Learning Baryonic Physics from Complete Merger Histories @Brown University

Efficiently mapping baryonic physics onto dark matter represents one of the major challenges of the current cosmological paradigm. Even as Semi-Analytical Models (SAMs) and hydrodynamical simulations have made impressive advances in reproducing key observables across cosmologically significant volumes, there is significant tension between the predictions of these two methods, and the increases in simulation volume has meant that simulation times are now of order $10^8$ CPU hours. However, with recent advances in Machine Learning (ML), key quantities of these simulations can now be reproduced in seconds. Graph Neural Networks (GNNs) have proven to be the natural choice for learning physical relations, and among the most inherently graph-like structures found in astrophysics are the merger trees that encode the evolution of dark matter haloes, which are used by SAMs to simulate baryonic physics. I’ll discuss how SAMs can be emulated precisely and quickly with a GNN for several different baryonic quantities of interest, and how this method offers key advantages over other ML methods in analyzing the interdependence between assembly history and the baryonic properties of galaxies.