Application of CNNs for classification of gamma ray Healpix images

This work was developed in the context of space-born gamma-ray astronomy, with particular focus on a multi-messenger approach to the detection of astrophysical sources. Gamma-ray astronomy studies violent and extreme phenomena in the universe. These events emit photons at the highest energies, alongside a number of other messengers such as gravitational waves, cosmic rays and neutrinos. Within dedicated networks, observatories share science alerts in real-time to allow follow-up observations from other facilities. The main objective of this thesis is to perform image classification with the use of a deep learning technique, to overcome the limits of standard analysis. Convolutional neural networks are applied to sky images in order to detect gamma-ray bursts, a class of transient astrophysical phenomena. The Healpix projection is used to prevent distortions that are usually caused by orthographic projections. A software pipeline is developed to generate the training dataset. Two spatial resolutions of the sky images are considered, one with more coarse and one with finer pixelation. Therefore, two convolutional neural network models are implemented and trained. P-values are evaluated to associate a statistical significance to the networks classification probability output in order to analyze the models capabilities to minimize the false positive rate with different classification thresholds, and then a test on real data is made selecting sky images centered on temporal and spatial coincidence with gamma-ray bursts detected by other instruments. To conclude, this work demonstrates that it is possible to solve the problem of gamma-ray bursts detection by applying a convolutional neural network on Healpix projected sky images. This application shows undoubtedly good potential, and further improvements have also been identified and planned.