Warped Disk Galaxies: Statistical Properties from DESI Legacy Imaging Surveys DR8

TL;DR

We address the origin and demographics of warped disks by constructing the largest warp catalog to date from DESI DR8 optical imaging. Using a CNN-based classifier (EfficientNet-B3) trained on visually labeled edge-on galaxies from Galaxy Zoo DESI, the study labels 23,996 warped and 288,562 non-warped systems among 595,651 edge-on disks. The analysis reveals that warped disks are larger, more disk-dominated (lower Sérsic index, smaller bulges), bluer, more gas-rich, and have higher SFRs and sSFRs, with warps preferentially found in denser small-scale environments ($R_{\mathrm{proj}} \lesssim 50$ kpc). These results provide empirical constraints on warp formation mechanisms, suggesting a role for external torques or gas accretion, and highlight the value of deep learning pipelines for large-scale galaxy morphology studies.

Abstract

Warped structures are often observed in disk galaxies, yet their physical origin is still under investigation. We present a systematic study of warped edge-on disk galaxies based on imaging data from the DESI Legacy Imaging Surveys DR8, with the expectation that this large sample size, enabled by wide-area surveys, will offer new perspectives on the formation of disk warps. Using a deep learning approach, we trained an EfficientNet-B3 convolutional neural network to classify the morphology of edge-on-disk galaxies into warped and non-warped categories. Our model was trained on a curated and visually verified set of labeled galaxy images and applied to a large dataset of over 595,651 edge-on disk galaxies selected from the Galaxy Zoo DESI catalog. Our results provide the largest warp catalog to date, consisting of 23996 warped edge-on disk galaxies, and reveal statistical trends between warp occurrence and galaxy properties. Compared to their non-warped counterparts, these warped disk galaxies tend to have bluer colors, lower stellar masses, higher gas fractions and star-formation rates, smaller Sérsic indices and larger disk sizes. In addition, warped disk galaxies show higher projected number densities of neighboring galaxies than their non-warped counterparts, particularly within \( R_{\mathrm{proj}} \lesssim 50~\mathrm{kpc} \), where the local number density is roughly twice as high.

Figures (11)

• Figure 1: Three two-dimensional projections of edge-on galaxies in the parameter space of merging_minor-disturbance_fraction, bulge-size_small_fraction, and redshift. Contours indicate the distribution of all edge-on galaxies in each projected plane, and scatter points show the subset selected for labeling using our 3D binned sampling strategy, with up to 20 galaxies randomly drawn per bin. This approach ensures that rare types of galaxies are represented, improving the completeness and diversity of the training set.
• Figure 2: The figure shows example galaxy images used for training. The top panel shows non-warped galaxies, and the bottom panel shows warped galaxies.
• Figure 3: Confusion matrix of the best-performing model among all saved epochs.
• Figure 4: Correct predictions (green) tend to have high confidence and low uncertainty, while incorrect predictions (red) show the opposite pattern. This validates the use of these metrics for high-confidence selection.
• Figure 5: Images of randomly selected galaxies with high-confidence model predictions: the top panel shows non-warped galaxies, and the bottom panel shows warped galaxies.