Speaker
Description
Machine learning techniques are widely implemented to predict galaxy properties based on halo or subhalo properties in semi-analytic models (SAM) and hydrodynamic simulations. The multivariate relation between galaxy and halo or subhalo can be captured and reproduced efficiently. However, the galaxy formation and evolution in galaxy formation models may deviate from those in the real Universe, as well as the galaxy-halo connections. With the ELUCID simulation which is constructed to reproduce the density field of the SDSS and the SDSS-ELUCID matched catalog which links SDSS galaxies to ELUCID subhalos, we train random forest (RF) models to predict SDSS galaxy properties based on ELUCID subhalo properties. The RF recovers the absolute magnitude of galaxies reasonably well but exhibits low accuracy in color prediction. To investigate the possible reasons for this, we perform similar RF predictions using galaxies from SAM and hydrodynamic simulation and compare the results with that from the SDSS-ELUCID sample. Uncertainties in the galaxy-subhalo connection of the SDSS-ELUCID sample arising from the limited accuracy of the ELUCID construction are also considered. Our analysis is helpful for a better understanding of the differences in the galaxy-subhalo connection between galaxy formation models and the real Universe.
