DeepJet : Generic physics object based jet multiclass classification for LHC experiments

The detectors at the Large Hadron Collider at CERN reconstruct, among other objects, collimated sprays of particles, which are referred to as “jets”. An important task is to identify the type of the elementary particle that initiated the jet, i.e. whether it is a light quark, a heavy quark or a gluon, leading to a multiclass classification problem. We present results from a realistic simulation of one of the two multi-purpose detectors at the LHC, the Compact Muon Solenoid. The basic network architecture relies heavily on using convolutional layers on low-level physics objects, like individual particle objects, and uses much more information than previous algorithms in the literature. It stands out as the first proposal that can be applied to multiclass classification for all types of jet initiators as well as for jets of different widths. We demonstrate significant improvements by the new approach on the classification capabilities for several of the tested particle classes. In one specific case, and at high momentum, a decrease of nearly 90% in the rate of false positives is achieved for a constant true positive rate of 40%.