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

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 $0.5 < z < 4$. Focusing on massive systems with stellar masses exceeding $10^{10}\,M_\odot$, 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.