Improving Four-Top-Quark Event Classification with Deep Learning Techniques using ATLAS Simulation

A study on the potential of different types of Deep Neural Networks as classifiers for four-top-quark production is presented. The used data are ATLAS simulated proton-proton collisions at a centre-of-mass energy of 13 TeV. Events are selected if they contain a same-sign lepton pair. A Feedforward Neural Network, using the jet multiplicity, b-tagging information, and features constructed from event kinematics, is optimized and compared to a Recurrent Neural Network, which uses similar information but the raw event kinematics. The largest area under the receiver-operating-characteristic curve observed is 0.852 ± 0.005, for the Feedforward Neural Network, and 0.838 ± 0.006, for the Recurrent Neural Networks. Further improvements of both the training of the Deep Neural Networks and the selected features are investigated.