CD&SE: Optimal Field Level Simulation Based Inference in cosmology

Artificial Intelligence (AI) applications to cosmological observations could provide significant improvements to our understanding of the universe. However, when it comes to analysis of real data there have been very few successes; real data contain many additional complications that need to be accounted for. This research program will develop AI methods that are sufficiently realistic to be applied to actual cosmological data of the next generation of surveys.

The research team will focus on weak lensing surveys, where we learn about the universe from the distortions of the galaxy images, which allows us to trace the dark matter through its gravitational effect on the galaxy images. The research program will engage a diverse group of students, and the lessons learned will be incorporated in the curriculum of Physics and Data Science education programs.

One of the promising AI venues is Simulation Based Inference, which aims to extract information from the data via simulations. The most common approach is discriminative learning, which first summarizes the information in the data via data compression, followed by their emulation to obtain the cosmological parameters of interest. An alternative is to learn the field level data likelihood directly using generative learning.

The goal of this proposal is to develop and compare these AI methods and apply them to actual weak lensing data of the next generation. The investigators will pursue Multiscale Normalizing Flows in the context of generative learning, and compare to Neural Network based compression architectures in the context of discriminative learning, to develop optimal methods for upcoming weak lensing surveys in terms of precision, reliability and robustness.

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.