Continuum background suppression using Deep Learning for the Belle II experiment

The underlying reason for matter dominance in the universe lies in the violation of CP symmetry, introduced to the SM by Kobayashi and Maskawa by proposing a six-quark model, which resulted in them winning the Nobel Prize for Physics in 2008. The proof of this fundamental achievement was made possible by the KEKB accelerator, located in Japan: a B-factory producing abundant B-mesons, which, decaying into other lighter mesons, are excellent candidates for investigating CP violation. KEKB is an asymmetric electron-positron collider with 10.58 GeV centre-of-mass energy, i.e. at Y(4s) resonance. The challenge today is to perform increasingly precise measurements of weak interaction parameters, which lurk the secrets to formulating physics beyond the SM. This is what SuperKEKB, the successor of KEKB, aims to achieve by working at the high-intensity frontier. The Belle II experiment runs at Y(4s) resonance and one of its goals is to study charmless decays of the B meson into specific channels. In particular, the signal of interest is not the most likely result of the positron-electron collision: it is necessary to devise efficient and robust techniques to discriminate the signal from the continuum background, namely the hadronisation of lighter quarks. Hence the purpose of the present work. By exploiting the topological differences between background continuum and signal, the parameters that can be used to perform the distinction are constructed. The goal of the work is to develop a more effective approach using Deep Neural Networks (DNN), optimising and comparing it with existing techniques (in particular BDT). The DNN, implemented with the Pytorch framework, aims to discriminate the continuum background from the signal with a simple sequential architecture and an approach that allows using a reduced number of features (instead of the total 361) is investigated, in order to limit potential mismodeling errors in Monte Carlo simulations and to avoid redundant information. Hyper-parameter tuning for the DNN is also performed, in order to choose the optimal number of nodes. The idea is to use low-level variables and check whether the distinction efficiency is comparable to the current one, and how the different variables can be combined in order to maximize performance and decrease correlations. The next step will be to investigate a more complex architecture, based on Convolutional Neural Networks and Graph Networks, exploiting the feature reduction carried out.