Galaxy mergers classification using CNNs trained on Sérsic models, residuals and raw images.
D. Chudy, W. J. Pearson, A. Pollo, L. E. Suelves, B. Margalef-Bentabol, L. Wang, V. Rodriguez-Gomez, A. La Marca · 2025
Read Galaxy mergers classification using CNNs trained on Sérsic models, residuals and raw images. by D. Chudy, W. J. Pearson, A. Pollo, L. E. Suelves, B. Margalef-Bentabol, L. Wang, V. Rodriguez-Gomez, A. La Marca (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-16603.txt.
Introduction
Chudy and co-authors ask which parts of a galaxy's image tell a neural network that two galaxies are merging. Mergers drive galaxy growth, and future surveys will produce far too many images for humans to sort. Networks are assumed to spot faint tidal features, but nobody had separated the clues cleanly.
Using mock images from the IllustrisTNG simulations dressed up to look like real telescope data, the team trained three identical networks on full images, smooth model images, and leftover residual images. The full images scored best at seventy-four percent, but models alone reached seventy and residuals sixty-eight. Position and shape information, not just faint features, carries real merger evidence.
Catalog record
| CREATOR | D. Chudy, W. J. Pearson, A. Pollo, L. E. Suelves, B. Margalef-Bentabol, L. Wang, V. Rodriguez-Gomez, A. La Marca |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-16603
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Galaxy mergers classification using CNNs trained on Sérsic models, residuals and raw images. D. Chudy, W. J. Pearson, A. Pollo, L. E. Suelves, B. Margalef-Bentabol, L. Wang, V. Rodriguez-Gomez, A. La Marca Abstract Context. Galaxy mergers are crucial for understanding galaxy evolution, and with large upcoming datasets, automated methods, such as Convolutional Neural Networks (CNNs), are needed for efficient detection. It is understood that these networks work by identifying deviations from the regular, expected shapes of galaxies, which are indicative of a merger event. Aims. Using images from the IllustrisTNG simulations, we aim to check the importance of faint features, source position and shape information present in galaxy merger images on the performance of a CNN merger vs. non-merger classifier. Methods. We fit Sérsic profiles to each galaxy in mock images from the IllustrisTNG simulations. We subtract the profiles from the original images to create residual images, and we train three identical CNNs on three different datasets – original images (CNN1), model images (CNN2), and residual images (CNN3). Results. We found that it is possible to conduct galaxy merger classification based only on faint features, source position and shape information present in residual images and model images, respectively. The results show that the CNN1 correctly classifies 74% of images, while CNN2 70%, and CNN3 68%. Conclusions. Source position and shape information is crucial for pre-merger classification, while residual features are important for post-merger classification. CNN3 classifies post-mergers in the latest merger stage the best out of all three classifiers. 1 Introduction In the Lambda-cold dark matter cosmology, dark matter (DM) halos form due to gravitational instability, allowing baryonic matter to cool, accumulate, and form galaxies. These galaxies then grow through both accretion within these halos and mergers. Hence galaxy mergers provide crucial insights into how galaxies evolve over time. The merging process proceeds differently depending on the properties of colliding galaxies. Major mergers, where the stellar mass ratio of the merging components is up to 4:1, can combine their structures, resulting in a significant change in morphology if the components collide at the right speed and angle ( Bournaud et al. 2005 ; Lotz et al. 2010 ) . Conspicuous merger features, such as bridges, double nuclei, and tidal features (tails and plumes) of different sizes and shapes may appear ( Toomre & Toomre 1972 ) . This process can feed supermassive black holes, activating galactic nucleus (AGN) activities ( Gao et al. 2020 ; Di Matteo et al. 2005 ) , and trigger star formation ( Pearson et al. 2019a ; Ellison et al. 2013 ) . When the mass of one of the merging galaxies is significantly larger than the mass of the other one (minor mergers), the merging process is less violent, and the less massive galaxy is usually incorporated into the larger one, leaving