Spatially Resolved Galaxy-Dust Modeling with Coupled Data-driven Priors

A notorious problem in astronomy is the recovery of the true shape and spectral energy distribution (SED) of a galaxy despite attenuation by interstellar dust embedded in the same galaxy. This problem has been solved for a few hundred nearby galaxies with exquisite data coverage, but these techniques are not scalable to the billions of galaxies soon to be observed by large wide-field surveys like the Legacy Survey of Space and Time, Euclid, and Roman. We present a method for jointly modeling the spatially resolved stellar and dust properties of galaxies from multiband images. To capture the diverse geometries of galaxies, we consider nonparametric morphologies, stabilized by two neural networks that act as data-driven priors: the first informs our inference of the galaxy’s underlying morphology, while the second constrains the galaxy’s dust morphology conditioned on our current estimate of the galaxy morphology. We demonstrate with realistic simulations of z ∼ 0 galaxies that we can recover galaxy host and dust properties over a wide range of attenuation levels and geometries. We apply our joint galaxy-dust model to three local galaxies observed by the Sloan Digital Sky Survey and find qualitatively good results. In addition to improving estimates of unattenuated galaxy SEDs, our inferred dust maps will facilitate the study of dust production, transport, and destruction. However, without informed priors on the inferred host spectrum, our method is vulnerable to the thin-screen limit; future work will need to address this degeneracy with stellar population synthesis modeling.