Abstract
Three deep tracking methods are presented for the BM@N experiment GEM track detector, which differ in their concepts. The first is a two-stage method with data preprocessing by a directional search in the k-d tree to find all possible candidates for tracks, and then use a deep recurrent neural network to classify them by true and ghost tracks. The second end-to-end method used a deep recurrent neural network to extrapolate the initial tracks, similar to the Kalman filter, which learns necessary parameters from the data. The third method implements our new attempt to adapt the neural graph network approach developed in the HEP.TrkX project at CERN to GEM-specific data. The results of applying these three methods to simulated events are presented.
Highlights
Three deep tracking methods are presented for the Baryonic Matter at Nuclotron (BM@N) experiment GEM track detector, which differ in their concepts
In fixed target experiments, such as the Baryonic Matter at Nuclotron (BM@N) [1], products of heavy ion interactions are recorded by the GEM track detector [2]
The model achieves 99 % accuracy, 92 % recall and 0.983 AUC score [5] for the segments obtained from the Minimum Branching Tree (MBT) algorithm
Summary
In fixed target experiments, such as the Baryonic Matter at Nuclotron (BM@N) [1], products of heavy ion interactions are recorded by the GEM track detector [2]. In the considered case it consists of six coordinate stations formed by sensitive electronic elements, in which a signal appears, induced by a passing particle Those elements are linear parallel thin strips on a silicon substrate. The intersection of the “fired” lines will give us a hit with the coordinates of a space point close to that at which the particle has flown It is worth noting a serious drawback of the strip detectors, due to the fact that, for a large number of particles, in addition to the real intersections corresponding to the point of the passage of a particle, there appear, by an order of magnitude more numerous, false intersections, called fakes, considerably contaminating the results of the measurement. Recall expresses the ability to find all true tracks in a dataset, while precision expresses the proportion of data, our model says was true, were true tracks
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have