DeepCMB: Lensing reconstruction of the cosmic microwave background with deep neural networks

João Caldeira, W. L. K. Wu, Brian Nord, Camille Avestruz, Shubhendu Trivedi, K. T. Story

TL;DR

This paper introduces DeepCMB, a novel deep learning approach using Residual U-Nets (ResUNets) to reconstruct the cosmic microwave background (CMB) lensing potential. DeepCMB outperforms traditional quadratic estimators (QE) in signal-to-noise ratio, achieving comparable performance to more complex maximum likelihood methods, paving the way for more precise cosmological parameter estimation from future CMB experiments.

What Problem Does This Paper Address?

The cosmic microwave background (CMB) is a treasure trove of information about the early universe. However, as CMB photons travel to us, their paths are deflected by the gravitational potential of intervening matter – a phenomenon called gravitational lensing. This lensing distorts the CMB, obscuring crucial information about primordial physics. To extract this information, scientists need to accurately reconstruct the lensing potential from CMB maps. Traditional methods like the quadratic estimator (QE) are effective but become suboptimal at the low noise levels expected from next-generation CMB experiments (like CMB Stage-4). While maximum likelihood estimators (MLE) offer better performance, they are computationally intensive and still under development. This paper addresses the need for a more efficient and accurate method for CMB lensing reconstruction, particularly for future low-noise experiments.

Methods

The authors propose using a deep convolutional neural network, specifically a Residual U-Net (ResUNet), for CMB lensing reconstruction. This approach treats the problem as an image-to-image regression task: the network takes lensed CMB polarization maps (Q and U) as input and outputs the unlensed CMB E-mode map and the gravitational convergence map (κ), which is directly related to the lensing potential.

Key aspects of their method include:

  • ResUNet Architecture: This architecture combines an encoder-decoder structure with skip connections, allowing the network to capture both high-level and low-level features from the input images. Residual connections are incorporated to improve training performance.
  • End-to-End Learning: Unlike traditional methods that often treat lensing reconstruction and delensing as separate steps, DeepCMB learns to perform both simultaneously.
  • Physical Model Independence: The network is trained on simulated data and does not explicitly incorporate physical parametrizations of CMB lensing, allowing it to potentially capture complex, hard-to-model astrophysical foregrounds and noise contributions in the future.
  • Noise Spectrum as a Metric: The authors use the noise spectrum of the reconstructed convergence map as a key metric for comparing their method against standard techniques.

Datasets

The authors used simulated CMB data for training and evaluation.

  • Simulated Data: 11,200 independent realizations of simulated CMB maps (Q, Ū, E, κ) were generated. These maps were 5 degrees by 5 degrees on the sky and pixelized into 128x128 images.
  • Data Splitting: The simulations were split into training (80%), validation (10%), and testing (10%) sets. Crucially, these sets were drawn from separate full-sky maps to prevent contamination.
  • Noise Levels: Simulations were generated with varying levels of white noise (0, 1, and 5 µK-arcmin), beam smoothing (1 arcmin), and an apodization mask to mimic realistic observational conditions.
  • Cosmological Variations: For testing the network's sensitivity to cosmology, noiseless maps generated with different cosmological parameters (specifically ΩCDMh²) were used as input.

Key Insights and Ideas

  • Deep Learning for Lensing: The paper demonstrates that deep neural networks can effectively learn the complex mapping required for CMB lensing reconstruction, offering a powerful alternative to traditional methods.
  • Performance Beyond QE: DeepCMB significantly outperforms the standard quadratic estimator, especially at the low noise levels relevant for future experiments. This suggests that neural networks can extract more signal from noisy data.
  • Approaching Maximum Likelihood: The performance of DeepCMB in terms of noise spectrum is comparable to that of iterative methods that approximate maximum likelihood estimators, indicating a high level of efficiency.
  • Cosmological Sensitivity: The network's output is sensitive to changes in the input cosmology, suggesting its potential for cosmological parameter estimation. This is a crucial validation for its use in scientific analysis.
  • Potential for Foreground Mitigation: The model-free nature of the neural network approach holds promise for future applications where complex foregrounds and noise are present and difficult to model analytically.

