Unlocking the Early Universe using HI 21cm Cosmology: Marginal Neural Ratio Estimation for 21-cm Power Spectra Analysis

Abstract

The cosmic 21-cm signal serves as a unique probe of the early universe, offering insights into the Epoch of Reionization (EoR), the Epoch of Heating (EoH), and the properties of the f irst galaxies. However, extracting meaningful astrophysical parameters from this complex, non-Gaussian signal poses significant challenges, particularly when traditional likelihood based methods like Monte Carlo Markov Chain (MCMC) are infeasible. In this work, I explore Neural Ratio Estimation (NRE), a cutting-edge Simulation-Based Inference (SBI) technique, to address these challenges. NRE leverages neural networks to approximate the likelihood-to-evidence ratio, enabling efficient posterior estimation without explicit likelihood evaluations. Our research highlights the practical advantages of implementing NRE using a custom PyTorch-based framework over existing libraries like Swyft, which impose restrictive data structures for its input, such as Zarr hierarchies, as well as a completely unknown input hierarchy, making it not user-friendly at all. By designing a streamlined, flexible data pipeline tailored to our needs, we eliminate unnecessary computational overhead and achieve faster training times while maintaining full control over the model architecture and data handling. This approach not only simplifies the workflow but also ensures that the codebase remains modular, transparent, and adaptable to diverse datasets. Weapply MNREtosimulated 21-cm power spectra and lightcones generated using 21cm FAST, demonstrating its ability to infer key astrophysical parameters such as the ionizing efficiency (ζ) and X-ray luminosity per star formation rate (LX,< 2keV/SFR). Our results underscore the potential of NRE to unlock deeper insights into the thermal and ionization history of the intergalactic medium (IGM). By combining the computational efficiency of Py Torch, this work provides a fresh, scalable framework for analyzing upcoming observations from next-generation telescopes like the Square kilometer Array (SKA). This study not only advances the field of 21-cm cosmology but also sets a new standard for flexible, efficient, user-friendly inference pipelines in astrophysics, paving the way for transformative discoveries about the early universe