Novel usage of deep learning and high-performance computing in long-baseline neutrino oscillation experiments

Deep-learning methods are playing a crucial role in numerous scientific and industrial applications. Over the past two decades, these techniques have helped in the collection, reconstruction, and analysis of large data samples in particle physics experiments. The main topic of this PhD research is the study of deep-learning techniques in long-baseline neutrino oscillation experiments. Neutrinos are mysterious light elementary particles, and their investigation is essential to shed light on some of the remaining open questions in physics. The work presented here describes an algorithm based on a convolutional neural network developed to provide highly accurate and efficient selections of electron neutrino and muon neutrino interactions in the Deep Underground Neutrino Experiment (DUNE). With this algorithm, the electron neutrino (antineutrino) selection efficiency peaks at 90% (94%) and exceeds 85% (90%) for reconstructed neutrino energies between 2-5 GeV. The selection efficiency for muon neutrino (antineutrino) interactions is found to have a maximum of 96% (97%) and exceeds 90% (95%) efficiency for reconstructed neutrino energies above 2 GeV. When considering all electron neutrino and antineutrino interactions as signal (both those appearing from oscillations and those intrinsic to the beam), a selection purity of 90% is achieved. These event selections are critical to maximise the sensitivity of the experiment to $CP$-violating effects, key to further understand the matter-antimatter asymmetry of the Universe. In high-energy physics experiments, deep learning has also been explored for producing fast simulations and physically-motivated manipulations of simulated images. Some of those simulations, such as the light production and detection, are very computationally expensive and require novel methods to produce the necessary samples while controlling the varied underlying physics model parameters. To do so, we invented the model-assisted generative adversarial network (MAGAN), first validated on simple generic case studies and then successfully applied to the DUNE photon-detector simulation. Moreover, we also developed graph neural networks for 3D-voxel classification of ambiguities and optical crosstalk for a different particle physics experiment, most precisely for the proposed SuperFGD. This novel 3D-granular plastic-scintillator neutrino detector will be used to upgrade the near detector of the T2K neutrino oscillation experiment, and our method reports efficiencies and purities of 94-96% per event in the classification of particle track voxels. Due to the growth and complexity of deep neural networks, researchers have been investigating techniques to train those networks in a more computationally-efficient way. Many efforts have been made by the community to optimise deep-learning models by parallelising or distributing their training computation across multiple devices. In this thesis, we study an approach based on data locality for those neural networks that cannot benefit from scaling their computation due to a significant bottleneck in the data I/O. The research also includes a detailed study on the performance of deep neural networks on hardware accelerator boards.