the larger galaxy fairly undisturbed. The galaxy merger fraction has evolved significantly during cosmic time. It is believed that mergers account for less than 10% of the low-redshift galaxies ( Ferreira et al. 2020 ; Mundy et al. 2017 ; Duncan et al. 2019 ) , with the percentage increasing to 20% within z\in[2,3] ( Tasca et al. 2014 ) . The merger fraction decreases to 5% at z\approx 6 ( Ventou et al. 2019 ) . Merger rates in Illustris ( Vogelsberger et al. 2014 ) simulations align with observational rates for medium-sized galaxies ( Rodriguez-Gomez et al. 2015 ) , therefore, the merger fraction in Illustris agrees with the merger fraction in observations. It is possible to use visual inspection ( Darg et al. 2010 ) to identify a merger, though this method is challenging to replicate consistently, time-consuming for large samples of galaxies, and may suffer from low accuracy, incompleteness ( Huertas-Company et al. 2015 ) , and human biases ( Lambrides et al. 2021 ) . Using morphological parameters is a way to quantify morphology with reliable and reproducible measurements. Concentration ( Okamura et al. 1984 ; Kent 1985 ; Conselice 2003 ) , Asymmetry ( Conselice et al. 2000 ) , and Smoothness ( Takamiya 1999 ; Conselice 2003 ) (CAS parameters) and the bulge-merger statistics - a combination of the Gini coefficient ( Abraham et al. 2003 ) and M_{20} (the second-order moment of the brightest 20% of the galaxy) ( Lotz et al. 2004 ) - serve as feature detectors and are often used for merger classification ( Okamura et al. 1984 ; Abraham et al. 1996 ; Bershady et al. 2000 ; Snyder et al. 2015 ) . Deriving those parameters requires images with sufficiently high spatial resolution. Conselice et al. 2000 has shown that the accuracy of CAS parameters begins to decline below the threshold value of 1\,\textrm{kpc}\,\textrm{pixel}^{-1} . Another commonly used method to identify mergers is the close pair method, which selects galaxy pairs that are close on the sky and in redshift ( Barton et al. 2000 ; Alonso et al. 2004 ; De Propris et al. 2005 ; Duncan et al. 2019 ) . This method can use both spectroscopic and photometric observations ( Lin et al. 2008 ) . While spectroscopic methods provide greater precision than photometric techniques ( Hildebrandt et al. 2008 ) , they are also more resource-intensive. Identifying galaxy mergers may also be possible using only photometric information. Suelves et al. 2023 found that photometric bands, colours, and their errors provide accurate classification of SDSS observations. With the advent of upcoming surveys such as Euclid ( Laureijs et al. 2011 ) and Legacy Survey of Space and Time ( Ivezić et al. 2019 , LSST;) that will collect large amount of data it becomes even more important to create algorithms to automate time-consuming, repetitive tasks such as identifying galaxy mergers. The application of machine learning (ML) in the field of astronomy has experienced a significant surge over the past few years. Recent studies have shown that it is possible to use Convolutional Neural Networks (CNN) to address the visual merger classification problem. Ackermann et al. 2018 used transfer learning to fine-tune CNN trained to classify real life objects, using SDDS observations. Pearson et al. 2019b trained one CNN with observations from SDSS and another with simulated galaxies from EAGLE observations receiving an accuracy of 91.5% and 65.2% respectively. Combination of CNNs and Linear Networks has led to creation of the North Ecliptic Pole merging galaxy catalogue ( Pearson et al. 2022 ) for z<0.3 . Ćiprijanović et al. 2020 used a CNN to distinguish between mergers and non-mergers in simulated images from Illustris-1 at z=2 with 79% accuracy. Bickley et al. 2021 showed that CNNs outperform Gini-M20 and asymmetry methods in post-merger and isolated galaxy classification. The performance of these networks vary due to many factors, including differences in setup. Margalef-Bentabol et al. 2024 conducted a systematic and consistent study between machine learning-based detection methods to understand the relative performance of different ML methods using the same training and test data. A total of six ML methods have been tested, based on the same mock IllustrisTNG simulation images generated to mimic the HSC data. All CNNs that were not pre-trained with other galaxy images show similar results, using different architectures. This may indicate that preprocessing the