Master Thesis Defense by Gustav Madsen
Title: Data-driven reconstruction of cluster mass profiles using convolutional and graph neural networks
Abstract:
This thesis aims to predict the mass profiles of galaxy clusters from the projected phase-space coordinates of galaxy clusters, through the implementation of convolutional and graph neural networks. The neural networks are trained, validated and tested on data from the dark matter N-body Uchuu simulation. Uncertainties on the network predictions are quantified using Monte Carlo dropout, a Bayesian approximation of the Gaussian process. In the majority of cases, both networks are able to predict the ground truth mass profiles of galaxy clusters within the uncertainties. The effect of interloper contamination is confirmed to have a noticeable effect on the predicted mass profiles, effectively degrading the accuracy of reconstruction. By applying the two-point angular correlation function to individual clusters from the Uchuu simulation, a complex relationship between the cluster mass, degree of clustering, cluster ellipticity and the quality of the predicted mass profiles is established. However, no clear conclusion is drawn here. A novel method involving the use of the splashback radius of galaxy clusters to infer the mass accretion rate is explored. The splashback radius is confirmed to be present in the ground truth mass profiles from the Uchuu simulation, as well as in the predicted mass profiles from both networks, allowing for the mass accretion rate of galaxy clusters to be estimated. Subsequently, the trained networks are applied to 7 clusters from the SDSS catalogue, 6 of these being Abell clusters with the remaining one being the Coma cluster. Half of the predictions involving the Abell clusters, i.e. 3 cluster mass profiles, are found to be in agreement with measurements from the literature on a 1σ level. The predicted mass profile of Coma cluster from both networks produce state-of-the-art results, agreeing with the majority of measurements from the literature to a 1σ level, and a 2σ level at most, for the remaining measurements. Evidence of the splashback radius is present as well in the predicted mass profiles of the 7 SDSS clusters. However, the mass accretion rate can only be estimated for a single one of these 7 clusters. As such, this method of estimating the mass accretion of galaxy clusters should as of now be used with caution.
Supervisor:
- Radoslaw Wojtak & Jens Hjorth, University of Copenhagen, Niels Bohr Institute
Censor:
- Hans Kjeldsen, Aarhus University
¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤¤
ZOOM-link for participants whom are not able to attend at Auditorium C