Speaker
Description
Detecting the signals of very low-pT muons with traditional track reconstruction algorithms, such as Kalman filters, is very challenging. In case of the decay of a tau lepton decaying into three muons, the signature includes three very low pT muons in the forward region of the CMS detector. Some or all of these muons might not carry enough pT to reach all stations of the CMS muon system. Even for muons reaching the CMS muon system, individual low momentum muon track reconstruction at hardware trigger level is prohibited with the traditional reconstruction algorithms, due to multiple scattering, the nonuniform magnetic field and large combinatorics especially in high pile up environment at the LHC/High-Luminosity LHC. An alternative approach is presented where a Graph Neural Network is trained to make use of the correlation between hits in the muon detectors to detect the presence of the τ -> 3 μ signature. The muon hits form the nodes of the graph and are connected by edges encoding their relative position. Based on this architecture a classifier is developed for use in the upgraded L1 trigger of the CMS detector for the HL-LHC. With this approach, a significant improvement of a factor of 5 to 10 in acceptance for τ -> 3 μ events is achieved compared to previous studies in CMS phase 2 muon system and Level-1 TDR. Support for GNNs designed in pytorch geometric is implemented into the hls4ml toolkit, which enables us to generate a FPGA implementation of the model for use in the L1 trigger.
