Towards a Robust Machine-Learning Pipeline for 21-cm Cosmology Data Analysis I: A Roadmap for Development and Demonstration of Robustness Against PSF Modeling Errors

Abstract

The 21-cm signal from the Epoch of Reionization (EoR) is a powerful probe of the evolution of the Universe. However, accurate measurements of the EoR signal from radio interferometric observations are sensitive to efficient foreground removal, mitigating radio-frequency interference and accounting for instrumental systematics. This work represents the first in a series of papers, where we will be introducing a novel ML based pipeline, step-by-step, to directly infer reionization parameters from 21-cm radio-interferometric images. In this paper, we investigate the impact of the variations in the point spread function (PSF) on parameter estimation by simulating visibilities corresponding to input 21-cm maps as observed by the 128-antenna configuration of the Murchison Widefield Array (MWA) Phase II. These visibilities are imaged to obtain dirty images, which are then used to train a 2D convolutional neural network (CNN) to predict $\rm x_{HI}$. To systematically assess the effect of PSF mis-modelling, we generate multiple test sets by varying the MWA's antenna layout, thereby introducing controlled variations in the PSF; we then feed these alternative PSF dirty images to our CNN trained using only dirty images with the PSF of the true antenna layout. Our results demonstrate that PSF variations introduce biases in the CNN's predictions of $\rm x_{HI}$, with errors depending on the extent of PSF distortion. We quantify these biases and discuss their implications for the reliability of machine-learning-based parameter inference in 21-cm cosmology and how they can be utilized to improve the robustness of estimation against PSF-related systematics in future 21-cm surveys. In concluding, we also discuss how this approach to incorporating realistic instrument error into an ML analysis pipeline can be expanded to include multiple other effects.