# Predicting CMB dust foreground using galactic 21 cm data

## TL;DR

This paper explores using galactic 21 cm emission data and neural networks to predict dust foregrounds in CMB observations, potentially improving foreground removal for primordial B-mode detection.

## Contribution

It demonstrates that galactic 21 cm velocity data can be used with neural networks to predict dust foregrounds, showing improved accuracy over simpler models.

## Key findings

- Velocity structure improves prediction accuracy.
- Significant improvement at arc-minute scales.
- Potential to enhance polarization foreground templates.

## Abstract

Understanding large-angular-scale galactic foregrounds is crucial for future CMB experiments aiming to detect $B$-mode polarization from primordial gravitational waves. Traditionally, the dust component has been separated using its different frequency dependence. However, using non-CMB observations has potential to increase fidelity and decrease the reconstruction noise. In this exploratory paper we investigate the capability of galactic 21 cm observations to predict the dust foreground in intensity. We train a neural network to predict the dust foreground as measured by the Planck Satellite from the full velocity data-cube of galactic 21 cm emission as measured by the HI4PI survey. We demonstrate that information in the velocity structure clearly improves the predictive power over both a simple integrated emission model and a simple linear model. The improvement is significant at arc-minute scales but more modest at degree scales. This proof of principle on temperature data indicates that it might also be possible to improve foreground polarization templates from the same input data.