Maximizing CNN potential to predict X-ray polarization through photoelectron production in GPDs

We present the results of our study which aims at maximizing the performance of Convolutional Neural Networks (CNN) to predict the X-ray polarization (in the 1-9 keV energy range) from photoelectron tracks detected in Gas Pixel Detectors (GPD). We developed a fast multi-head ensemble hexagonal CNN and trained it with Geant4 simulated events. We find that providing the initial impact point, derived analytically, as additional input parameter for each track to the CNN significantly improves the modulation factor, and hence the minimum detectable polarization (MDP), with respect to the analytic reconstruction. This is particularly useful for lower-energy tracks given the typical power-law spectra of the astrophysical sources of X-rays. Through our multi-head ensemble we estimate the uncertainty on the even-by-event polarization predictions, which can be exploited to further improve the polarization estimation by means of known weighted maximum likelihood methods to estimate the polarization state through the Stokes parameters.