Coherent Muon to Electron Transition (COMET) phase-i local filtering by catboost algorithm for track reconstruction

Abstract

Coherent muon to electron transition (COMET) experiment is an exclusive beamline for studying charge lepton flavor violation through investigation of neutrinoless muon to electron transition. The present work aims to classify signal electron and background from the truth level data generated from GEANT4 simulation (MC5 file) using CatBoost algorithm. This data was first simulated in the Integrated Comet Experimental Data User Software Toolkit (ICEDUST) framework to extract electron and background samples of the main COMET detector, CyDet. Both electron and background samples are merged and the detector response towards this sampling are calibrated using the previous MC4 file. Subsequently, the muon stopping region, bunch width effect, overflow of hits, trigger acceptance and occupancy parameters are observed. The data was sanitized by applying energy cut to the energy deposited on cylindrical drift chamber (CDC) and Cherenkov trigger hodoscope (CTH). Four local features (charge deposited on CDC wire, radial distance of hit from muon stopping target (MST), relative time to the trigger signal, and angle of hit from x-axis) and four neighbour features (charge deposited on right wire, charge deposited on left wire, time relative to the trigger signal on right wire, and time relative to the trigger signal on left wire) are calculated. Using these selected features along with CatBoost algorithm, 94.2% of background hits are removed, whereas 93.7% of hits signal are retained. Performance study using confusion matrix and features importance shows that radial distance from MST gives the highest contribution in the classification of signal and background. Application of machine learning in particle physics is very useful in predicting the experimental sensitivities and processing of big data analysis.