The multi-dimensional halo assembly bias can be preserved when enhancing halo properties with HALOSCOPE

Abstract: Over $90$\% of dark matter haloes in cosmological simulations have unresolved properties. This can hinder the dynamical range of simulations and result in systematic biases when modelling cosmological tracers. We aim to more precisely determine unresolved structural and dynamical halo properties while preserving the correlations with environment and halo assembly bias found in simulations. We have developed HALOSCOPE, a machine learning technique that uses multi-variate conditional probability distribution functions. This method ensures that correlations among various halo properties, as well as their dependence on the local environment, are preserved. In this work, we trained HALOSCOPE with a high-resolution (HR) simulation and used it to better determine the properties (concentration, spin, and two shape parameters) of unresolved dark matter haloes in an eight times lower resolution simulation. HALOSCOPE is able to recover the multi-dimensional halo assembly bias, that is, the correlations of different combinations of halo properties with the large-scale environment, measured in the HR simulation. This is achieved by including the linear halo-by-halo bias and tidal anisotropy in the set of input training parameters. HALOSCOPE, by design, also recovers the joint distribution of the halo properties. To study how resolution effects propagate into the clustering of model galaxies, we generated catalogues of central galaxies using two implementations of the assembly bias in a halo occupation distribution model. The clustering of central model galaxies is improved by a factor of three at $0.009<k ({\rm Mpc}^{-1}h)<0.6$ when the unresolved haloes are enhanced with HALOSCOPE. HALOSCOPE can improve the accuracy of cosmological tracer catalogues produced with approximate methods when many realisations are needed.