Cross-Attention¿Based Reconstruction of Muon Detector Showers in the CMS Endcap Muon Detectors from Low-Level Inputs

We present a machine-learning based reconstruction of muon detector showers (MDS) in the endcap cathode strip chambers (CSCs) of the CMS detector. The standard CSC hit reconstruction is tuned for single muon tracks. In the dense showers that long-lived particles produce when they decay inside the muon system, the number of reconstructed hits (RecHits) decorrelates from the number of true energy deposits, and since the cluster size is the primary discriminating observable of MDS searches, this directly limits their sensitivity. We train a cross-attention transformer that is aware of the detector identifier (DetID) on the digitised cathode strip and anode wire signals, and let it predict a set of hits for each chamber layer. The training target is the set of simulated hits (SimHits) from the Geant4 detector simulation, matched to the predictions by set prediction, so that no per-hit labels are needed. On samples simulated for 2023 detector conditions, the reconstructed hit multiplicity follows the true one with a Pearson correlation coefficient of 0.95 and a 68% response width of 0.05, compared with 0.66 and 0.31 for the classical RecHit reconstruction, and this holds both for long-lived particle signal and for prompt muon bremsstrahlung clusters. In the dense regime that matters for MDS searches the predicted hits are matched to SimHits more efficiently than RecHits at an equal or lower fake rate, and the cluster shape and timing observables rebuilt from them follow the SimHit distributions. Expanding the SimHit head to the deposited energy additionally recovers the summed energy of the cluster, with a Pearson correlation coefficient of 0.89 for long-lived particle showers and 0.82 for prompt-muon showers.