Improving 21-cm Radio Interferometry Mock Observation for Cosmology Inference with Machine Learning

Abstract: The Square Kilometre Array (SKA-Low) is set to revolutionize our understanding of the reionization era by providing extensive 3D tomographic data on the distribution of neutral hydrogen. However, the analysis of this data faces significant challenges in separating the desired 21-cm signal from the unwanted foreground and instrumental noise contaminations. In this study, we introduce SERENEt, a deep learning approach designed to tackle these issues using mock observations from the SKA with an observation time of 1000 hours, accounting for the presence of the Galactic synchrotron foreground. Our network effectively identifies neutral hydrogen regions during reionization, enabling recovery of the associated 21-cm signal. The results demonstrate that our approach achieves an accuracy of over 87% in identifying neutral regions during reionisation and an average accuracy of 95% in recovering the 2D power spectra of the 21-cm signal. The SERENEt pipeline is participating in the ongoing SKA Data Challenge 3a (SDC3a), organised by the Square Kilometre Array Observatory (SKAO). Here, I will present our most recent results. Moreover, I highlight the limitations of current mock observation approaches for the Square Kilometre Array, particularly in neglecting essential modulations and additive effects associated with radio interferometry experiments for cosmology. My work aims to overcome these limitations by incorporating multiplicative and additive modulations, including direction dependence, direction independence, and system equivalent flux density (SEFD) parameters, which are crucial in accurately simulating the telescope's synthesised beam. In the SDC3a context, I present our ongoing work to address these effects and improve the understanding and modelling of radio interferometric observations.