A Simulation-driven Deep Learning Approach for Separating Mergers and Star-forming Galaxies: The Formation Histories of Clumpy Galaxies in All of the CANDELS Fields
Abstract
Being able to distinguish between galaxies that have recently undergone major-merger events, or are experiencing intense star formation, is crucial for making progress in our understanding of the formation and evolution of galaxies. As such, we have developed a machine-learning framework based on a convolutional neural network to separate star-forming galaxies from post-mergers using a data set of 160,000 simulated images from IllustrisTNG100 that resemble observed deep imaging of galaxies with Hubble. We improve upon previous methods of machine learning with imaging by developing a new approach to deal with the complexities of contamination from neighboring sources in crowded fields and define a quality control limit based on overlapping sources and background flux. Our pipeline successfully separates post-mergers from star-forming galaxies in IllustrisTNG 80% of the time, which is an improvement by at least 25% in comparison to a classification using the asymmetry (A) of the galaxy. Compared with measured Sérsic profiles, we show that star-forming galaxies in the CANDELS fields are predominantly disk-dominated systems while post-mergers show distributions of transitioning disks to bulge-dominated galaxies. With these new measurements, we trace the rate of post-mergers among asymmetric galaxies in the universe, finding an increase from 20% at z = 0.5 to 50% at z = 2. Additionally, we do not find strong evidence that the scattering above the star-forming main sequence can be attributed to major post-mergers. Finally, we use our new approach to update our previous measurements of galaxy merger rates ${ \mathcal R }=0.022\pm 0.006\times {(1+z)}^{2.71\pm 0.31}$ .
- Publication:
-
The Astrophysical Journal
- Pub Date:
- May 2022
- DOI:
- 10.3847/1538-4357/ac66ea
- arXiv:
- arXiv:2204.05786
- Bibcode:
- 2022ApJ...931...34F
- Keywords:
-
- Galaxy mergers;
- Convolutional neural networks;
- 608;
- 1938;
- Astrophysics - Astrophysics of Galaxies;
- Astrophysics - Instrumentation and Methods for Astrophysics
- E-Print:
- Accepted for publication in ApJ. 28 pages, 18 Figures, 2 Tables. Pretrained models available at https://github.com/astroferreira/FERREIRA2022