Robust machine learning model of inferring the ex-situ stellar fraction of galaxies from photometric data

Abstract

We search for parameters defined from photometric images to quantify the ex situ stellar mass fraction of galaxies. We created mock images using galaxies in the cosmological hydrodynamical simulations TNG100, EAGLE, and TNG50 at redshift $z=0$. We define a series of parameters describing their structures. In particular, the inner and outer halo of a galaxy are defined by sectors ranging from $45-135$ degrees from the disk major axis, and with radii ranging from $3.5-10$ kpc and $10-30$ kpc, respectively, to avoid the contamination of disk and bulge. The surface brightness and colour gradients are defined by the same sectors along the minor axis and with similar radii ranges. We used the Random Forest method to create a model that predicts $f_{\rm exsitu}$ from morphological parameters. The model predicts $f_{\rm exsitu}$ well with a scatter smaller than 0.1 compared to the ground truth in all mass ranges. The models trained from TNG100 and EAGLE work similarly well and are cross-validated; they also work well in making predictions for TNG50 galaxies. The analysis using Random Forest reveals that $\nabla \rho_{\rm outer}$, $\nabla (g-r)_{\rm outer}$, $f_{\rm outerhalo}$ and $f_{\rm innerhalo}$ are the most influential parameters in predicting $f_{\rm exsitu}$, underscoring their significance in uncovering the merging history of galaxies. We further analyse how the quality of images will affect the results by using SDSS-like and HSC-like mock images for galaxies at different distances. Our results can be used to infer the ex situ stellar mass fractions for a large sample of galaxies from photometric surveys.