Flatiron Seminar Series: Francisco Villaescusa

Title: Simulating universes. Exploring them with AI

Abstract: Cosmology is a branch of astrophysics dedicated to the study of the origin, fate, laws, and constituents of the Universe. To provide answers to these questions, cosmologists collect data from surveys that sample vast regions of the Universe. An important question is: What is the optimal way to extract information from that data? In this seminar, I will show how cosmologists have been extracting information from cosmic surveys from a historical perspective. First, I will describe the classical, pen-and-paper, calculations the community has employed for decades. Next, I will illustrate the limitations of such an approach and discuss how cosmological numerical simulations can provide theoretical predictions where analytical calculations break down. From the analysis of these simulations, the community has found that the structures in the Universe on small (cosmological) scales contain a wealth of information about the laws and constituents of the Universe. This has served as a motivation to look at the Universe on smaller scales. However, as we push to smaller and smaller scales, we need to worry about the impact of uncertain astrophysical processes, such as the energy released by supernovae and supermassive black-holes, on the theoretical predictions. I will then present the CAMELS project, which contains the largest set of cosmological hydrodynamic simulations ever created. I will show how this petabyte-large collection of Universes was created to combine machine learning with state-of-the-art hydrodynamic simulations to push the frontiers of cosmology on small scales. Finally, I will present the DREAMS project, whose goal is to learn about the nature and properties of dark matter by combining particle physics, astrophysics, and machine learning. During the talk, I will illustrate the potential of combining deep learning with massive sets of numerical simulations, showing applications ranging from using SE(3)-invariant graph neural networks to deriving universal equations governing cosmic systems.