PineTree: A generative, fast, and differentiable halo model for wide-field galaxy surveys

¹ Sorbonne Université, CNRS, UMR 7095, Institut d’Astrophysique de Paris, 98 bis boulevard Arago, 75014 Paris, France
² The Oskar Klein Centre, Department of Physics, Stockholm University, Albanova University Center, SE 106 91 Stockholm, Sweden

^★ Corresponding author; simon.ding@iap.fr

Received: 2 July 2024
Accepted: 1 August 2024

Abstract

Context. Accurate mock halo catalogues are indispensable data products for developing and validating cosmological inference pipelines. A major challenge in generating mock catalogues is modelling the halo or galaxy bias, which is the mapping from matter density to dark matter halos or observable galaxies. To this end, N-body codes produce state-of-the-art catalogues. However, generating large numbers of these N-body simulations for big volumes, especially if magnetohydrodynamics are included, requires significant computational time.

Aims. We introduce and benchmark a differentiable and physics-informed neural network that can generate mock halo catalogues of comparable quality to those obtained from full N-body codes. The model design is computationally efficient for the training procedure and the production of large mock catalogue suites.

Methods. We present a neural network, relying only on 18 to 34 trainable parameters, that produces halo catalogues from dark matter overdensity fields. The reduction in network weights was realised through incorporating symmetries motivated by first principles into our model architecture. We trained our model using dark-matter-only N-body simulations across different resolutions, redshifts, and mass bins. We validated the final mock catalogues by comparing them to N-body halo catalogues using different N-point correlation functions.

Results. Our model produces mock halo catalogues consistent with the reference simulations, showing that this novel network is a promising way to generate mock data for upcoming wide-field surveys due to its computational efficiency. Moreover, we find that the network can be trained on approximate overdensity fields to reduce the computational cost further. We also present how the trained network parameters can be interpreted to give insights into the physics of structure formation. Finally, we discuss the current limitations of our model as well as more general requirements and pitfalls of approximate halo mock generation that became evident from this study.