Machine learning technique for signal-background separation of nuclear interaction vertices in the CMS detector

The CMS inner tracking system is a fully silicon-based high precision detector. Accurate knowledge of the positions of active and inactive elements is important for simulating the detector, planning detector upgrades, and reconstructing charged particle tracks. Nuclear interactions of hadrons with the detector material create secondary vertices whose positions map the material with a sub-millimeter precision in situ, while the detector is collecting data from LHC collisions. A neural network (NN) with two hidden layers was used to separate secondary vertices due to combinatorial background from those arising from nuclear interactions with material. The NN was trained and tested on data from proton-proton collisions at a center-of-mass energy of 13 TeV, recorded in 2018 at the LHC. NN training is performed using Keras and Matplotlib in a Jupyter notebook. Secondary vertices in the training data are classified as signal or background, based on their geometrical position. Even though the variables used in training show only small differences between background and signal, the NN has impressive separation power. Hadrographies of the CMS inner tracker detector before and after background cleaning are presented.