A Machine Learning Approach to Correct for Mass Resolution Effects in Simulated Halo Clustering Statistics

The increase in the observed volume in cosmological surveys imposes various challenges on simulation preparations. Firstly, the volume of the simulations required increases proportionally to the observations. However, large-volume simulations are quickly becoming computationally intractable. Secondly, on-going and future large-volume survey are targeting smaller objects, e.g. emission line galaxies, compared to the earlier focus, i.e. luminous red galaxies. They require the simulations to have higher mass resolutions. In this work we present a machine learning (ML) approach to calibrate the halo catalogue of a low-resolution (LR) simulation by training with a paired high-resolution (HR) simulation with the same background white noise, thus we can build the training data by matching HR haloes to LR haloes in a one-to-one fashion. After training, the calibrated LR halo catalogue reproduces the mass-clustering relation for mass down to $2.5\times 10^{11}~h^{-1}M_\odot$ within $5~{\rm per~cent}$ at scales $k<1~h\,\rm Mpc^{-1}$. We validate the performance of different statistics including halo mass function, power spectrum, two-point correlation function, and bispectrum in both real and redshift space. Our approach generates high-resolution-like halo catalogues ($>200$ particles per halo) from low-resolution catalogues ($>25$ particles per halo) containing corrected halo masses for each object. This allows to bypass the computational burden of a large-volume real high-resolution simulation without much compromise in the mass resolution of the result. The cost of our ML approach ($\sim 1$ CPU-hour) is negligible compared to the cost of a $N$-body simulation (e.g. millions of CPU-hours), The required computing time is cut a factor of 8.