Recurrent neural network TrackNETv3 for building of the track-candidates on the BM@N experiment

The work is devoted to the development of a neural network approach for the problem of track recon-struction. This approach is successfully used for Monte-Carlo simulation of the BESIII experiment in the form of TrackNETv2 model. However, due to the peculiarities of the GEM detector of the BM@N experi-ment, a number of problems arise. Modifications of the TrackNETv2 neural network model are proposed to solve them, and changes are also made to the training process of the model. At the moment, the best results are achieved by a fully recurrent neural network model and a network using causal convolution. As a result of testing, the best model showed the track efficiency equal to 0.9830. Also, the process of the model was optimized by using the Faiss library for efficient similarity search.