Multi-class classification and population studies of unassociated Fermi-LAT gamma-ray sources with machine learning

• Applicant: Dr. Dmitry Malyshev
• Subject Area: Astrophysics and Astronomy
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 524110727

About one-third of the gamma-ray sources detected with the Fermi Large Area Telescope (LAT) are not associated with known astrophysical objects. The main goal of this proposal is to investigate the nature of these puzzling gamma-ray sources using multi-class classification with machine learning (ML). In the past, ML algorithms have been used to perform the classification of unassociated gamma-ray sources into 2 or 3 classes, which is too few to capture the rich variety of different source classes in the Fermi LAT data. We propose to develop a framework that paves the way for a multi-class classification of gamma-ray sources into more than 3 classes. The proposed approach consists of several parts: (1) Develop a procedure for multi-class classification using hierarchical definition of classes. This will allow us to control the minimal size of the classes used in the classification and to compare the performance of the classification with different numbers of classes. (2) Domain adaptation for unassociated sources. In the training and classification evaluation, the distributions of the samples in the training and target sets are assumed to be identical, which is not the case for associated and non-associated gamma-ray sources. Here, we will adapt the multi-class classification method and realistically estimate its performance taking into account the differences in the distributions of associated and unassociated sources. As a result of the first two parts, we will produce a catalog with probabilistic multi-class classification of Fermi-LAT gamma-ray sources. (3) Probabilistic population studies of unassociated sources. We will determine the distributions of different classes of sources including the unassociated sources as a function of source parameters, such as the position on the sky. In particular, we will compare the expectations for the pulsar explanations of the 3 GeV excess of gamma-rays near the Galactic center with the distribution of pulsar candidates near the Galactic center. We will also search for new classes of gamma-ray sources (such as dark matter haloes) by looking for new components in the distributions of sources. (4) Probabilistic association of sources using Bayesian and machine learning methods. A more detailed insight into the nature of unassociated sources will be obtained by investigating the probabilistic association of sources using Bayesian and machine learning methods. We will take into account several source parameters in addition to the location on the sky in order to improve the associations of sources compared to methods, which use only the spatial coincidence of sources.

• DFG Programme: Research Grants