Key Results

  • Signal-to-Noise Improvement: DeepCMB achieves a higher signal-to-noise ratio for lensing potential reconstruction compared to the QE method, with improvements of 50-70% in noise spectra across a wide range of angular scales (L).
  • Comparison to MLE: The noise spectra from DeepCMB are comparable to those obtained from iterative methods that approximate maximum likelihood estimators.
  • Cosmological Parameter Sensitivity: When tested with different cosmological parameters, the network's recovered κ spectra were significantly different, allowing the authors to rule out the null hypothesis of cosmological insensitivity at the 2.9-4.5 σ level.
  • E-mode Recovery: For noiseless inputs, the network recovers large-scale structure in the E-mode map well, with residuals within a few tens of percent of the input map's maximum pixel value.
  • κ Recovery: In the noiseless case, DeepCMB recovers over 90% of the map fluctuations (80% in power spectrum) for the entire L range considered. However, κ recovery is more sensitive to noise than E-mode recovery.

Limitations and Future Work

  • Small-Scale Performance: The paper notes that the network's performance degrades at higher multipoles (L ≥ 2000 for 5 µK-arcmin noise), where the signal-to-noise ratio of the input maps becomes low. This suggests that the network might be more sensitive to local features than standard methods.
  • Finite Map Size: The observed decrease in recovered power for low-L modes is attributed to the finite size (5 degrees) of the simulated maps. Future work could involve training on larger sky patches.
  • Randomness in Training: The inherent randomness in neural network training can lead to variations in learned mappings. The authors suggest using ensembles of networks or weight averaging to mitigate this.
  • Foregrounds and Systematics: The current simulations do not include galactic and extragalactic foregrounds. Future work will incorporate these to study their effects and potentially extract foreground components.
  • Primordial B-modes: The authors plan to extend this work to recover primordial B-modes, which are crucial for constraining inflation parameters.
  • Direct Parameter Estimation: Investigating the possibility of directly estimating cosmological parameters from the input CMB maps using neural networks is a future direction.
  • Spherical CNNs: For analysis of larger sky regions, the use of spherical convolutional networks will be necessary as the flat-sky approximation will no longer be valid.

Summary updated

Abstract

Next-generation cosmic microwave background (CMB) experiments will have lower noise and therefore increased sensitivity, enabling improved constraints on fundamental physics parameters such as the sum of neutrino masses and the tensor-to-scalar ratio r. Achieving competitive constraints on these parameters requires high signal-to-noise extraction of the projected gravitational potential from the CMB maps. Standard methods for reconstructing the lensing potential employ the quadratic estimator (QE). However, the QE performs suboptimally at the low noise levels expected in upcoming experiments. Other methods, like maximum likelihood estimators (MLE), are under active development. In this work, we demonstrate reconstruction of the CMB lensing potential with deep convolutional neural networks (CNN) - ie, a ResUNet. The network is trained and tested on simulated data, and otherwise has no physical parametrization related to the physical processes of the CMB and gravitational lensing. We show that, over a wide range of angular scales, ResUNets recover the input gravitational potential with a higher signal-to-noise ratio than the QE method, reaching levels comparable to analytic approximations of MLE methods. We demonstrate that the network outputs quantifiably different lensing maps when given input CMB maps generated with different cosmologies. We also show we can use the reconstructed lensing map for cosmological parameter estimation. This application of CNN provides a few innovations at the intersection of cosmology and machine learning. First, while training and regressing on images, we predict a continuous-variable field rather than discrete classes. Second, we are able to establish uncertainty measures for the network output that are analogous to standard methods. We expect this approach to excel in capturing hard-to-model non-Gaussian astrophysical foreground and noise contributions.
BibTeX
@inproceedings{caldeira2019deepcmb,
  author = {João Caldeira and W. L. K. Wu and Brian Nord and Camille Avestruz and Shubhendu Trivedi and K. T. Story},
  title = {{DeepCMB: Lensing reconstruction of the cosmic microwave background with deep neural networks}},
  booktitle = {Astron. Comput},
  year = {2019},
  url = {https://openreview.net/forum?id=Vl9PWLbfYs},
}

Loading related content...

Figure