Line intensity mapping (LIM) is an observational technique that probes the large-scale structure of the Universe by collecting light from a wide field of the sky. This book demonstrates a novel analysis method for LIM using machine learning (ML) technologies. The author develops a conditional generative adversarial network that separates designated emission signals from sources at different epochs. It thus provides, for the first time, an efficient way to extract signals from LIM data with foreground noise. The method is complementary to conventional statistical methods such as cross-correlation analysis. When applied to three-dimensional LIM data with wavelength information, high reproducibility is achieved under realistic conditions. The book further investigates how the trained machine extracts the signals, and discusses the limitation of the ML methods. Lastly an application of the LIM data to a study of cosmic reionization is presented. This book benefits students and researchers who are interested in using machine learning to multi-dimensional data not only in astronomy but also in general applications.
Author(s): Kana Moriwaki
Series: Springer Theses
Publisher: Springer
Year: 2022
Language: English
Pages: 125
City: Singapore
Supervisor’s Foreword
Acknowledgments
Contents
1 Introduction
References
2 Observations of the Large-Scale Structure of the Universe
2.1 Large-Scale Structure of the Universe
2.2 Observations of Large-Scale Distribution of the Galaxies
2.2.1 Galaxy Surveys
2.2.2 Line Intensity Mapping
2.3 Observations of the Cosmic Reionization
2.3.1 Current Observational Constraints on the Reionization
2.3.2 Observations of the 21-cm Lines at the EoR
References
3 Modeling Emission Line Galaxies
3.1 Line Emissions from Hii Regions
3.2 Emission Line Model
3.3 Mock Observational Line Intensity Maps
References
4 Signal Extraction from Noisy LIM Data
4.1 Machine Learning Algorithms
4.1.1 Basics of Neural Networks
4.1.2 Convolutional Neural Networks
4.1.3 Generative Adversarial Networks
4.2 Methods: Training Data and Network Architecture
4.3 Extracted Signals from Noisy Maps
4.4 Discussions
4.4.1 Different Emission Line Models
4.4.2 Choice of Training Data
4.5 Conclusion
References
5 Signal Separation from Confused LIM Data
5.1 Line Confusion Problem in Line Intensity Mapping Observations
5.2 Methods: One-to-Many Translation Network Architecture
5.3 Separation of Multiple Emission Line Signals
5.4 Discussions
5.4.1 Different Emission Line Models
5.4.2 Combining Multiple Networks
5.4.3 Convolutional Filters and Hidden Layers
5.5 Conclusion
References
6 Signal Extraction from 3D LIM Data
6.1 Methods
6.1.1 Data Preparation
6.1.2 Physics-Informed Network Architecture
6.2 Reconstruction of Three-Dimensional Large-Scale Structures
6.3 Understanding the Networks
6.4 Conclusion
References
7 Application of LIM Data for Studying Cosmic Reionization
7.1 Methods
7.1.1 Reionization Simulation
7.1.2 [Oiii] Line Emission
7.2 Cross-Power Spectra
7.3 Discussions
7.3.1 Small-Scale Signals
7.3.2 Large-Scale Signals
7.3.3 Detectability of the Signals
7.4 Conclusion
References
8 Summary and Outlook
References
Appendix A Training of the Generative Models
A.1 Loss Functions of GANs
A.2 Choice of Training Models and Datasets
Appendix B 21-cm Line from Intergalactic Medium
B.1 Brightness Temperature
B.2 Noise Power Spectrum
