Improving the electron identification of the ALICE detector using machine learning

Abstract

A few moments after the Big Bang, the universe was in a state of matter called the quark-gluon plasma. Nowadays, this state is recreated by colliding heavy ions in particle accelerators. Heavy quarks are used as probes to study this medium created in heavy-ion collisions. Since the quarks decay quickly into other particles information about the properties of the quarks is not directly accessible. However, information about these heavy quarks can be accessed through electrons decayed from hadrons containing heavy quarks. In ALICE (A Large Ion Collider Experiment), one of the main experiments at CERN's Large Hadron Collider (LHC), past research has been performed using rectangular selections for the electron identification. In this project we have developed electron identification models through the use of machine learning to improve upon these rectangular selections. We have applied the machine learning models to detector simulation data of a minimum bias pp collision generated by PYTHIA8 provided by the ALICE collaboration. By fixing the model purity to match the purity of the best performing rectangular selection, the efficiency was improved by up to 37%, depending on the model and transverse momentum of the electron. While fixing the efficiency of the best performing rectangular selection resulted in a 1.0% increase in the purity. In general, the best performing model was XGBoost. To check for the possible bias in differences between simulation and data, a test dataset was shifted with respect to the training sample and the impact on the performance of the models was evaluated. These results imply that machine learning models have the potential to more accurately identify electrons and help in the determination of the properties of heavy quarks and the quark-gluon plasma.