions at low and intermediate stellar masses, while finding a pronounced decline in disk dominance at the highest masses, where spheroid-like morphologies become increasingly prevalent ( Lee et al. 2024 ; Carreira et al. 2026 ) . Extending disk studies to even earlier epochs, Mirzoyan et al. 2026 found that disk-like morphologies constitute a significant fraction of the population at 4\lesssim z\lesssim 7.7 , emphasizing the importance of noise suppression for robust high-redshift inference. In parallel, new unsupervised machine-learning approaches have improved the efficiency of large-scale morphology analysis, such as the ConvNeXt+UMAP framework, which combines hierarchical CNN Convolutional Neural Network feature extraction based on the ConvNeXt (Convolutional Neural Network for the 2020s) architecture ( Liu et al. 2022 ) with nonlinear manifold learning using UMAP (Uniform Manifold Approximation and Projection) ( McInnes et al. 2018 ) to classify nearly 10^{5} COSMOS (Cosmic Evolution Survey) galaxies ( Scoville et al. 2007 ) into physically meaningful morphology groups using an optimized set of 20 clusters ( Fang et al. 2026 ) . More generally, representation-learning approaches applied to JWST data suggest that a fraction of visually identified disks at z\gtrsim 3 may occupy a more structurally ambiguous regime when compared against simulated galaxy populations, highlighting potential systematics in visually driven or supervised classifications ( Vega-Ferrero et al. 2024 ) . Large-scale unsupervised clustering analyses further demonstrate that disk-like galaxies populate multiple distinct regions of morphology space rather than forming a single homogeneous class, underscoring the diversity of disk structures and the sensitivity of disk identification to the adopted representation and clustering strategy ( Fang et al. 2026 ) . Together, these developments highlight both the rapid evolution of galaxy structures across cosmic time and the growing need for automated, noise-robust classification methods capable of operating across heterogeneous imaging datasets. Machine learning (ML) has become central to modern galaxy morphology studies, lens discovery, and classification pipelines ( Mirzoyan 2025a ; Mirzoyan 2025b ; Mirzoyan et al. 2019 ; Ghaderi et al. 2025 ; Tian et al. 2025 ; Atemkeng et al. 2025 ) . CNN and group-equivariant architectures have shown excellent performance on benchmark galaxy datasets ( Pandya et al. 2023 ) , though input noise and imaging artifacts continue to limit accuracy—especially for high-redshift sources. To address this limitation, in Paper I of this series Mirzoyan 2025a we introduced a denoising framework based on a U-Net Variational Autoencoder (VAE) ( Ronneberger et al. 2015 ; Kingma & Welling 2014 ) . Trained on realistically contaminated EFIGI (Extraction de Formes Idéalisées de Galaxies en Imagerie) images ( Baillard et al. 2011 ) , the model effectively suppressed noise while preserving morphology-relevant structures, yielding consistent gains in downstream classification metrics across multiple deep-learning architectures. Paper I established the methodological foundation that denoising is not merely cosmetic, but functionally beneficial for morphological inference. In Paper II Mirzoyan 2025b , we extended this denoising pipeline to JWST Near-Infrared Camera (NIRCam) ( Rieke et al. 2005 ; Rieke et al. 2023 ) observations up to z\approx 4 . There we demonstrated that VAE-based denoising improves the recoverability of faint structural components and enhances disk/non-disk separability in early-universe galaxies. Analysis of a pilot subsample revealed that disk-like

Read the complete text