deep21: a Deep Learning Method for 21cm Foreground Removal (2010.15843v2)

Abstract: We seek to remove foreground contaminants from 21cm intensity mapping observations. We demonstrate that a deep convolutional neural network (CNN) with a UNet architecture and three-dimensional convolutions, trained on simulated observations, can effectively separate frequency and spatial patterns of the cosmic neutral hydrogen (HI) signal from foregrounds in the presence of noise. Cleaned maps recover cosmological clustering statistics within 10% at all relevant angular scales and frequencies. This amounts to a reduction in prediction variance of over an order of magnitude on small angular scales ($\ell > 300$), and improved accuracy for small radial scales ($k_{\parallel} > 0.17\ \rm h\ Mpc^{-1})$ compared to standard Principal Component Analysis (PCA) methods. We estimate posterior confidence intervals for the network's prediction by training an ensemble of UNets. Our approach demonstrates the feasibility of analyzing 21cm intensity maps, as opposed to derived summary statistics, for upcoming radio experiments, as long as the simulated foreground model is sufficiently realistic. We provide the code used for this analysis on Github https://github.com/tlmakinen/deep21 as well as a browser-based tutorial for the experiment and UNet model via the accompanying http://bit.ly/deep21-colab Colab notebook.

• The paper introduces a novel deep21 method that leverages a 3D UNet CNN to separate cosmic signals from overwhelming foreground emissions.
• It outperforms traditional PCA by recovering clustering amplitude and phase within 20% accuracy across all angular scales.
• The approach includes an ensemble of networks for uncertainty quantification, paving the way for improved analyses in upcoming radio astronomy surveys.

Essay on "deep21: a Deep Learning Method for 21cm Foreground Removal"

The paper "deep21: a Deep Learning Method for 21cm Foreground Removal" presents a sophisticated approach leveraging deep learning to improve the separation and cleaning of 21cm intensity maps by mitigating foreground contaminants. The main focus is on utilizing a convolutional neural network (CNN) with a 3D UNet architecture to disentangle cosmic signals from overwhelming foreground noise, a persistent challenge impeding the efficacy of 21cm cosmology in elucidating the universe's formative epochs.

Key Merit and Approach

This paper's core innovation is the application of a UNet-based CNN to simulate and clean maps of cosmic neutral hydrogen (HI) emissions. The deep21 model's architecture is fine-tuned with convolutional layers capable of processing spatial and frequency patterns. This network is trained on artificial datasets which simulate the complexities of foregrounds and cosmic signals, allowing it to learn distinguishing features effectively.

The choice of a UNet architecture is pertinent given its success in image segmentation tasks, offering robust performance in identifying and segregating features spread over spatially complex high-dimensional data. The 3D convolution allows deep21 to capitalize on variances across frequency, enhancing the particular application of astrophysical signal processing where cosmic signal and foreground separation is crucial.

Performance and Findings

The authors provide empirical evidence that the deep21 model outperforms traditional methods like Principal Component Analysis (PCA) in both variance reduction and signal preservation. The model recovers cosmological clustering amplitude and phase within 20% accuracy for all angular scales, signifying a marked improvement over the PCA, which is notably susceptible to signal loss and bias when handling smooth spectral foregrounds.

Furthermore, the use of an ensemble of networks facilitates uncertainty quantification, a crucial step for ensuring predictive robustness when handling real-world noisy data. The outcomes highlight that the model maintains improved signal separation even in altered foreground parameter settings, demonstrating generalizability.

Implications and Future Directions

The findings underscore the feasibility of employing 21cm intensity maps, instead of limited summary statistics, to encapsulate the universe's large-scale structure development. This shift holds promise for enhancing the accuracy of cosmological measurements from upcoming radio astronomy initiatives, such as the Square Kilometer Array (SKA). As these maps encapsulate pivotal epochs like Reionization and structure formation, clean signal interpretation is central to probing the underlying cosmological parameters and physics.

Future advancements could involve extended training on more realistic datasets incorporating complex noise models and polarized foregrounds, simulating scenarios more akin to observational data. Additional measures might include adapting the network to handle larger-scale maps by leveraging greater computational resources, thus further sharpening its efficacy on practical data.

This paper contributes significantly to the toolbox of astrophysical signal processing, showcasing the strength of deep learning in overcoming intricate measurement challenges in radio cosmology. Addressing current limitations and expanding on this research could catalyze further breakthroughs in our understanding of the universe's history and structure.