Discovery of the Polar Ring Galaxies with deep learning
D. V. Dobrycheva, O. O. Hetmantsev, I. B. Vavilova, A. Shportko, O. Gugnin, O. V. Kompaniiets
TL;DR
The paper tackles the rarity of polar ring galaxies (PRGs) in large surveys by combining a visually curated catalog with a CNN-based image analysis pipeline and transfer learning from GALFIT-simulated galaxies. It demonstrates that a CNN, aided by image segmentation, augmentation, and synthetic data, can identify PRG candidates, yielding 3 CNN-discovered PRGs and several visually confirmed cases, while constructing a catalog of 179 inspected PRGs. A multiwavelength case study of SDSS J140644.42+471602.0 using CIGALE indicates a star formation rate of 71 $M_{ m 5}$ per year and a total stellar mass of 8.34×10^{10} $M_{ m 5}$ with an older stellar mass-dominated population, suggesting ongoing interactions. The work highlights the viability of deep learning for PRG discovery in big sky surveys and outlines future directions, including GAN-based data augmentation and hybrid analyses combining imaging with photometry/spectroscopy, to improve robustness and expand the PRG census.
Abstract
The aim of our research is to create a catalog of strong and good candidates for PRGs using existing catalogs of PRGs, develop an image-based approach with machine learning methods for the search and discovery of PRGs in a big sky survey, and explore the capability of the CIGALE software for determining their multiwavelength properties. For the first time, we applied a deep learning method to the search for PRGs. We visually inspected galaxies from existing catalogs of PRGs to create a training sample based on high-quality SDSS images. Since the resulting training sample was extremely small (87 strong and good PRGs), we applied augmentation, image segmentation, and ensemble learning techniques. However, most effective method was transfer learning with its ability to enlarge the training sample by synthetic images generated by GALFIT. To examine deep learning approach for finding new PRGs we used the SDSS catalog of galaxies at z < 0.1. The method with synthetic images showed that even with overtraining we were able to find galaxies with a ring pattern. Our deep learning approach has resulted in the discovery of three PRGs (SDSS J140644.42+471602.0; SDSS J133650.48+492745.3; SDSS J095717.30+364953.5). Also, we visually inspected the Catalog of the SDSS Ring galaxies at z < 0.1 and discovered four PRGs among ~2,200 ring galaxies (SDSS J095851.32+320422.9; SDSS J104211.05+234448.2; SDSS J162212.63+272032.2; SDSS J104600.10+090627.2). One of the discovered galaxies with transfer learning, SDSS J140644.42+471602.0, was studied with CIGALE software to determine its spectral energy distribution in IR-UV bands. The current SFR is 71 M_sun per year, although the lack of FUV data limits this estimate. The total stellar mass is 8.34x10^{10} M_sun. The predominance of an old stellar population (two-thirds of the total mass) suggests that this PRG is undergoing interaction process.