Title:Reconstruction of Continuous Cosmological Fields from Discrete Tracers with Graph Neural Networks

Abstract:We develop a hybrid GNN-CNN architecture for the reconstruction of 3-dimensional continuous cosmological matter fields from discrete point clouds, provided by observed galaxy catalogs. Using the CAMELS hydrodynamical cosmological simulations we demonstrate that the proposed architecture allows for an accurate reconstruction of both the dark matter and electron density given observed galaxies and their features. Our approach includes a learned grid assignment scheme that improves over the traditional cloud-in-cell method. Our method can improve cosmological analyses in situations where non-luminous (and thus unobservable) continuous fields need to be estimated from luminous (observable) discrete point cloud tracers.