Application of Machine Learning to 21 cm Cosmology

Abstract

This chapter reviews applications of machine learning (ML) to redshifted 21 cm cosmology, focusing on cosmic dawn, the Epoch of Reionization, and SKA-Low science. The redshifted 21 cm line directly probes diffuse neutral hydrogen, but the measured signal is not a simple astrophysical observable: density, ionization, heating, radiation backgrounds, foreground treatment, and instrumental response are coupled. The chapter first summarizes the physical ingredients needed in later sections, including the global signal, spatial fluctuations, morphology-sensitive statistics, and the 21 cm forest. It then discusses the main barriers to interpretation: bright foregrounds, radio-frequency interference, ionospheric and calibration effects, incomplete sampling, and the cost of forward modeling in high-dimensional parameter spaces. ML applications are organized by their role in the analysis chain. Observation-domain methods act on contaminated data products; theory-domain methods accelerate or compress forward modeling; and inference-domain methods connect observables to astrophysical and cosmological parameters.