Euclid preparation. LXVIII. Extracting physical parameters from galaxies with machine learning

The Euclid mission is generating a vast amount of imaging data in four broadband filters at high angular resolution. This will allow the detailed study of mass, metallicity, and stellar populations across galaxies, which will constrain their formation and evolutionary pathways. Transforming the Euclid imaging for large samples of galaxies into maps of physical parameters in an efficient and reliable manner is an outstanding challenge. We investigate the power and reliability of machine learning techniques to extract the distribution of physical parameters within well-resolved galaxies. We focus on estimating stellar mass surface density, mass-averaged stellar metallicity and age. We generate noise-free, synthetic high-resolution imaging data in the Euclid photometric bands for a set of 1154 galaxies from the TNG50 cosmological simulation. The images are generated with the SKIRT radiative transfer code, taking into account the complex 3D distribution of stellar populations and interstellar dust attenuation. We use a machine learning framework to map the idealised mock observational data to the physical parameters on a pixel-by-pixel basis. We find that stellar mass surface density can be accurately recovered with a $\leq 0.130 {\rm \,dex}$ scatter. Conversely, stellar metallicity and age estimates are, as expected, less robust, but still contain significant information which originates from underlying correlations at a sub-kpc scale between stellar mass surface density and stellar population properties.