The universe is a noisy place when it comes to radio transmissions. Cosmological objects from stars to black holes emit radio waves that can sound like eerie electronic ringing or high-pitched squeaks.
Of particular interest to cosmologists are radio emissions outside the Milky Way coming from neutral hydrogen atoms. These are found within galaxies and in interstellar clouds. These cosmic emissions are important because they can give us clues about the distribution and structure of celestial matter in space, along with the universe’s evolution.
But these hydrogen emissions are relatively faint and tend to be drowned out by stronger, closer radio signals coming from our home galaxy. But a team led by researchers at Princeton University and affiliated with the school’s Center for Statistics and Machine Learning (CSML) have developed a deep learning neural network that can separate cosmic neutral hydrogen from these “foreground contaminants.” This research, “deep21: a Deep Learning Method for 21cm Foreground Removal,” was recently published in April in the Journal of Cosmology and Astroparticle Physics.
The method can be used to interpret data from upcoming radio telescopes that are under construction. This includes the Square Kilometre Array, based in South Africa and Australia, said T. Lucas Makinen ’20, the first author of the paper and currently a master’s student in physics at Sorbonne Université Paris. The work on this paper formed the basis of his senior thesis at Princeton and his independent project for CSML’s undergraduate certificate.
“The ability to better read the data coming from these telescopes will give us a deeper understanding of the growth of structure in the universe,” said Makinen.
Neutral Hydrogen Atoms in the Universe
Before outlining this deep learning technique, we first need to understand why it’s essential to look at the radio emissions of neutral hydrogen atoms.
Neutral hydrogen atoms are stable, meaning they do not frequently emit electromagnetic radiation. However, each atom has a tiny probability of undergoing a spin-flip transition in its electron cloud and emitting a photon with a radio wavelength of 21 cm (a frequency of 1,420 megahertz). This occurrence is possible, but it would take millions of years for this to happen spontaneously. But space is so full of hydrogen atoms that radio cosmologists can detect these rare spin flips. These emissions can penetrate the thick swirls of interstellar dust that hampers optical observations of the universe.
“It’s an excellent tracer for baryon content (ordinary atoms) and the dark matter which presumably makes up most of our universe – where there’s matter, there’s hydrogen,” said Makinen.
Scientists can use the emissions to study matter in galaxies. For example, the emissions can yield estimates of a celestial body’s mass and were used to confirm the Milky Way’s spiral structure. Importantly, scientists can use hydrogen emissions to understand better the universe’s early history and how structures like the first galaxies and stars formed through cosmic time.
Issues with Existing Methods
To study these faint hydrogen radio emissions, cosmologists first remove “foreground” contaminants – radio signals from the Milky Way and extragalactic sources.
“These contaminants tend to be three to four orders of magnitude brighter than the interesting cosmological signal – think of trying to pick out a set of anthills on top of Mount Everest,” said Makinen.
One technique that radio cosmologists currently use to accomplish this separation is called principal component analysis (PCA). This statistical method reduces the high dimensionality of a big data set but still retains prevalent patterns and hopefully relevant information. PCA removes most foreground signals, but it is not physically precise, is unreliable, and may even over-subtract from the underlying cosmological signal that scientists are seeking, Makinen said.
A Deep Learning Neural Network for Radio Cosmology
“Component separation is a critical prerequisite of precision cosmology,” said Peter Melchior, assistant professor jointly appointed to astrophysical science and CSML and one of the paper’s authors.
To achieve this aim, researchers need to recover the cosmological signal over-subtracted during the PCA process, Makinen said. The researchers turned to machine learning, specifically by constructing a convolutional neural network to retrieve the interesting hydrogen emission signals from data discarded during the PCA process.
This neural network, called deep21, has a UNet architecture, originally developed to identify cancer cells in biomedical imaging.
“U-Nets were developed for image segmentation, that means for separating features in images spatially. This research work generalizes it to the multi-channel separation of spatially overlapping features: the foreground emissions from our Galaxy from the cosmological 21cm signal of neutral hydrogen,” said Melchior.
The neural network picked out the cosmological 21cm emissions, which tend to have different spectral and statistical qualities and patterns from foreground contaminants.
To test and verify this method, the researchers developed a suite of 100 full-sky simulations of both cosmological signal and foregrounds and then had the neural network analyze these datasets. Various tests showed that the neural network was effective at recovering 21cm emissions. For example, brightness temperature maps of the 21cm cosmological signal processed by the neural network were closer in appearance to and bore the same statistical structure as the expected simulated signal than did maps processed by PCA.
“Our method demonstrates that cosmological analyses on previously irretrievable 21cm intensity maps may be possible in an observational setting,” said Makinen.
The team also tested the deep21 network on various foreground environments and showed that the network could generalize well despite being trained on a single cosmological model.
The next steps for deep21 are to incorporate more detailed observational effects and test on larger datasets and in different environments with various constraints. These trials will help prepare deep21 for the analysis of actual observations.
In addition to Makinen and Melchior, authors of the paper are Lachlan Lancaster, Princeton graduate student in Astrophysical Sciences; Francisco Villaescusa-Navarro, Associate Research Scholar in Astrophysical Sciences at Princeton; Shirley Ho, group leader of the Cosmology X Data Science group at the Center for Computational Astrophysics in the Flatiron Institute; Laurence Perreault-Levasseur, assistant professor at Univesité de Montréal; and David N. Spergel, the Charles A. Young Professor of Astronomy on the Class of 1897 Foundation, Emeritus, Professor of Astrophysical Sciences, Emeritus, (both at Princeton), and the Director of the Center for Computational Astrophysics at the Flatiron Institute.
Funding and Support
Funding sources for this project included the WFIRST program through NNG26PJ30C and NNN12AA01c, as well as the Simons Foundation. A large portion of the work was completed to satisfy requirements for Makinen’s Bachelor of Arts degree at Princeton University.
For researchers interested in accessing and using the software package, the code used for training and generation of results are publicly available at https://github.
A browser-based tutorial for the experiment and UNet module is available via an interactive accompanying Colab notebook.