RAISE: ADAPT : Novel AI/ML methods to derive CMB temperature and polarization power spectra from uncleaned maps

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Cosmic Microwave Background (CMB) radiation is one of the pillars of modern cosmology – it confirmed the standard theory of the Big Bang and helped reveal the structure and content of the universe. This project investigates the next major potential discovery from the CMB, the detection of primordial gravity waves (PGW). However, the masking of the CMB signal by contaminants is detrimental to high-precision measurements and in particular to the expected faint PGW signal. Extracting and analyzing the CMB signal from the overwhelming amounts of data expected from experiments such as CMB-Stage-4 presents challenges that require sophisticated new methods. The interdisciplinary team of computer scientists and astrophysicists at the University of Texas, Dallas, will develop and apply novel Machine Learning (ML) methods for this endeavor. Their methodology combines the predictive power of modern Deep Neural Networks (DNNs) with statistical tools to produce powerful and efficient models that incorporate domain expertise and respect known physical constraints. The team will complement this research with outreach efforts to promote and increase public engagement with science and technology within the Dallas-Fort Worth (DFW) area, including (1) organizing yearly workshops for cosmology and ML at the high-school teacher conference Mini-CAST, which is affiliated with the Science Teachers Association of Texas, and (2) actively participating in science camps and exchanges in low-socioeconomic communities as well as the broader DFW area to reach out and recruit students from underrepresented groups in STEM fields. This ADAPT RAISE project includes an amalgamation of expertise and joint efforts from astrophysicists and computer scientists that goes beyond a simple combination of the subjects and aims to transform each of them to provide a fast and scientifically informed ML model to deal with CMB contaminants and analysis. The team will first develop a novel method that produces CMB clean temperature and polarization power spectra directly from uncleaned maps. This comes from the realization that application of ML to CMB should not try to replicate the processing steps of traditional methods but rather take full advantage of what ML is exactly good at – extracting rich patterns from data. Second, the ML approach builds DNNs that incorporate soft scientific domain knowledge via statistical models to regularize and inform the model. An immediate consequence of this approach is that the CMB power spectra harmonic components can be used in the ML loss function allowing one to take full advantage of their physical and mathematical properties during the model training. While this investigation is focused on developing and applying DNN ML methods to the CMB, the tools and approaches developed here have far-reaching applications in sciences. 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.