Deep Learning generated observations of galaxy clusters from dark-matter-only simulations

Hydrodynamical simulations play a fundamental role in modern cosmological research, serving as a crucial bridge between theoretical predictions and observational data. However, due to their computational intensity, these simulations are currently constrained to relatively small volumes. Therefore, this study investigates the feasibility of utilising dark matter-only simulations to generate observable maps of galaxy clusters using a deep learning approach based on the U-Net architecture. We focus on reconstructing Compton-y parameter maps (SZ maps) and bolometric X-ray surface brightness maps (X-ray maps) from total mass density maps. We leverage data from $\text{textsc}TheThreeHundred$ simulations, selecting galaxy clusters ranging in mass from ${10}^{13.5}{h}^{-1}{M}_{\odot }\le M_{200}\le {10}^{15.5}{h}^{-1}{M}_{\odot }$. Despite the machine learning models being independent of baryonic matter assumptions, a notable limitation is their dependency on the underlying physics of hydrodynamical simulations. To evaluate the reliability of our generated observable maps, we employ various metrics and compare the observable-mass scaling relations. For clusters with masses greater than $2\times {10}^{14}{h}^{-1}{M}_{\odot }$, the predictions show excellent agreement with the ground-truth datasets, with percentage errors averaging (0.5 $\pm$ 0.1)% for the parameters of the scaling laws.