Neural-network-based Emulators for Accelerating Cosmological Inference

Student thesis: Doctoral Thesis


A key aspect of modern cosmology is cosmological parameter inference. This is a task involving determining the values of free parameters present in theoretical cosmological models using observations of the Universe. The three-dimensional clustering of galaxies and its compressed representation in the two-point correlation function or power spectrum (PS) is one such observation. Inferring cosmological parameters from the PS involves evaluating theoretical predictions for the PS for many different sets of cosmological parameters. These theoretical models can be computationally expensive to evaluate. This means that cosmological parameter inference is a task that requires significant resources. A method for reducing the computational cost is to replace the direct evaluation of the theoretical model with a more efficient emulator. This emulator often takes the form of a machine learning algorithm like a Gaussian process or neural network (NN) and is trained to replicate the predictions of the expensive theoretical model.
In this thesis, we present a suite of NN-based emulators that have been developed to predict components of theoretical models for the nonlinear galaxy PS. The first set in the suite is tailored to producing an emulator with a suite of numerical simulations in the halo model (HM) framework. A common technique for improving prediction accuracy when producing an emulator for the dark matter PS is to emulate the ratio of the simulation prediction to an analytic prediction. The analytic prediction itself comes with some non-negligible computational cost. As such, we produce emulators that predict several components of the analytic HM. We demonstrate that, when combining the component emulators for the analytic HM with a final component emulator for the nonlinear ratio, predictions for the nonlinear galaxy PS can be made in ∼ 0.1 s.
The next set of emulators has been developed for making predictions for the effective field theory of large-scale structure (EFTofLSS) PS multipoles. Each component emulator predicts a different set of redshift-space galaxy kernels of the EFTofLSS model. With the suite of emulators, predictions can be made over three orders of magnitude faster than those coming from direct evaluation of the EFTofLSS. We exploit the accelerated predictions coming from the suite of emulators to examine the impact of analyses setup when conducting parameter inference with data from the 6dFGS, BOSS, and eBOSS galaxy surveys. Via a set of mock full-shape analyses of synthetic power spectrum multipoles, designed to approximate measurements from the surveys above, we demonstrate that the use of alternative priors on nuisance parameters and restricted model complexity reduces many of the biases previously observed in marginalised cosmological constraints coming from EFTofLSS analyses. The alternative priors take the form of a Jeffreys prior; a non-informative prior that can mitigate against biases induced by marginalising over poorly constrained nuisance parameters. When performing a joint analysis of all synthetic multipoles, we see an improvement in the level of agreement between the marginalised ln 1010As constraints and the truth; from ∼ 2.0σ to ∼ 0.42σ. Using our pipeline to analyse the measured multipoles, we find an improvement in the level of agreement with cosmic microwave background (CMB) results; from ∼ 2.4σ to ∼ 0.5σ.
Date of Award1 Dec 2023
Original languageEnglish
SupervisorKazuya Koyama (Supervisor) & Florian Beutler (Supervisor)

Cite this