Reconstructing galaxy star formation histories from COSMOS2020 photometry using simulation-based inference

Abstract: We propose a novel method to reconstruct the full posterior distribution of the star formation histories (SFHs) of galaxies from broad-band photometry. Our method combines simulation-based inference (SBI) using a neural network trained with SFHs and photometry from the {\sc Horizon-AGN} hydrodynamical cosmological simulation. We apply it to reconstruct SFHs using COSMOS2020 photometry at redshift $0<z\<3$. We are able to accurately estimate the SFH and quantify the uncertainty on simulated data, with an unbiased posterior mean and well calibrated credible intervals. Our SFHs are in broad agreement with independent literature measurements The SFHs of galaxies as a function of location in the $\mathrm{NUV}-r$ versus $r-J$ diagram are in agreement with expectations. We extract summary statistics to quantify the shape of the SFH, number of peaks, and formation redshift. The slopes of the SFHs of passive galaxies show only a weak trend with stellar mass at $z\<1.35$ but a significant scatter, indicating that other factors than mass could drive the suppression of star-formation. Nevertheless, star-forming galaxies show a clear mass-dependent SFH, with lower-mass galaxies undergoing more vigorous recent star-formation. Low-mass galaxies have more peaks in their mass assembly histories than high-mass ones. At a given mass, we find many different formation redshifts, but for passive galaxies the mass dependency of formation redshifts is weak. Most passive galaxies with stellar mass $\log M_*/M_\odot > 9$ had a first event of mass assembly around $z\sim 3$, independent of mass. This work represents a pilot study for the future analysis of the \textit{Euclid} Deep fields that will reach similar depths in alike set of photometric bands, but with over an order-of-magnitude larger area, opening the possibility of deriving SFHs for millions of galaxies in a robust manner.