CNN‑Enhanced Radiative Transfer Emulation for Dust Attenuation in High‑Redshift Galaxies

Anisotropic scattering by the dust couples all lines of sight, and dust absorption/emission couples all wavelengths, making the radiative transfer equation highly non-local and non-linear (Steinacker et al. 2013). Modern radiative transfer codes have paved the way for synthetic observations, leading to a robust comparison of observations and simulations. Slowly, these codes are also being coupled with leading hydrodynamical simulations of galaxies to better model the interplay of gas, dust, and radiation. More accurate modeling of radiative heating, cooling, and ionization changes will lead to more constrained star formation and gas dynamics. With all its benefits it will provide, radiative transfer still remains very computationally expensive to run. We present our approach of harnessing CNN’s to use the results of dust RT to modify the results of cheaper approaches, such as a Line of Sight (LoS) model (Vijayan et al. 2021), to faster generate reasonably accurate attenuation maps compared to traditional RT modeling. LoS models calculates the effect of dust attenuation by calculating the intervention of gas and dust particles in front of the light-emitting particle by relating optical depth with metal column density. Using the SKIRT radiative transfer code on First Light And Reionization Simulations’ (FLARES) galaxies at the Epoch of Reionisation (EoR, z>5), we generated multi-wavelength photometry of massive galaxies (≳ 10^9M⊙). We also produce LoS model photometry at the same resolutions. The neural networks map the output of the cheaper (LoS) model to the RT outputs to calculate and generate the impact of scattering events.