- Main
Model-Driven Cosmology With Bayesian Machine Learning and Population Inference
- Ho, Ming-Feng
- Advisor(s): Bird, Simeon
Abstract
This thesis presents new directions for cosmological data analysis using Bayesian techniques and machine learning.
First, I introduce a novel machine learning spectroscopic analysis technique to detect absorption systems in the Lyman-α forest. Using Gaussian processes, I build data-driven models for the quasar emission and apply Bayesian model selection to classify the damped Lyman alpha absorbers (DLAs), which are high column density absorption systems found in quasar spectra. The Gaussian process DLA finder (GP-DLA) is applied to the Sloan Digital Sky Survey (SDSS) quasar spectra and and now adopted by the Dark Energy Spectroscopic Instrument (DESI) collaboration. This GP-DLA technique allows us to construct probabilistic catalogs of damped Lyman-α absorbers, offering a new approach to studying the intergalactic medium and cosmology at z = 2 - 5.
Next, I present a new method to infer cosmological parameters using Bayesian surrogate modeling with multi-fidelity emulators. Multi-fidelity emulators are a type of surrogate model that use information from multiple levels of fidelity to improve the accuracy of the surrogate model. This approach accelerates both the analysis of cosmological simulations and the inference of cosmological parameters, providing a probabilistic method to quantify and correct the resolution in cosmological simulations. Multi-fidelity emulators make it possible to perform fast and accurate parameter inference on large-scale structure data, such as the matter power spectrum, using computationally expensive simulations in high-dimensional parameter spaces.
Finally, I discuss population inference of gravitational wave (GW) data using a mixture model approach. The population statistics of GW events can provide insights into the formation and evolution of binary black holes (BBHs). I present a data-driven method to infer the mixing fraction between BH populations, along with a Bayesian hierarchical approach to correct for selection effects. The results of the mixing fraction analysis suggest that the population of 35 M⊙ BHs is likely separate from the rest of the population, indicating that current formation channels for this mass bump need to be revised to include explanations for the separation of these massive 35 M⊙ binaries.
Main Content
Enter the password to open this PDF file:
-
-
-
-
-
-
-
-
-
-
-
-
-
-