ABCMB: Deep Delensing Assisted Likelihood-Free Inference from CMB Polarization Maps

ABCMB is a deep learning framework for CMB B-mode polarization delensing and tensor-to-scalar ratio inference. This repository contains the implementation of the method described in our paper.

Installation

Environment Setup

Create a conda environment using our provided requirements:

# Create and activate environment
conda create --name ABCMB --file requirements.txt
conda activate ABCMB

Getting Started

Download Pre-trained Models

Download our pre-trained model weights

Data Generation

Generate simulation data for training or inference:

# Generate validation data
python lensing_sim.py --output_path 'output/val/all_map/'

# Generate training data
python lensing_sim.py --output_path 'output/train/all_map/'

Running Inference

For detailed examples of delensing Q/U maps to obtain B-mode maps and computing power spectra, see our inference notebook.

Citation

If you consider our codes and datasets useful, please cite:

@article{yi2024ab,
  title={AB $$\backslash$mathbb $\{$C$\}$ $ MB: Deep Delensing Assisted Likelihood-Free Inference from CMB Polarization Maps},
  author={Yi, Kai and Fan, Yanan and Hamann, Jan and Li{\`o}, Pietro and Wang, Yuguang},
  journal={arXiv preprint arXiv:2407.10013},
  year={2024}
}