AbstractThis thesis focuses on applying machine learning methods to both astronomical and cosmological problems. Regarding the application to astronomy, I use data analysis techniques new to astronomy to detect strong correlations in observed data to perform feature pre-selection, machine learning techniques (four tree-based methods including Random Forests) to classify astronomical objects, and novel software packages to interpret a machine learning model in an attempt to understand how it is correctly classifying objects. I showcase these techniques by applying them to the problem of star-galaxy separation using data from the Sloan Digital Sky Survey (hereafter SDSS) and the results show that the rate of misclassifications can be reduced by up to ≈ 33% over the standard SDSS frames approach.
In reference to the application to cosmology, I seek to answer the question: 'can we distinguish between cosmological/gravitational models using machine learning, and if so, what features are useful discriminants?'. To approach this, I use an image classi cation machine learning method called Convolutional Neural Networks (CNNs) to classify dark matter particle simulations created with different theories of gravity. The results show that these simulations can be classified to a high degree of accuracy. I then investigate the model, using generated datasets with known parameters to probe the decision boundaries of the CNN and determine where the model breaks down. I also manipulate the CNN into creating representations of dark matter particle simulations to understand which features of the simulations it has been able to learn about - showing that CNNs do not have to simply focus on matter density variance and can learn about higher order statistics such as isodensity curvature. All of these methods are new to the analysis of different theories of gravity in cosmology.
|Date of Award||Mar 2019|
|Supervisor||David Bacon (Supervisor) & Kazuya Koyama (Supervisor)|