Searching for compaction in the TNG50 cosmological simulation using deep learning

Abstract

We optimize a convolutional neural network, intending to study an astrophysical process known as ‘blue nuggets’ (BN), which consists of a compaction followed by a central quenching that occurs in young galaxies at high redshifts. This network is evaluated with mock ‘observed’ images of galaxies at three phases of evolution (Pre-BN, BN and Post-BN), generated by the zoom-in hydro-cosmological simulation VELA. We then use this to classify galaxies from the TNG50 simulation in these three phases, and finally, we study their physical properties such as the redshift, the effective radius and the star formation rate (SFR), as well as the masses of gas, of stars, and of the central supermassive black holes. The network successfully detects this compaction phase in the new simulation, consistent with the features observed in VELA galaxies. We highlight the existence of a temporal sequence, together with the fact that the BN phase forms stars while the Post-BN does not. Furthermore, the BN phase is associated with a gas mass peak at z ∼ 2 and with a smaller radius.