Thesis

Machine-Learned Statistical Inference on Cosmological Survey Data

• Call:

  PT-CERN Call 2020/2

• Academic Year:

  2020/2021

• Domain:

  Astroparticle Physics

• Supervisor:

  Andrew Liddle

• Co-Supervisor:

  Antonio da Silva

• Institution:

  FCUL (Universidade de Lisboa)

• Host Institution:

  IA - Instituto de Astrofísica e Ciências do Espaço

• Abstract:

  The overall objective is the development, implementation and deployment of novel algorithms for statistical analysis of large cosmological survey datasets, with particular emphasis on machine/deep learning innovations emerging in the literature. We propose the application of state-of-the-art machine-learning techniques to address the following most pressing topics of modern cosmology, via model-level and model-independent inferences. One initial direction will be the development of machine-learned sampling for Bayesian evidence calculations in cases where likelihood evaluations are extremely costly, as is increasingly becoming the case with large-scale structure data. A second direction is the use of symbolic regression to seek model-independent information on Dark Matter, Dark Energy and Inflation directly from survey data, building on the work of Schmidt, Udresco, Tegmark and others [SCHMIDT, UDRESCO, LIU]. A third direction is to use the obtained results and developed pipelines to address the tensions [LAHAV] between: 1. Measurements of the Hubble constant (Planck vs. SH0ES + H0licow datasets) via early and late Universe probes: 4 to 5 sigma mismatch; 2. Measurements of the amplitude of perturbations via CMB radiation and matter power spectra (Planck vs. DES + KiDS): 2 to 3 sigma mismatch; A considerable degree of a-priori model assumption / bootstrapping is employed in existing analyses. We will use our model-independent approach to reconcile the different datasets, and further limit the model parameter space in order to shed light on the nature of Dark Matter, Dark Energy and Inflation.