Populating dark matter halos with galaxies using machine learning

Abstract

Robust understanding of the connection between galaxies and the dark matter (DM) halos that host them is achieved using modern hydrodynamic simulations. However, such simulations are generally computation-intensive and require a lot of computing time. In this study, we develop a machine learning (ML) model to map baryonic properties of galaxies onto DM halos from the SIMBA cosmological hydrodynamic simulation, at redshifts z = 0 to z = 2, in a computationally efficient manner. In particular, we trained a suite of extremely randomized trees (ERT) models to predict central galaxy properties using DM halo properties. We find that ML predictions are able to successfully recover the mean relations between halo mass and galaxy properties, as well as the fundamental metallicity relation (FMR). However, it fails to reproduce the observed scatter around the mean relations observed in studies based on hydrodynamic simulations. Results from this study illustrate both the potential and current limitations of traditional ML approaches for generating realistic galaxy catalogs.