A Machine Learning approach for astrophysical X-ray polarimetry with Gas Pixel Detectors

In astrophysics four different techniques are used to study the information carried by electromagnetic radiation: photometry, imaging, spectroscopy, and polarimetry. Photometry refers to the measurement of the flux or intensity of light emitted by celestial objects, whereas imaging focuses on analyzing spatially resolved representations of these objects. Spectroscopy aims at studying the radiation properties as a function of energy, while polarimetry examines the geometrical orientation of electromagnetic field oscillations. X-ray astrophysics has studied and demonstrated the dynamic nature and complex structures of celestial sources through the combined use of the first three techniques, whereas polarimetry has played a marginal role until recent years. Still in the early years of the 21st centuries, the challenging and time-consuming nature of X-ray astrophysical polarization measurements lagged behind the advancements in spectroscopy and imaging. The development of the Gas Pixel Detector (GPD) [1], and with it the opportunity to exploit the photoelectric effect for polarization measurements, changed this narrative, achieving unprecedented sensitivity in the 1-10 keV energy range. The interest of the astrophysical community for the X-ray polarimetry reached a climax with the launch of the Imaging X-ray Polarimetry Explorer (IXPE) on December 9th, 2021, a NASA mission designed to perform astrophysical X-ray polarimetry in the 2-8 keV energy range [2]. Since the beginning of data acquisition, IXPE demonstrated the ability of X-ray polarimetry to access new and groundbreaking information about the properties of a wide variety of celestial objects. IXPE employs three GPDs to determine the polarization properties of incident photons, by analyzing the tracks of photo-electrons (PEs) generated through the interaction of the photons with the gas contained in the detector. The state-of-the-art IXPE data analysis relies on the analytic reconstruction of these tracks [3]. While reading the early chapters of this thesis, it will become evident how this kind of reconstruction well aligns with the use of Machine Learning (ML) algorithms. The application of these techniques to IXPE data reconstruction has already been explored by other few groups, leading to interesting and promising results [4; 5; 6]. However, these works demonstrated that there is substantial room for improvement as well, for both the analysis of IXPE data and for future X-ray polarimetry missions. This thesis addresses the scenario just described, presenting a new algorithm that leverages the promising performance of Convolutional Neural Networks (CNNs) while maintaining the core structure of the standard analytic algorithm. The goal is to create a stable and high-performing model, that could be applicable to current IXPE data, and also benefit future X-ray polarimetry missions and similar applications. Chapter 1 introduces the physics of polarimetry, alongside with the main experimental techniques involved in astrophysical measurements. Chapter 2 starts with an overview of astrophysical X-ray polarimetry and then focuses on IXPE mission and instrument. The Gas Pixel Detector, its components and functioning, as well as the reconstruction of the events, are discussed. Chapter 3 is dedicated to IXPE science, and particular attention is directed towards sources which are functional to test the application of the algorithm developed in this work. Chapter 4 gives a brief introduction to Machine Learning and Convolutional Neural Networks, describing in greater details the models and methods involved in the project of this thesis. Chapter 5 depicts the developing phase of the algorithm and reports the performance achieved with simulated data. Chapter 6 validates the results obtained with simulations by applying the algorithm to laboratory data. Finally, Chapter 7 reports examples of its application with real IXPE data.