Title: Stellar Masses of Giant Clumps in CANDELS and Simulated Galaxies Using Machine Learning

Abstract: A significant fraction of high redshift star-forming disc galaxies are known to host giant clumps, whose nature and role in galaxy evolution are yet to be understood. In this work we first present a new method based on neural networks to detect clumps in galaxy images. We use this method to detect clumps in the rest-frame optical and UV images of a complete sample of $\sim1500$ star forming galaxies at $1<z<3$ in the CANDELS survey as well as in images from the VELA zoom-in cosmological simulations. We show that observational effects have a dramatic impact on the derived clump properties leading to an overestimation of the clump mass up to a factor of 10, which highlights the importance of fair comparisons between observations and simulations and the limitations of current HST data to study the resolved structure of distant galaxies. After correcting for these effects with a mixture density network, we estimate that the clump stellar mass function follows a power-law down to the completeness limit ($10^{7}$ solar masses) with the majority of the clumps being less massive than $10^9$ solar masses. This is in better agreement with recent gravitational lensing based measurements. The simulations explored in this work overall reproduce the shape of the observed clump stellar mass function and clumpy fractions when confronted under the same conditions, although they tend to lie in the lower limit of the confidence intervals of the observations. This agreement suggests that most of the observed clumps are formed in-situ.