A Simulation Driven Deep Learning Approach for Separating Mergers and Star Forming Galaxies: The Formation Histories of Clumpy Galaxies in all the CANDELS Fields
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 (CNN) to separate star forming galaxies from post-mergers using a dataset 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 neighbouring 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 disc-dominated systems while post-mergers show distributions of transitioning discs 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 (SFMS) can be attributed to major post-mergers. Finally, we use our new approach to update our previous measurements of galaxy merger rates $\mathcal{R} = 0.022 \pm 0.006 \times (1+z)^{2.71\pm0.31}$