Neural network implementation of the global method of reconstruction of primary vertices of events for the BESIII inner tracking detector

In this paper, we consider a method for predicting the position of the vertex of an event using a convolutional neural network, and even before the stage of track reconstruction. The recognition of particle trajectories (tracks) from experimental measurements plays a key role in the reconstruction of events in high-energy physics. The internal detector of the BESIII experiment has only 3 cylindrical GEM stations, so the track in the magnetic field registered by only two stations cannot be restored without additional information. The two hits can only determine the direction of movement of the particle, so you need to know the primary vertex of the event. The average absolute deviation (MAE) of the coordinates of the predicted vertex from the true primary vertex of the event known in advance from the simula-tion was used as a metric for determining the quality of the neural network. The trained model can predict the primary vertex of an event with an average absolute error of 0.009. To solve the problem of finding the vertex of an event, in this article we propose to apply the deep convo-lutional neural network LOOT model, which processes all event tracks at once, as a 3D image, and after training, is able to predict the primary vertex with acceptable accuracy.