Revealing the Hierarchical Structure of Galactic Haloes with Novel Data Mining Algorithms

Abstract

The research within this thesis is directed at developing, training, and testing unsupervised astrophysical clustering algorithms that extract meaningful structures from their input data. It is a well-studied consequence of the ΛCDM cosmological model of the Universe that these structures form hierarchically through the continual merging of smaller structures. As such, galaxies are expected to contain a myriad of substructure that act as fossil records of the galaxies themselves. As larger and more advanced surveys continue to be conducted, we are faced with the task of unearthing these galaxies and their substructures over a vast range of ever-more-complicated data sets. To tackle this issue, it is necessary to prepare ourselves with appropriate tools that can sift through these data sets and discover new structures. This is the goal that motivates the works within this thesis. First I developed Halo-OPTICS, a new algorithm designed to hierarchically classify astrophysical clusters within N-body particle simulations. I showed that it performs well against a current state-of-the-art code (e.g. VELOCIraptor) even though it uses comparatively less of the available information within the simulation data. Next I developed CluSTAR-ND and in doing so I made various algorithmic improvements upon its predecessor Halo-OPTICS. These upgrades dramatically improved CluSTAR-ND's computational footprint, its sensitivity to relevant clusters, and its capacity to operate over any size data set containing any number of dimensions. Finally, I developed CluSTARR-ND which boasts all the operational virtues of CluSTAR-ND while also providing an OPTICS-style representation of clustering structure and identifying clusters as statistically distinct overdensities of the input data. CluSTARR-ND therefore opens up the opportunity for adaptively providing a meaningful hierarchical astrophysical clustering of any n-point d-dimensional data set with an extremely modest computational demand.