Event reconstruction of Compton telescopes using a multi-task neural network

We have developed a neural network model to perform event reconstruction of Compton telescopes. This model reconstructs events that consist of three or more interactions in a detector. It is essential for Compton telescopes to determine the time order of the gamma-ray interactions and whether the incident photon deposits all energy in a detector or it escapes from the detector. Our model simultaneously predicts these two essential factors using a multi-task neural network with three hidden layers of fully connected nodes. For verification, we have conducted numerical experiments using Monte Carlo simulation, assuming a large-area Compton telescope using liquid argon to measure gamma rays with energies up to 3.0 MeV. The reconstruction model shows excellent performance of event reconstruction for multiple scattering events that consist of up to eight hits. The accuracies of hit order prediction are around 60% while those of escape flags are higher than 70% for up to eight-hit events of 4 pi isotropic photons. Compared with two other algorithms, a classical model and a physics-based probabilistic one, the present neural network method shows high performance in estimation accuracy particularly when the number of scattering is small, 3 or 4. Since simulation data easily optimize the network model, the model can be flexibly applied to a wide variety of Compton telescopes.