Inferring the Physics of Reionization Using Non-linear Data Compression Techniques

Abstract

The 21-cm signal from neutral hydrogen during the Epoch of Reionization (EoR) is one of the most promising probes of the first luminous sources and the process by which they ionized the intergalactic medium. However, the signal is intrinsically high-dimensional and non-Gaussian, posing significant challenges for parameter inference using traditional summary statistics such as the power spectrum. In this thesis, we develop and evaluate an inference pipeline that combines Information Maximizing Neural Networks (IMNNs) for non-linear data compression with Simulation-Based Inference (SBI) for likelihood-free posterior estimation, and benchmark it against power-spectrum-based approaches using both SBI and Markov Chain Monte Carlo (MCMC). We generate mock 21-cm observations at redshift z = 7 using the semi-numerical reionization code SCRIPT, targeting two key parameters: the mean ionized fraction Q_M^HII and the minimum halo mass log10Mmin . Three analysis workflows of increasing observational realism are constructed: compression of the spherically averaged power spectrum, compression of mean-subtracted 3D ionization maps, and compression of noisy mean-subtracted 3D brightness temperature maps incorporating SKA-like instrumental noise. Our key result is that the IMNN+SBI framework delivers the most accurate and tightest param- eter constraints in the most observationally realistic scenario. When applied to noisy brightness temperature maps, IMNN+SBI recovers parameters consistent with the fiducial values ((Q_M^HII = 0.548 (+0.032, −0.034), log10Mmin = 8.936 (+0.238, −0.242 )), whereas both power-spectrum-based methods produce posteriors that are significantly biased and broader. This demonstrates that learned, non-linear summary statistics can access information in the full 3D field that is lost when the data are reduced to a one-dimensional power spectrum. In the complementary regime of clean power spectra, all three methods perform comparably. However, for noise-free ionization maps, IMNN+SBI already demonstrates its ability to extract non-Gaussian spatial information, yielding notably tighter posteriors than power spectrum-based approaches. These findings establish the IMNN+SBI pipeline as a promising framework for extracting reionization physics from upcoming 21-cm imaging experiments such as the SKA, with natural extensions to foreground-contaminated data and richer astrophysical parameter spaces.