Speaker
Description
The redshifted 21-cm signal from the Epoch of Reionization (EoR) is a powerful probe of the structure formation and ionization history of our Universe. However, extracting the reionization parameters from the observations of this signal requires an efficient and robust inference framework which can perform under the influence of strong parameter correlations and computationally expensive forward models. In this work, we developed and tested a Metropolis-within-Gibbs (MWG) sampling framework to infer three key EoR parameters—minimum halo mass Mmin ,ionizing efficiency Nion, and mean free path Rmf — using the 21-cm power spectrum as the observable. To enable fast likelihood estimation, we used an uncertainty-aware Bayesian Neural Network (BNN) emulator that predicts both the power spectrum and its predictive uncertainty. The likelihood was modelled as a multivariate Gaussian with a diagonal covariance including SKA-Low noise and emulator uncertainty. We compared MWG against a standard Metropolis-Hastings (MH) sampler that proposes all parameters jointly. Across three test cases considering observations at different stages of reionization, MWG recovered posterior constraints consistent with MH while improving sampling efficiency, yielding higher acceptance ratios, fewer sampling steps to converge, and smoother mixing. Building on this inference framework, we are currently developing a foreground-robust extension to perform inference of the 21-cm signal from a contaminated observation cube by masking foreground dominated low-k∥ line-of-sight Fourier modes. The remaining safe modes (which lie in the EoR window) are expected to have only signal, while the masked modes are reconstructed using Gaussian Constrained Realisations (GCR). A prior LoS covariance is expected to be learned from a large ensemble of signal-only simulations, enabling conditional sampling of masked modes for each pixel and generating an ensemble of reconstructed cubes. From these reconstructions, we estimate the power spectrum and its uncertainty, discarding the k-bins dominated by reconstruction variance, and finally infer the EoR parameters using a Bayesian emulator-based likelihood.
