From dark to light: The connection between galaxies and their dark matter haloes through cosmic time

• Applicant: Dr. Benjamin Moster
• Subject Area: Astrophysics and Astronomy
• Term: from 2015 to 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 274846616

The formation and evolution of galaxies ranks among the most exciting and promising areas of physical sciences. How do galaxies form and what determines their properties? What can galaxies teach us about the nature of the Universe and its main components such as dark matter and dark energy? To answer these central questions we need to compare observations of stars and gas with the prediction of cosmological models. Empirical galaxy formation models provide a unique and direct link between galaxies and dark matter haloes, and do not depend on model assumptions on baryonic physics, only on gravity and observed galaxy properties. During the first phase of the Emmy Noether project we have developed the state-of-the-art empirical galaxy formation model EMERGE, which follows the evolution of individual galaxies by tracing the growth histories of their haloes as extracted from numerical simulations, and uses empirical relations to derive galaxy properties such as star formation rate and gas mass. The main difficulty in this approach is finding suitable parameterised relations between galaxy and halo properties. In this project, we will utilise the enormous advances that the field of artificial intelligence has made over the last years, and employ machine learning methods that are able to detect patterns like the connection between galaxy and halo properties automatically. To this end, we will develop a deep neural network that takes the properties of a dark matter halo as input values, and computes the properties of the galaxy in the centre of the halo. The network will first be trained in a supervised manner using the results of EMERGE. In the next step, we will train the neural network directly on observed data, using a reinforcement learning approach. For a set of network weights, we will compute mock observations like stellar mass functions, cosmic star formation rate densities and clustering, and compare them to the observed data with a loss function. We will then use particle swarm optimisation and evolutionary algorithms to minimise the loss function, and choose the weights such that the observed data are reproduced. Once the network has been trained, we will extract suitable parameterisations between galaxy and halo properties and implement them in EMERGE so that the evolution of galaxies can be followed self-consistently. Finally, we will make the trained neural networks available to the community. In this way, the scientific achievements of our Emmy Noether group can be easily accessed and used, providing a legacy.

• DFG Programme: Independent Junior Research Groups