Observations indicate that a continuous supply of gas is needed to maintain observed star formation rates in large, disky galaxies. To fuel star formation, gas must reach the inner regions of such systems. Despite its crucial importance for galactic evolution, how/where gas joins disks is poorly constrained observationally and rarely explored in fully cosmological simulations. To investigate gas accretion onto galaxies at low redshift and how it fuels star formation in central regions, we analyze the FIRE-2 cosmological zoom-in simulations, focusing on 4 Milky Way mass galaxies (M_halo~10^12 M_solar).

In Chapter 2, we present a phenomenological analysis of how gas accretes onto simulated Milky Way mass disks and transports inwards, fueling star formation. We find that at z~0, gas approaches the disk with angular momentum similar to the gaseous disk edge and average radial speeds of 10-20 km/s, piling up near the edge and settling into full rotational support. These flows are largely hidden from observations that search for inflows via large deviations from galactic rotation. Within the disk, average radial speeds slow down to a few km/s of inward motion. Radial motion of the gas in the disk is complex, dominated by spiral arm-induced oscillations and feedback effects.

In Chapter 3, we present a study on the torques that drive these flows. Gas in the disk is fully rotationally supported, so must lose angular momentum to move inward. In this chapter, we characterize gravitational torques from stars, dark matter, and gas-gas interactions; magnetohydrodynamical torques from pressure gradients, non-continuum hydrodynamical terms, magnetic fields, and viscous shearing; and torques from stellar feedback.

In Chapter 4, we present an analysis of how to constrain these radial flows in observations, focusing on the use of Convolutional Neural Networks (CNNs) to infer radial mass fluxes from HI 21 cm spectral data cubes. Constraining flows within the disks of galaxies is particularly difficult, with large observational uncertainties. We additionally compare the use of more traditional tilted ring modeling to infer radial mass flux rates with CNN performance and find that our networks offer similar or better accuracy with dramatically enhanced speeds.