New Deep Learning Framework Enhances 21-cm Signal Recovery in Reionization Studies

A new deep learning framework, named SERENEt, has been developed to enhance the recovery of the 21-cm signal from observations made by the Square Kilometre Array Observatory (SKA-Low). This framework is particularly aimed at addressing the challenges posed by foreground contamination, which complicates the detection of the 21-cm signal during the reionization era.

The research, conducted by Michele Bianco and a team of nine other authors, demonstrates that SERENEt can recover the signal distribution with an average accuracy of 75% during the early stages of reionization, improving to 90% as reionization progresses. Conversely, the detection accuracy for neutral hydrogen (HI) regions starts at 92% but decreases to 73% as reionization advances.

In addition to improving image recovery, SERENEt produces cylindrical power spectra with an average accuracy exceeding 93% throughout the reionization period. The framework was tested on a simulation covering a 10-degree field of view, showing consistent improvements when prior maps were utilized. Notably, providing prior information about the locations of HII regions enhanced the recovery of the 21-cm signal by approximately 10%.

The findings suggest that high-redshift galaxy surveys can optimize foreground mitigation and enhance the construction of 21-cm images, which is crucial for understanding the early universe. This research is detailed in the paper titled "Deep learning approach for identification of HII regions during reionization in 21-cm observations -- III. image recovery," available on arXiv (arXiv:2408.16814).