2025MNRAS.540.2081F

We present the results of a Machine Learning study to measure the dust content of galaxies observed with JWST at z > 6 through the use of trained neural networks based on high-resolution IllustrisTNG simulations. Dust is an important unknown in the evolution and observability of distant galaxies and is degenerate with other stellar population features through spectral energy fitting. As such, we develop and test a new spectral energy distribution (SED)-independent Machine Learning method to predict dust attenuation and sSFR of high redshift (z > 6) galaxies. Simulated galaxies were constructed using the IllustrisTNG model, with a variety of dust contents parametrized by E(B-V) and A(V) values. These simulated galaxies were then used to train Convolutional Neural Network (CNN) models using supervised learning through a regression model. We demonstrate that within the context of these simulations, our single and multiband models are able to predict dust content of distant galaxies to within a 1$\sigma$ dispersion of A(V) $\sim 0.1$. On spectroscopically confirmed z > 6 galaxies from JADES and CEERS programmes, our models predicted attenuation values of A(V) < 0.7 for all systems, with a lower average [A(V) = 0.28]. The predictions of dust attenuation values that have an average error of 0.26 ($\sigma$ = 0.36) larger than SED fitted values, but for star formation an average error of 0.18 ($\sigma$ = 0.2) smaller. Both results show that distant galaxies at $z > 6$ with confirmed spectroscopy are not very dusty, although this sample is potentially biased. We discuss these issues and present ideas on how to accurately measure dust features at the highest redshifts using a combination of Machine Learning and SED fitting.