New Small Wheel Muon Detector Trigger Processor Hardware and Beyond Standard Model Searches Using Machine Learning Techniques in the ATLAS Experiment at LHC

The Standard Model of particle physics represents the culmination of human understanding in terms of the fundamental building blocks of the universe and how they interact at a fundamental level. Yet, it is still missing several pieces such as solutions for the gauge hierarchy problem, and the mechanisms behind neutrino masses and dark matter candidates, all of which require new physics and Beyond Standard Model (BSM) theories. Pushing those frontiers implies constant improvements both in the experimental capabilities and theoretical understanding of the High-Energy Particle Physics (HEPP) scientific community. While the last breakthrough on this front was arguably one decade ago with the discovery of the Higgs boson, those last 10 years brought explosive growth in terms of technological development in scientific computation, both on the hardware and software side. Machine Learning (ML) methods have seen widespread adoption and success across many fields, aided by the increasingly better performing Graphics Processing Unit (GPU) hardware acceleration. This Ph.D. research aims to study the potential effectiveness of applying Neural Network (NN) implementations to improve different areas of HEPP. On the detector side, I focused on evaluating the feasibility of hardware-level trigger for the newly installed ATLAS New Small Wheel (NSW) end-cap muon detector. The Trigger Processor (TP) operates in-situ on Field-Programmable Gate Array (FPGA) hardware within a 200 ns time budget allowed for making a Level-0 (L0) trigger decision. Those are very stringent requirements and considering how resource-hungry NN-inference is, the model was designed specifically to mitigate those inherent limitations. In the end, a Convolutional Neural Network (CNN) approach was chosen to interpret hits from the NSW’s eight MicroMegas (MM) detector planes and trained to output quantities related to the likelihood for a muon track to originate from the interaction point (IP). The second main topic of this thesis involves the exploration of unsupervised learning as a means to model-agnostic BSM searches. Provided access to data generated by well-understood SM processes, a NN is trained to compress the events into a lower-dimensional representation and use it to reconstruct the original input. This type of model is called Autoencoder (AE), and a high error of an event’s reconstruction is a potential indicator that the event is not part of the same distribution as the training data, namely SM data. Combining such a model with NN-based density estimation results in an ensemble called Probabilistic Autoencoder (PAE) and this represented the central technique employed in this research. The PAE was tested as a potential tool for identifying new physics in a model-independent way in the context of jet physics and showed promising results.