Title:

Early universe cosmology and its observational effects on the cosmic microwave background

This Thesis is written in three parts. The first part describes the analytic calculation of the unequaltime correlator of cosmic strings and superstrings. The first efficient constraint analysis of all string and superstring network parameters is performed. By studying the effect of cosmic strings on the cosmic microwave background (CMB) radiation it is discovered that cosmic strings must make up a vanishingly small proportion of the energy density of the universe. The constraints on string network parameters are all skewed toward reducing the magnitude of energy density arising from strings. Also in this Part, a better comprehension of the unconnected segment model (USM) was gained. In particular, a greater understanding of the string scaling parameter $L_f$ was garnered, as well as finding the reason why the USM tends to provide greater power than simulations of NambuGoto cosmic strings. The second part contains a detailed description of statistical cosmology and how differences between parameter constraints from different data sets can lead to misleading quantification of discordance. The majority of this part describes different methods of quantifying differences between probability distributions and how these can be interpreted. In particular, using the most uptodate data possible, differences between parameter constraints using the CMB and probes of large scale structure (LSS) in the universe can be measured. With current data the discordance can be interpreted as a low level of disagreement, but the application of prior ranges on well known parameters can force the tension to be greater. Using data from earlier work, this issue is considered in greater detail, with extensions to the accepted LCDM model added to test if the discordance can be alleviated. These extensions include the addition of active or sterile neutrinos and even adhoc changes to the primordial power spectrum. Although there are slight hints that these may help, when considering only the new data it might be unwise to believe that the discordance between parameter distributions from different data sets exists to a degree where the modifications are necessary. Finally, application of deep learning to astrophysical observations is discussed. Using neural networks to learn about specific problems is de rigueur and their use in astronomy and cosmology is a promising field of study. In particular, applying raw data to neural networks can often outperform, or add enhanced features, to what is possible with current, nonempirical feature detection. The classification of supernovae from their light curves can be achieved using a specific machine learning architecture called a recurrent neural network (RNN). Using the raw data from supernova light curves, the RNN is able to learn about features in sequences which can be used to classify types of supernova. Although a large training set is needed to perform as well as current techniques, one major advantage the RNN method has is the possibility of early detection. Rather than needing the entire light curve to perform statistical fits to categorise the supernova type, relatively little information from the early observation data is needed to classify using the RNN. Installing RNN on machinery for observation would save a vast amount of time by early classification since only supernovae of interest can be concentrated on.
