UCLA

Cosmological weak lensing probes can inform us of the contents and evolution of the universe, including the properties of dark matter and dark energy, which collectively make up $\sim 95\%$ of the universe. We live in an exciting period in scientific history; large scale astronomical surveys such as the Legacy Survey of Space and Time (LSST) will soon provide imaging for over a billion celestial objects, which timely coincides with recent advancements in probabilistic image-based machine learning. It is incumbent on scientists to leverage recent advancements to extract as much information as possible from large scale astronomical surveys to probe our universe. This thesis contains my contribution toward this objective.

Precision cosmological measurements require accurate data analysis with precise uncertainties. The two critical data analysis tasks for weak lensing cosmological probes are 1) photometric redshift (photo-z) estimation and 2) galaxy shear estimation. These quantities allow us to map the distribution of galaxies in the sky and quantify the distribution of dark matter. Here we present results for photo-z estimation and galaxy shape estimation using probabilistic neural networks, using a novel dataset derived from the Hyper Suprime-Cam (HSC) Survey.

In Chapter 1, we provide an introduction to weak lensing cosmological probes, photo-z estimation, and shear estimation. In Chapter 2, we introduce the machine-learning-ready dataset derived from HSC consisting of galaxy photometry, galaxy images, and spectroscopic redshifts. We make this dataset publicly available and utilize it for all photo-z estimation analyses in this work. In Chapter 3, we present a probabilistic photo-z estimation model using a Bayesian neural network (BNN) and compare its performance to alternative methods. In Chapter 4, we present an image-based probabilistic photo-z estimation model using a Bayesian convolutional neural network (BCNN) and compare its performance to alternative methods. In Chapter 5, we present an image-based probabilistic model for galaxy ellipticity estimation (as a proxy for shear estimation) evaluated on HSC galaxy images using a custom BCNN. In the Appendix we provide a roadmap by which one can utilize the photo-z and potential shear estimation models in this thesis to perform a weak lensing measurement.