CDS&E: Probabilistic modeling of fields and point clouds in cosmology

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Cosmological observations provide a unique window into fundamental physics, such as the ultra-high energy physics of the primordial universe. Upcoming galaxy surveys such as Rubin Observatory will produce vast amounts of data, but the extraction of fundamental physics from this data will be difficult. To fully exploit the statistical sensitivity of upcoming experiments, new methods which leverage high performance computing and machine learning must be developed. In this project scientists from the University of Wisconsin, Madison will make use of novel techniques from probabilistic machine learning and apply them to the dark matter and galaxy distribution of the universe. Using these tools, the team will be able to measure fundamental physics parameters from galaxy data more precisely than was previously possible. This research project will also provide exciting research opportunities for several undergraduate students from UW Madison's URS program, which supports and encourages students with non-traditional backgrounds, contribute to outreach efforts, and improve education in the important field of artificial intelligence for science. The scientific goal of this project is to develop normalizing flows to model two types of cosmological data, as well as their statistical connection: field level data, such as the non-linear matter field, and point cloud data such as halos and galaxies. For this task, the team will adapt recently developed normalizing flows for point clouds. Similar tasks appear in 3-dimensional computer vision and molecular design, but cosmology has unique properties that will feed into new machine learning developments. The team will design a scale decomposition to treat very large point clouds and will then use their normalizing flows for two important applications in cosmology. The first application will be to generate super-resolution simulations, where a conditional normalizing flow is used to augment the resolution of dark matter simulations, as well as to include baryonic physics. The second application will be to use the point cloud flow to establish a probabilistic dark matter to galaxy connection in a forward modeling framework, which will improve the reconstruction of cosmological initial conditions with respect to previous approaches. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.