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

¹ CNRS and Sorbonne Université, UMR 7095, Institut d’Astrophysique de Paris, 98 bis, Boulevard Arago, F-75014 Paris, France
² Aix Marseille Univ., CNRS, CNES, LAM, Marseille, France

^⋆ Corresponding author.

Received: 1 October 2024
Accepted: 28 April 2025

Abstract

We propose a novel method for reconstructing the full posterior distribution of the star formation histories (SFHs) of galaxies from broad-band photometry. Our method combines the simulation-based inference (SBI) framework using a neural network trained with SFHs and photometry from the HORIZON-AGN hydrodynamical cosmological simulation. We applied our technique for reconstructing SFHs in the COSMOS Treasury field using only COSMOS2020 photometry in the redshift range 0<z<3. The method is able to accurately estimate the SFH and quantify the Bayesian uncertainty on simulated data, with an unbiased posterior mean, σ_err≤0.16 dex for all formation times and properly calibrated posterior intervals. Our SFHs broadly agree with literature measurements derived by different methods using combined photometric and spectroscopic datasets. The SFHs of galaxies as a function of location in the near-UV−r versus r−J colour-colour diagram agree in general with expectations. They vary smoothly from star-forming to passive and quiescent galaxies that are properly localised in the red part of the diagram. We extracted summary statistics to quantify the shape of the SFH, the number of peaks, and the 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 factors other than mass might drive the suppression of star-formation. Nevertheless, star-forming galaxies show a clearly mass-dependent SFH, with lower-mass galaxies undergoing more vigorous recent star-formation. Overall, the SFH slopes in COSMOS vary over a wider range than in HORIZON-AGN. Low-mass galaxies have more peaks in their mass assembly histories than high-mass galaxies, and the trend is clearer in COSMOS than in HORIZON-AGN. At a given mass, we find many different formation redshifts, but the mass dependence on the formation redshifts is weak for passive galaxies. Most passive galaxies with a stellar mass log M_*/M_⊙>9 had a first event of mass assembly around z∼3 (2.2<z<5.8), regardless of their mass. This work represents a pilot study for the future analysis of the Euclid Deep fields that will reach similar depths in a similar set of photometric bands, but with an area that is larger by more than an order of magnitude. This opens the possibility of deriving SFHs for millions of galaxies in a robust manner.