Constraining reionization parameters through Bayesian neural network & Bayesian inference

Abstract

The Epoch of Reionization (EoR) is the least-known era in the universe’s history. Many next-generation radio interferometers, e.g., the Square Kilometre Array (SKA), are expected to observe this era via the redshifted 21-cm line emitted by the neutral hydrogen. Once such observations are available, the next challenge is to interpret them. Given the versatility of the possible reionization models (arising due to the various complex physical processes that drive it), it is a considerably significant and non-trivial challenge. Further, the realistic reionization models are computationally very expensive to run if one wants to build them in a self-consistent manner. In this project, we try to address this issue by building Neural Network-based emulators for the observable 21-cm power spectrum. Here, we first focus on improving the performance of such an existing Artificial Neural Network (ANN)-based emulator (Tiwari et al. [1]) for the target signal statistic by training it on a simulated signal power spectrum library that covers a larger portion of the model parameter space. Next, we identify the region of the parameter space in which the emulator is more prone to emulating erroneous signal statistics and try to optimally increase the training sample size in that region to improve its performance. Finally, using this emulator, we constrain the reionization parameters via Bayesian inference with the Metropolis-Hastings Markov Chain Monte Carlo (MH-MCMC) framework.