UC Irvine

Buried among millions of galaxies, large sky surveys embed strong gravitational lenses essential to cosmological studies. For instance, individual lenses have been used to study the properties of dark matter and early-type galaxies. Lensed supernovae studies have resulted in model-independent constraints on the Hubble constant and the spatial curvature. If a statistically significant number of lenses in sky surveys are discovered and studied, some of the persisting mysteries of cosmology such as the Hubble tension and the cosmic curvature may be resolved. Toward this end, in the first part of this thesis, I will demonstrate deep learning can recover lenses identified by, or missed by, other methods. This proof of concept was performed with a deep neural net with only three convolutional blocks and was applied to the HST/ACS i-band observations of the COSMOS field, a well-studied region of the sky. In the second half, I present a set of fully developed computational techniques as a ready-to-use software package for lens mining, or for mining other rare events in large sky surveys using more sophisticated deep learning methods. Inspired by DenseNet, I have designed a simple and easily modifiable neural net, with more than 50 convolution blocks, which with the help of a Generative Adversarial Neural Network (GAN) augmentation can learn from a mere 50 examples. After training on 50 known lenses, I was able to identify 42 lens candidates in the Subaru/HSC-PDR1. For the first time, I provide the selection function for a lens mining algorithm using a fast and realistic lens synthesizer. This is a crucial step in gravitational lens studies allowing us to provide a useful link between observational catalogs and theoretical cosmological predictions via counterbalancing the inevitable biases of the selection process.