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 $3.0\,\mathrm{MeV}$. 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 $4π$ 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.