FLAMINGO: Using Gaussian process machine learning to calibrate sub-grid models for cosmological simulations
FLAMINGO (Full-hydro Large-scale structure runs with All-sky Mapping for the Interpretation of Next Generation Observations) aims to do a Hydrodynamical simulation in a volume of (3.2 Gpc)^3 at a gas resolution of ~10^9Msun, with several variations in smaller volumes. With such a large volume, the mock observables that can be generated are well suited for next generation surveys, like EUCLID and LSST. An important step in designing a hydro-simulation with such a goal is to ensure that the sub-grid models reproduce the baryonic observables that have a large influence on the matter power spectrum, cluster counts, galaxy-galaxy clustering, etc. The method that will be described, uses Gaussian process machine learning to tune the sub-grid models of our hydro-simulation to match two key observables, the stellar mass function and the gas fraction in massive halos (M>10^{13.5} Msun). We train the emulators with a latin hypercube consisting of 100 simulations in which the five most important sub-grid parameters are varied. This processes leads us to be able to emulate the SMF to a precision of 2% and the gas fractions to a precision of 5%. We then use MCMC to sample emulated parameter space to find the best-fit models compared with observations. We then test these parameters by running a simulation with the found parameters. The resulting simulations closely match the chosen observables.
