Title: Determining Satellite Infall Times Using Machine Learning

Abstract: A key unknown of the Milky Way (MW) satellites is their orbital history, and, in particular, the time they were accreted onto the MW system since it marks the point where they experience a multitude of environmental processes. We present a new methodology for determining infall times, namely using a neural network (NN) algorithm. The NN is trained on MW-analogues in the EAGLE hydrodynamical simulation to predict if a dwarf galaxy is at first infall or a backsplash galaxy and to infer its infall time. The resulting NN predicts with 85\% accuracy if a galaxy currently outside the virial radius is a backsplash satellite and determines the infall times with a typical 68\% confidence interval of 4.4 Gyrs. Applying the NN to MW dwarfs with Gaia EDR3 proper motions, we find that all of the dwarfs within 300 kpc had been inside the Galactic halo. The overall MW satellite accretion rate agrees well with the theoretical prediction except for late times when the MW shows a second peak at a lookback time of 1.5 Gyrs corresponding to the infall of the LMC and its satellites. We also find that the quenching times for ultrafaint dwarfs show no significant correlation with infall time and thus supporting the hypothesis that they were quenched during reionisation. In contrast, dwarfs with stellar masses above $10^5~M_\odot$ are found to be consistent with environmental quenching inside the Galactic halo, with star-formation ceasing on average at $0.5^{+0.9}_{-1.2}$ Gyrs after infall.