Development and commissioning of a machine learning algorithm for real time reconstruction of electromagnetic showers with a scintillating fibres tracker

This thesis approaches the problem of reconstructing electromagnetic showers in real time using a tracking detector interleaved with other layers serving as absorbing material. It finds immediate application in experiments such as SHiP and SND@LHC, which use calorimeters made of emulsion bricks interleaved with scintillating fibre tracking stations. We developed a machine learning algorithm able to reconstruct the energy of an electron which showers through such a detector. The spatial location of the hits detected by the tracker planes are given as input to the algorithm. We train the algorithm using a simulated data sample with electrons energies in the [0.3, 100] GeV range. We also developed a downsizing procedure which allows the algorithm to be used with tracking planes of arbitrary, larger dimensions, such as the ones used in SHiP and SND@LHC. We achieved to use the studied detector as a sampling calorimeter, with a measured stochastic term of $a = 0.665 \pm 0.02$ GeV$^{1/2}$ . We obtain a relative energy resolution of about 10% in an energy range [40, 70] GeV. Even if the main calorimetry information in these experiment will be provided, at the analysis stage, by the emulsion bricks, this algorithm allows to reconstruct showers in real time using information from the tracker. As such, it will prove useful both for neutrino physics, and to spot early hints of the possible presence of dark matter.