Classifying photons with machine learning in ALICE

Abstract

The ALICE detector at CERN is used to study collisions between heavy ions, which can create a high-energy quark-gluon plasma as they collide inside the detector. In this research the Boosted Decision Tree algorithm is applied to distinguish electron-positron pairs created by the conversion of photons emitted by this plasma, from background consisting of falsely identified ”pairs” of electrons and positrons which do not originate from a photon. The algorithm is trained on over 1.5 million photon candidates generated by a Monte Carlo simulation. Suitable variables for training are determined, data separated into bins to ensure consistency and a K-S test is performed to confirm that the algorithm is not subject to overtraining. Comparison with traditional cuts on the same data show that this BDT method provides a 30% purity increase at maximum significance.