Machine Learning Implementation for Tau Neutrino Appearance in the DUNE Far Detector

The Deep Underground Neutrino Experiment (DUNE) is an upcoming long-baseline neutrino oscillation experiment representing the next milestone of steady growth in neutrino physics. With many features surpassing current long-baseline experiments, DUNE would be the first experiment capable of generating a large sample of visible ντ for study of the νμ →ντ oscillation channel. In this project, the DUNE far detector (FD) monte carlo simulation sample is used to study selection efficiency and purity for ντ event classification. Results show that the current classification routine for ντ event selection in the DUNE analysis framework performs poorly and must be improved if a reliable distinction between ντ signal (S) and background (B) is to be achieved. Using a set of weak classifiers derived from reconstructed event statistics, an adaptive boosted decision tree (BDT) is implemented into the FD classification procedure. BDT classification on neutrino events for the tau flavor results in a tenfold increase in selection purity and a threefold increase in signal significance (S/√B) for a 3.5-year simulation sample. A separate χ2 analysis on the BDT output suggests near-discovery-level ντ signal significance in 3.5 years of data-taking across 99% of possible oscillation parameters.