Classifying Unidentified Gamma-ray Sources

TL;DR

This paper applies machine learning techniques to classify unidentified gamma-ray sources from Fermi-LAT data, aiming to identify their nature and guide follow-up observations, especially targeting radio-quiet millisecond pulsars.

Contribution

It introduces a statistical classification approach using logistic regression and neural networks to determine the nature of unassociated gamma-ray sources.

Findings

Successful classification of sources based on gamma-ray data

Identification of potential radio-quiet millisecond pulsar candidates

Enhanced target selection for multi-wavelength follow-up

Abstract

During its first 2 years of mission the Fermi-LAT instrument discovered more than 1,800 gamma-ray sources in the 100 MeV to 100 GeV range. Despite the application of advanced techniques to identify and associate the Fermi-LAT sources with counterparts at other wavelengths, about 40% of the LAT sources have no a clear identification remaining "unassociated". The purpose of my Ph.D. work has been to pursue a statistical approach to identify the nature of each Fermi-LAT unassociated source. To this aim, we implemented advanced machine learning techniques, such as logistic regression and artificial neural networks, to classify these sources on the basis of all the available gamma-ray information about location, energy spectrum and time variability. These analyses have been used for selecting targets for AGN and pulsar searches and planning multi-wavelength follow-up observations. In particular, we have focused our attention on the search of possible radio-quiet millisecond pulsar (MSP) candidates in the sample of the Fermi-LAT unidentified sources. These objects have not yet been detected but their discovery would have a formidable impact for our understanding of the MSP gamma-ray emission mechanism.