GCE NN

This repository contains the Tensorflow implementation of the papers

• Extracting the Galactic Center Excess' source-count distribution with neural nets, Florian List, Nicholas L. Rodd, and Geraint F. Lewis, Phys. Rev. D 104, 123022, 2021 [arXiv:2107.09070].
  → Branch "prd"
• Galactic Center Excess in a New Light: Disentangling the γ-Ray Sky with Bayesian Graph Convolutional Neural Networks, Florian List, Nicholas L. Rodd, Geraint F. Lewis, and Ishaan Bhat, Phys. Rev. Lett. 125, 241102, 2020 [arXiv:2006.12504].
  → Branch "prl"

The "main" branch contains the most up-to-date version, which is under development.

Author: Florian List (University of Vienna).
Contributions: Nicholas L. Rodd (CERN).

For any queries, please contact me at florian dot list at univie dot ac dot at.

 

Disclaimer: The code in this repository borrows from several other GitHub repositories and other publicly available sources. In particular, the neural network architecture is built upon DeepSphere (Perraudin et al. 2019, Defferrard et al. 2020).

Data

The Fermi dataset and the templates contained in this repository have been generated for the following papers:

• Spurious point source signals in the galactic center excess, Rebecca K. Leane and Tracy R. Slatyer, Phys. Rev. Lett. 125, 121105, 2020 [arXiv:2002.12370].
• The enigmatic Galactic Center excess: Spurious point sources and signal mismodeling, Rebecca K. Leane and Tracy R. Slatyer, Phys. Rev. D 102, 063019, 2020 [arXiv:2002.12371].

The data is made available with the permission of the authors, and everybody using the data for a publication should cite these papers.

(The data selection criteria can be found in arXiv:2107.09070.)

Getting started

Option 1: Try out the code in Google Colab

The folder examples contains an example Jupyter notebook gce_nn_example_notebook_colab.ipynb that can be run in Google Colab. The Jupyter notebook will clone this Github repository into the (temporary) /content/ folder on Google Colab, so it is not needed to manually clone or download this repository to get started.

1. Open Google Colab.
2. Click on "Github" in the top panel.
3. Enter https://github.com/FloList/GCE_NN and click on "Search".
4. Select examples/gce_nn_example_notebook_colab.ipynb.
5. Select a GPU runtime (Runtime -> Change runtime type -> GPU).
6. Run the notebook. (Note: After this Github repository has been cloned by the Jupyter notebook and the packages have been installed, the Jupyter notebook kernel needs to be restarted as explained in the notebook.)

Option 2: Clone the repository and run it locally on your computer

First, clone the repository via

git clone https://github.com/FloList/GCE_NN.git

Warning: This github repository is quite large (several hundred MBs) because it contains Fermi data and templates.

Then, cd into the directory

cd GCE_NN

We highly recommend using a new virtual environment for the GCE NN analysis in which all the required packages can be installed in isolation from the globally installed packages.

This can be done using venv

python3.8 -m venv venv_gce_nn  # create the environment
source venv_gce_nn/bin/activate  # activate it
# deactivate   # to deactivate the environment

or if you are using pyenv

pyenv virtualenv 3.8.0 venv_gce_nn  # create the environment
pyenv activate venv_gce_nn  # activate it
# pyenv deactivate   # to deactivate the environment

or if you are using conda

conda create -n venv_gce_nn python=3.8.0 anaconda  # create the environment
conda activate venv_gce_nn  # activate it
# conda deactivate  # to deactivate the environment

Once you are inside the virtual environment, all the required dependencies can be installed from the requirements.txt file with

pip install -r requirements.txt

Afterwards, install the GCE NN package with

python setup.py install

Then, a good starting point is the Jupyter notebook gce_nn_example_notebook.ipynb in the examples folder, which performs a convolutional neural network-based analysis of γ-ray photon-count maps for a simple scenario. To consider a different scenario, generate a new parameter file in the parameter_files folder (for example by copying the file parameters.py and modifying the relevant settings).

Results for the Fermi data (from arXiv:2107.09070)

• In the Fermi data, we find a faint Galactic Center Excess (GCE) described by a median source-count distribution (SCD) peaked at a flux of ~ 4 × 10⁻¹¹ counts / cm² / s (corresponding to ~ 3 - 4 expected counts per PS), which would require N ~ O(10⁴) sources to explain the entire excess (median value N = 29,300 across the sky).
• Although faint, this SCD allows us to derive the constraint ηₚ ≤ 66% for the Poissonian fraction of the GCE flux ηₚ at 95% confidence, suggesting that a substantial amount of the GCE flux is due to PSs.

Citations

@article{List_et_al_2021,
archivePrefix = {arXiv},
arxivId = {2107.09070},
author = {List, Florian and Rodd, Nicholas L. and Lewis, Geraint F.},
eprint = {2107.09070},
journal = {Physical Review D},
volume = {104},
number = {12},
pages = {123022},
title = {{Extracting the Galactic Center Excess' source-count distribution with neural nets}},
url = {https://link.aps.org/doi/10.1103/PhysRevD.104.123022},
year = {2021}
}

@article{List_et_al_2020,
archivePrefix = {arXiv},
arxivId = {2006.12504},
author = {List, Florian and Rodd, Nicholas L. and Lewis, Geraint F. and Bhat, Ishaan},
eprint = {2006.12504},
journal = {Physical Review Letters},
volume = {125},
number = {24},
pages = {241102},
title = {{Galactic Center Excess in a New Light: Disentangling the γ-Ray Sky with Bayesian Graph Convolutional Neural Networks}},
url = {https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.125.241102},
year = {2020}
}