Automated Detection of Galaxy Mergers using Cosmological Simulations and Deep Learning (Dr. Fernando Caro; LERMA, Observatoire de Paris)

Desde Abril 01, 2019 13:15 hasta Abril 01, 2019 13:45

En Instituto de Astrofísica, Pontificia Universidad Católica, Vicuña Mackenna 4860, Santiago, Chile

Categorías: Seminarios

One of the main predictions derived from ?-CDM cosmology is that structure grows hierarchically as consequence of the gravity-driven assembly of dark matter halos and the galaxies they host, making galaxy mergers a fundamental element within the current framework of galaxy evolution. Despite this, the specific role that galaxy mergers play in the different processes observed in galaxy evolution along with the rates that are estimated, from either observations or simulations, remain not fully understood. In that regard, the development of accurate methods to detect galaxy mergers is something extremely relevant in order to improve our comprehension of the merging process, but also to bring into agreement galaxy formation models with the available observational data.

Inspired by the recent success of multiple deep learning techniques used to address different astrophysical problems, we present a novel method based on the utilization of Convolutional Neural Networks (CNNs) and the Horizon-AGN cosmological simulation for identifying galaxy mergers in an automated and accurate manner. The main idea behind this method consists on generating large training sets of HST-like multiband mock observations of mergers, isolated galaxies (no-mergers) and projection effects (fake-mergers) employing as input the virtual galaxies of the Horizon-AGN simulation, which then are used to train a CNN that is able of identifying galaxy mergers. These mock observations are generated considering massive galaxies (M_star <= 10^10 M_sun) in the redshift range 0.5 < z < 3.5, and in the case of galaxy mergers, only those that correspond to major ones (stellar mass ratio <= 1:4) are taken into account.

The performance exhibited by this new method outperforms not only the classic approaches based on non-parametric morphologies employed for identifying galaxy mergers, such as CAS and G-M20, but also the extended and improved versions of these methods that rely on the utilization of machine learning. Moreover, the use of a simulated training set allowed us to relate the observability timescales of the different methods analyzed with the actual timescales of the galaxy mergers selected from the simulation.