Improving our view of the Universe using Machine Learning

Abstract

Increasingly ambitious photometric surveys have driven the field forward, but upcoming surveys are so complex (requiring active optical control) and sensitive (requiring deblending of overlapping galaxies) that current approaches are challenged. In this work we apply machine learning to both problems with promising results. The Legacy Survey of Space and Time (LSST) at the Vera C. Rubin Observatory (Rubin) will be deepest optical/infrared survey ever to cover the majority of the sky. LSST will provide new insights into the mysteries of dark matter and dark energy, as well as transient phenomena, solar system objects, and Milky Way structure. In order to achieve its scientific goals, the telescope must deliver atmospheric distortions limited image quality over the 3.5 degree field of view. In this thesis, we describe deep learning methods to optimize the performance of the Active Optics System of LSST by reducing the aberrations of the system. For decades, astronomers have scanned the sky with increasingly powerful telescopes and cameras, collecting millions of digital images of billions of galaxies. These ``sky surveys'' collect so much data that no human could ever examine all images directly, so the objects are characterized by automated software pipelines. Such pipelines work well for isolated objects, but survey images are often crowded with overlapping objects that must be "deblended." In this thesis, we address this problem with a conditional auto-encoder. This encoder characterizes galaxy images with a semantically meaningful latent space including the physical parameters of each galaxy in an image, and achieves nearly unbiased regression errors with a variance close to the statistical limit. We extend the conditional auto-encoder to be a galaxy deblender, which takes in the blended galaxy images, and reconstructs the deblended and denoised center galaxy and outputs its physical parameters.