Deep Neural Network Extension to Kinematic Likelihood Fitter

As a part of the experiments of the ATLAS collaboration, protons are accelerated and brought to collision in the Large Hadron Collider (LHC) at CERN. So-called top-quark pair production events take place in which a top-quark and an antitop-quark are created from the parton interactions from the underlying proton collision. These particles then decay into new particles in the final state. To study the top-quark, the decay products must be reconstructed in the detector which is complicated by background processes and measurement uncertainties. In order to obtain the best possible reconstruction of the top-quark events, the permutations of possible particle assignments and their probabilities are calculated with the help of the Kinematic Likelihood Fitter. Based on Kinematic Likelihood Fitter and a boosted decision tree enhancement to it, in a first step a deep neural network was developed to reconstruct and improve the evaluation of the most probable particle permutation. Subsequently, another deep neural network architecture was developed which allows statements about parts of the reconstruction. This enables users to filter out a larger dataset of relevant events from the originally measured data, which can be tailored to their problem and has a higher purity and precision.