Application of 3D U-Net Neural Networks in Extracting the Epoch of Reionization Signal from SKA-Low Observations Based on Real Observations of NCP Field from LOFAR

Summary

The 21-cm signal from the Cosmic Dawn and Epoch of Reionization is crucial for understanding the early universe. However, observations are challenged by thermal noise and systematic effects. This study simulates SKA-Low-depth images and processes them with a deep learning method, utilizing foreground residuals, thermal and excess variances, and 21-cm signals generated with 21cmFAST.

Highlights

• The study uses a 3D U-Net neural network architecture for image analysis to overcome foreground, thermal noise, and excess variance components.
• The foreground residuals are obtained from actual LOFAR observations of the North Celestial Pole field.
• The thermal noise and excess variance are calculated via Gaussian process regression.
• The 21-cm signals are generated with 21cmFAST for signal extraction tests.
• The U-Net provides reliable 2D power spectrum predictions and robustness tests ensure realistic EoR signals.
• The study evaluates the impact of different observation times on the signal extraction.
• The results show that the U-Net can recover the 21-cm signal in the presence of strong foreground and excess variance.

Key Insights

• The 3D U-Net neural network is effective in extracting the 21-cm signal from simulated SKA-Low-depth images, demonstrating its potential for processing real SKA observational data.
• The study highlights the importance of reducing residual foreground and excess variance to improve the recovery of the 21-cm signal below the horizon delay line.
• The results demonstrate that the U-Net can recover the 21-cm signal in the presence of strong foreground and excess variance, but the presence of these contaminations can impact the recovery of the signal.
• The study shows that the U-Net can produce reliable 2D power spectra for the EoR signal, even in the presence of strong foreground and excess variance.
• The results suggest that the U-Net can be used to process real SKA observational data, but further improvements in data calibration and foreground subtraction methods are necessary.
• The study demonstrates the potential of deep learning techniques for processing large datasets and extracting signals from noisy data.
• The results highlight the importance of considering the impact of different systematic effects on the recovery of the 21-cm signal.

Mindmap

Citation

Gao, L.-Y., Koopmans, L. V. E., Mertens, F. G., Munshi, S., Li, Y., Brackenhoff, S. A., Ceccotti, E., Chege, J. K., Acharya, A., Ghara, R., Giri, S. K., Iliev, I. T., Mellema, G., & Zhang, X. (2024). Application of 3D U-Net Neural Networks in Extracting the Epoch of Reionization Signal from SKA-Low Observations Based on Real Observations of NCP Field from LOFAR (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2412.16853