Cleaning our own Dust: Simulating and Separating Galactic Dust Foregrounds with Neural Networks

TL;DR

This paper introduces a neural network-based approach to model and separate galactic dust foregrounds from CMB maps, improving uncertainty quantification in cosmological observations.

Contribution

It develops a DCGAN to model non-Gaussian dust emission and uses it to enhance CMB signal estimation from contaminated maps.

Findings

DCGAN effectively models complex dust emission distributions

Neural network improves CMB signal extraction accuracy

Potential applications include polarized emission analysis

Abstract

Separating galactic foreground emission from maps of the cosmic microwave background (CMB), and quantifying the uncertainty in the CMB maps due to errors in foreground separation are important for avoiding biases in scientific conclusions. Our ability to quantify such uncertainty is limited by our lack of a model for the statistical distribution of the foreground emission. Here we use a Deep Convolutional Generative Adversarial Network (DCGAN) to create an effective non-Gaussian statistical model for intensity of emission by interstellar dust. For training data we use a set of dust maps inferred from observations by the Planck satellite. A DCGAN is uniquely suited for such unsupervised learning tasks as it can learn to model a complex non-Gaussian distribution directly from examples. We then use these simulations to train a second neural network to estimate the underlying CMB signal from dust-contaminated maps. We discuss other potential uses for the trained DCGAN, and the generalization to polarized emission from both dust and synchrotron.