Populating Large N-body Simulations with LRGs Using Neural Networks

The analysis of state-of-the-art cosmological surveys like the Dark Energy Spectroscopic Instrument (DESI) survey requires high-resolution, large-volume simulations. However, the computational cost of hydrodynamical simulations at these scales is prohibitive. Instead, dark matter (DM)-only simulations are used, with galaxies populated a posteriori, typically via halo occupation distribution (HOD) models. While effective, HOD models are statistical in nature and lack full physical motivation. In this work, we explore using neural networks (NNs) to learn the complex, physically motivated relationships between DM haloes and galaxy properties. Trained on small-volume, high-resolution hydrodynamical simulations, our NN predicts galaxy properties in a larger DM-only simulation and determines which galaxies should be classified as luminous red galaxies (LRGs). Comparing the original LRG sample to the one generated by our NN, we find that, while the subhalo mass distributions are similar, our NN selects fewer low-mass subhaloes as LRG hosts, possibly due to the absence of baryonic feedback effects in DM-only simulations. This feedback could brighten or redden galaxies, altering their classification. Finally, we generate a new LRG sample by fitting an HOD model to the NN-generated LRG sample. We verify that both the HOD- and NN-generated samples preserve a set of bias parameter relations, which assume that the higher-order parameters, $b_{s2}$ and $b_{3\rm{nl}}$, are determined by the linear bias parameter $b_{1}$. These relations are commonly used to simplify clustering analyses.