Speaker
Description
Understanding how galaxies form and evolve is essential for advancing our knowledge of cosmic structure formation. However, generating large ensembles of mock galaxy catalogs through traditional semi-analytic models (SAMs) remains computationally prohibitive, especially when exploring wide ranges of physical parameters. In this work, we present a graph-neural-network-based emulator that efficiently reproduces SAM outputs by learning the mapping from dark matter halo merger trees and model parameters to galaxy properties.
Our framework is trained on large-scale data derived from the UchuuMicro simulation (with a 100 Mpc/h box size) and conditioned on outputs from the Galacticus SAM spanning 10,000 distinct parameter sets. The core innovation lies in a shared GNN backbone that effectively captures hierarchical and temporal information from complex merger trees. Crucially, we integrate the physical SAM parameters using a Feature-wise Linear Modulation (FiLM) mechanism, allowing the network to condition its feature learning on diverse physical priors. A multi-headed decoder simultaneously predicts multiple galaxy properties, including stellar mass, star formation rate, gas metal abundance, and black hole mass.
In an independent test set containing 600 SAM realizations (each generating 10,000 galaxies), the emulator reproduces the SAM stellar mass predictions with an average scatter of ~0.33 dex, and emulates all six galaxy properties for the entire 6 million-galaxy test set in just 9 seconds, an orders-of-magnitude acceleration over traditional approaches. This framework provides a fast and physically informed avenue for generating realistic galaxy catalogs and accelerating cosmological modeling pipelines.
