Deep learning analysis of the gamma-ray Galactic Centre Excess

Supervisor: Florian List

Contact information: florian.list@univie.ac.at 

Expected duration: 9 months

Project description & Goals:

Measurements of the Fermi gamma-ray space telescope show an unexplained excess of photon counts coming from the region around the Galactic Centre, peaking at an energy of ~2 GeV (see [1] for a recent review). The origin of this so-called "Galactic Centre Excess" (GCE) has been extensively discussed in the literature, and it was argued that a population of faint millisecond pulsars or alternatively dark matter annihilation could explain the GCE. While millisecond pulsars are expected to produce a point-like GCE, dark matter annihilation would give rise to a smooth GCE morphology. Traditionally, template-based methods such as the Non-Poissonian Template Fit (NPTF) have been used to distinguish between these two possibilities (e.g. [2]). It has been pointed out, however, that mismatch between the templates and the data can bias the results (e.g. [3]), and it cannot be ruled out at the moment that at least a fraction of the GCE is due to annihilating dark matter.

Recently, it was shown that convolutional neural networks (CNNs) are able to disentangle the different emission components in gamma-ray maps [4] and to characterise the point-source populations [5], and machine learning techniques have also been proposed for a flexible modelling of the background emission [6]. In contrast to template fitting methods, which look at each pixel individually, CNNs analyse many patches of the sky, potentially leading to more robust results.

Currently, neither the NPTF nor CNN-based methods make use of the energy information in the data. Instead, all the photon counts within a selected energy range are put into a single bin for the analysis, thereby throwing out the information provided by the energy spectrum. The goal of this master thesis is to extend current CNN-based analysis methods for the gamma-ray sky by incorporating energy information into the analysis pipeline (i.e., for the training data generation and the CNN-based analysis) and to apply the resulting method to the GCE in the Fermi data, which will possibly lead to new hints regarding the nature of the GCE.

Working plan & Milestones (including final thesis):

  1. Literature study and introduction to the GCE and the methodology
  2. Include energy information in training data generation (e.g. building on NPTFit-Sim github.com/nickrodd/NPTFit-Sim)
  3. Develop and implement a strategy for incorporating the energy associated with each photon count into recent CNN-based analysis frameworks for the gamma-ray sky
  4. Generate training data, train neural network, optimise neural network architecture and hyperparameters
  5. Validate the method using simulated photon-count maps; then, apply it to the real Fermi data and draw conclusions as to the origin of the GCE
  6. Write master thesis (and, depending on the progress, a short paper on the results)

Requirements / special skills: 1) Should have followed lectures on astrophysics. 2) Advanced Python programming skills are a bonus. 3) Some prior experience with machine learning techniques (in particular neural networks) is a plus.

References:

[1] www.annualreviews.org/doi/10.1146/annurev-nucl-101916-123029

[2] arxiv.org/abs/1506.05124
[3] arxiv.org/abs/2002.12370
[4] arxiv.org/abs/2006.12504
[5] arxiv.org/abs/2107.09070
[6] arxiv.org/abs/2010.10450