Deep Learning Applications to Particle Physics: from Monte Carlo simulation acceleration to ProtoDUNE reconstruction

Abstract

The thesis arises in the context of deep learning applications to particle physics. The dissertation follows two main parallel streams: the development of hardware-accelerated tools for event simulation in high-energy collider physics, and the optimization of deep learning models for reconstruction algorithms at neutrino detectors. Event generation is a central concept in high-energy physics phenomenology studies. The state-of-the-art software dedicated to Monte Carlo simulation is often written for general-purpose computing architectures (CPUs), which allow great flexibility but are not compatible with specialized accelerating devices, GPUs. We present two original tools, PDFFlow and MadFlow, that manage to combine these two aspects in Python. PDFFlow, is a Parton Distribution Functions interpolator, while MadFlow aims at building a complete tool suite to accelerate the whole event generation framework. The reconstruction pipeline at neutrino detectors is comprised of many different algorithms that work in synergy to extract a high-level representation of detector data. All the most important experiments in neutrino physics are developing software to automatically process and extract this information. This work describes the implementation of deep learning techniques to improve neutrino reconstruction efficiency at the ProtoDUNE-SP detector. Two original contributions are presented concerning raw data denoising and a hit-clustering procedure named "slicing". Both denoising and slicing involve the implementation and the training of novel neural network architectures, based on state-of-the-art models in machine learning, such as feed-forward, convolutional and graph neural networks. They represent a proof of concept that these models are indeed capable of providing an important impact on signal reconstruction at neutrino detectors.