Dipartimento di Fisica

Descrizione del progetto

Deep learning (DL) has been successfully applied in various domains such as medical image analysis, remote sensing, computer vision, engineering and astronomy [1-5]. Also the astroparticle physics community has recently started to successfully adopt DL algorithms for a wide range of tasks [6]. Most of these attempts are nevertheless in an embryonic phase and the many technical difficulties often prevent the full exploitation of DL potentialities.

The aim of our project is to ENHANCE THE APPLICATION OF DEEP LEARNING TECHNIQUES to the field of EXPERIMENTAL ASTROPARTICLE PHYSICS, and in particular to develop new and more performing algorithms, which are expected to improve the quality and the reliability of experimental measurements and even extend the existing ones. The first important application concerns the event reconstruction task. We want to investigate the use of DL, specifically Convolutional Neural Networks (CNN) [7, 8], in the context of X-ray polarization measurement from astrophysical sources, as observed by IXPE [9], a NASA space mission launched in 2021, and eXTP [10], a future Chinese mission to be launched in 2027. We will conduct a study aimed to maximize the reconstruction of the X-ray polarization with a pure DL approach. A similar context, where CNN will significantly improve the reconstruction accuracy, involves Imaging Atmospheric Cherenkov Telescopes (IACT), i.e. HESS, VERITAS, MAGIC [11] and the future CTA [12, 13], conceived for the study of very high-energy gamma-rays from ground. We plan to optimize a CNN to extract the species, energy, and incoming direction of primaries from the raw detected images, without introducing any cleaning or analytic parameterization.

DL could be efficiently employed also for the challenging task of event selection, to discern among the different species of cosmic rays (CR) hitting a detector. For this purpose we want to introduce an Unsupervised Learning approach [15, 16], exploiting its potential in detecting patterns and relations without any guidance and enabling an approach completely independent from simulations. We will test this technique on the selection of CR electrons in the events detected by the Fermi-LAT gamma-ray telescope [14]. We will possibly extend this approach also to the IACT event selection.