Electron Identification with a Convolutional Neural Network in the ATLAS Experiment

The identification of electrons plays an important role for a large fraction of the physics analyses performed at the ATLAS experiment. An improved electron identification algorithm is presented that is based on convolutional neural networks (CNN), a type of machine learning architecture specialized in image recognition. It takes as input the images of the deposited energy in the calorimeter cells around the reconstructed electron candidates for each of the electromagnetic and hadronic calorimeter layers. Additional input features include the same high-level variables that are used by the likelihood (LLH) and deep neural network (DNN) algorithms developed in ATLAS, as well as the information of up to five inner detector tracks that are matched to an electron candidate during its reconstruction. The output of the network corresponds to the probability that a reconstructed electron belongs to six classes of signals and backgrounds. A significant improvement in identification performance is observed when the CNN algorithm is used in the simulation. For example, for a working point that corresponds to the same signal efficiency as the LLH "Loose" working point, the CNN improves the rejection against charged hadrons faking signal electrons, the dominant electron background at the LHC, by factors of 5 to 8 (depending on the electron kinematics) with respect to the LLH. For the most difficult background constituted of electrons originating from heavy flavour hadron decays, the background rejection of the CNN is improved by factors varying between about 2 to 3.5 with respect to the LLH.