Subject Area

Physics

Abstract

Next-generation cosmic microwave background (CMB) surveys will map the microwave sky with unprecedented precision presenting a wide range of opportunities for new insights into cosmology. To maximize the scientific value of upcoming surveys, new methods that can disentangle primary and secondary CMB anisotropies are required. The CMB secondaries encode information about the cosmological sources that cause them and act as a source of noise in the observation of primary anisotropies. Quadratic estimators are the current standard techniques for reconstruction of the sources of CMB secondaries. While successful for reconstruction of distortion fields with current data, quadratic estimators will become sub-optimal for forthcoming data at the new expected level of sensitivity. In this dissertation, I will present a convolutional neural network, ResUNet-CMB, which is able to reconstruct multiple distortion fields simultaneously. I focus on three sources of CMB secondary anisotropies: gravitational lensing, patchy reionization, and cosmic polarization rotation. I also detail a process for using the machine learning reconstructions for the detection of primordial gravitational waves. I show that the ResUNet-CMB network significantly outperforms the quadratic estimator at low instrument noise levels and is not subject to an estimator bias from the presence of multiple distortion fields seen with a straightforward application of the quadratic estimator. Additionally, I show that when using the ResUNet-CMB reconstructions to reduce secondary B-modes, the reduction in B-mode spectrum power and the inferred tensor-to-scalar ratio match those of an estimated ideal iterative method.

Degree Date

Spring 5-13-2023

Document Type

Dissertation

Degree Name

Ph.D.

Department

Physics

Advisor

Dr. Joel Meyers

Second Advisor

Dr. Robert Kehoe

Third Advisor

Dr. Fredrick Olness

Fourth Advisor

Dr. Alexander Van Engelen

Format

.pdf

Creative Commons License

Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.

Available for download on Sunday, May 04, 2025

Share

COinS