Dissertation

Programming the Cosmic Time Machine: New Ideas and Applications of Machine Learning for Stage-IV Spectroscopy and Beyond

Green, Dylan Andrew

2025

Overview

The nature of Dark Energy is one of the most significant and compelling unsolved problems in physics today. Stage-IV spectroscopic surveys like the Dark Energy Spectroscopic Instrument (DESI) aim to try probe this unknown sector of physics using observables like Baryon Acoustic Oscillations (BAO). DESI recently released its Year Three results, with some of the tightest constraints on the composition of the universe recorded to date.DESI uses a state of the art software and data processing pipeline that makes significant use of a variety of machine learning algorithms. I present in this work a selection of four different machine learning and algebraic techniques developed for use as part of the DESI survey. The first algorithm applies deep learning to the problem of cosmic ray identification and rejection. Cosmic rays are a persistent source of error when extracting spectroscopy from raw CCD images and thus necessitate an accurate and fast algorithm for detection and masking.The second algorithm is QuasarNET, a deep convolutional neural network designed to automatically classify quasars in raw DESI spectra. QuasarNET also estimates a quasar redshift for each quasar identified. In my work we retrain QuasarNET using an algorithm called Active Learning, which automatically determines which unlabeled spectra would be beneficialto label. We then use those newly labeled spectra as a training dataset for QuasarNET, in the process discovering and later fixing a systemic problem with QuasarNET redshift estimates. This new weights file was used for DESI Year Three analysis and will be used in Year Five and beyond.The last two algorithms are centered on linear algebra and matrix decomposition. I present new unified coaddition scheme called “Bayesian Coaddition” that unifies three different coadds into a single Bayesian likelihood with a single hyper parameter prior. This work includes a coadd that reconstructs the true image without any telescope transmission effects, as well as a coadd with a diagonal covariance that is the statistically optimal way to weight exposures with different seeing when coadding. Finally I present an algorithm for non-negative matrix factorization that generates non-negative decomposed matrices of templates and coefficients that does not require the input data to be non-negative.

• Publisher: ProQuest Dissertations & Theses
• Subject: Artificial intelligence / Astronomy / Astrophysics / Physics
• ISBN: 9798291538906