Deep Learning for Dynamical Mass Estimation of Galaxy Clusters

Galaxy clusters are the most massive gravitationally bound systems in the universe, consisting of hundreds of luminous galaxies and hot gas embedded in dark matter halos. Due to their immense size, galaxy clusters are excellent tracers of the large-scale matter distribution of the universe and function as highly sensitive probes of gravity and the growth of structure. However, utilizing galaxy cluster abundance in precision cosmology requires large, well-defined cluster samples and precise, robust mass measurement methods.

This thesis consolidates the introduction, validation, and observational extrapolation of novel deep learning estimators of dynamical cluster mass. Using state-of-the-art computer vision models, we have defined new machine learning methodology for inferring masses from dynamical observables which mitigate classical sources of systematic error. We show that these models outperform traditional techniques by greater than a factor of two, and modern techniques by up to 30%. We have defined techniques for accurately reconstructing predictive uncertainties of these mass estimators such that we can confidently recover mass percentiles to within ±1% of their empirical value. With this method, we produce new mass posteriors on the Coma cluster and eight CLASH clusters, and define rigorous inference pipelines for future applications of machine learning models to observational data.