Analysis of Ring Galaxies Detected Using Deep Learning with Real and Simulated Data

Abstract Understanding the formation and evolution of ring galaxies, which possess an atypical ring-like structure, is crucial for advancing knowledge of black holes and galaxy dynamics. However, current catalogs of ring galaxies are limited, as manual analysis takes months to accumulate an appreciable sample of rings. This paper presents a convolutional neural network (CNN) to identify ring galaxies from unclassified samples. A CNN was trained on 100,000 simulated galaxies, transfer learned onto a sample of real galaxies, and applied to a previously unclassified data set to generate a catalog of rings, which was then manually verified. Data augmentation with a generative adversarial network to simulate images of galaxies was also employed. The resulting catalog contains 1967 ring galaxies. The properties of these galaxies were then estimated from their photometry and compared to the Galaxy Zoo 2 catalog of rings. However, the model’s precision is currently limited due to a severe imbalance of rings in real data sets, leading to a significant false-positive rate of 41.1%, which poses challenges for large-scale applications in surveys imaging billions of galaxies. This study demonstrates the potential of optimizing machine learning pipelines for low training data in rare morphologies and underscores the need for further refinements to enhance precision for extensive surveys like the Vera Rubin Observatory Legacy Survey of Space and Time.