A Convolutional Neural Network (CNN) Approach to Galaxy Merger Classification
JURPA 2026 Cover Image.
Abstract. Galaxy mergers provide critical insights into cosmic evolution, yet identification in survey data remains challenging. We introduce a rigorously verified dataset of 228 hand-classified mergers and 292 nonmergers, curated from Galaxy Zoo DR2 and DR5, VV objects, and Toomre merger candidates. Using this data, we trained a convolutional neural network with four convolutional blocks, utilizing L2 regularization, dropout, and augmentations like random rotations and flips. Our model achieved 74% test accuracy, demonstrating the dataset’s utility. The findings indicate that improvements through synthetic data or domain adaptation would enhance machine learning performance. We present this dataset as a resource to advance automated merger classification.
