Speaker
Description
We present a set of algorithms purely based on machine learning to reconstruct the impact position, the arrival direction and the energy of particles arriving on a cube of plastic scintillator, readout by 32 photo-multiplier tubes, evenly distributed on the edges of the cube. The algorithms allow also distinguish electrons from protons and nuclei. The main element of the reconstruction chain is a fully connected neural network, taking as input the signal of photo-multiplier tubes, and giving as output polar and azimuthal angles in the local frame, energy and particle-type flag. The network is trained with a sample of events produced with a Geant4-based Monte Carlo simulation. Remarkably, significant information is drawn out from a block of scintillator as small as 15 cm, a distance over which plastic scintillators are almost transparent and light production gradients are difficult to measure.
