Using Cutting Edge Software and Techniques to Model and Measure Experimental Neutrino Data

Abstract

Neutrinos are fascinating particles whose behaviour could have cosmological implications. For example, by studying them we could understand the imbalance of matter and antimatter in the Universe, one of the crucial puzzles in astrophysics. In the field of neutrino physics, there are many open questions, such as the possibility of CP violation in the lepton sector, the ordering of the neutrino mass states, and the existence of one or more so-called sterile neutrinos. In order to answer these questions, several technological improvements are required. This thesis represents the work done by the author towards this goal. Most of these results will be used to better understand the behaviour of the neutrinos, and those techniques involving artificial intelligence could also aid the technological development of our society in general. Sterile neutrinos are hypothetical particles that cannot interact with anything. To investigate the existence of sterile neutrinos, a new and comprehensive analysis of the MiniBooNE low energy excess of electronlike events is presented. A new model called the “dipole model” is used, which is an extension of the 3+1 model. Alongside the three neutrinos in the Standard Model and a fourth, sterile neutrino participating in the oscillations, a second set of neutrinos Nj with j = 1, 2, 3 is introduced. Every member of this set of neutrinos can interact with photons via the so-called "dipole portal interaction" and decay. Among them, the heaviest, called Heavy Neutral Lepton (HNL) or simply N and with a mass on the MeV scale, might be responsible for the anomaly. Firstly, the creation of the HNL and its subsequent decay N → ν + γ was simulated, obtaining the possible photon distribution inside the detector responsible for the excess. Then, in order to remove contributions to the excess coming from other sources, a global fit using a 3+1 model without MiniBooNE is performed and results are removed from the anomaly. The remaining excess is fitted with the dipole model in both energy and angular distribution. We found a region of parameter space consistent with both distributions at the 95% confidence level near d = 3 × 10−7 GeV−1 and mN = 400 MeV. Recognising neutrino interactions is of fundamental importance in the field, and this has been done with the help of Machine Learning (ML) and its subfield Deep Learning (DL). I developed a new ML tool able to characterise each hit in a detector as a track or shower-like topology, to be used inside the framework “Pandora”. Results showed that a good separation is possible, with a 71% efficiency (tracks) and a 91% background rejection (showers). This work then continued testing the new Edge TPU, a portable device created by Google Coral to perform inference with DL. For the first time, a neutrino physics dataset was used with positive results identifying the Edge TPU as a credible competitor to GPUs. Moreover, I developed a new framework called Collaborative Learning in partnership with the British company Fetch.AI to train a convolutional neural network with different datasets simultaneously. The last part of the thesis is one of the first analyses of ProtoDUNE-SP data and simulation using the Transverse Kinematic Imbalance (TKI) techniques. This novel methodology can give great insights into the internal structure of argon, which is necessary to create better simulations of neutrino-argon interactions. Starting from raw data, I focused on the channels p+Ar → p+p selecting a proton beam and two outgoing child protons. I then obtained the four TKI distributions, with good alignment between data and simulation. These results will help improve the next generation of simulations and our understanding of the interactions of particles with liquid argon. Next-generation experiments such as DUNE and Hyper-Kamiokande are going to have unprecedented numbers of neutrino interactions, so statistical uncertainties will be small and systematic uncertainties are more important than ever.