Displaced muon detection using Machine Learning

In the search of new physics scenarios, many modern particle physics theories have proposed the extension of the Standard Model (SM), the current understanding of the elementary particles and their forces consisting the universe. Some of these extensions predict the existence of long-lived particles (LLPs), which decay further away from the Interaction Point (IP) producing displaced particle signatures. Unfortunately, studies around them have been hindered by a severe lack of available statistics. To address this issue, this analysis focuses on utilising machine learning (ML) algorithms in the form of Convolutional (CNN) and Recurrent Neural Networks (RNN) with the goal of triggering not only on prompt, but also on displaced muon event signatures using hit data from the Monitored Drift Tube (MDT) chambers of the ATLAS Muon Spectrometer (MS). The algorithms developed took into account the available hardware memory and latency limitations, achieving in some cases a simultaneous background rejection and signal detection efficiency for both prompt and displaced muons greater than 90%, although failing to make a distinction between the two. Training and validation on simulated data from dark photon samples of varying mass and lifetimes, as well as real background data, indicated good generalisation capabilities. The models were successfully converted into High Level Synthesis (HLS) for implementation on Field Programmable Gate Arrays (FPGAs) and post-conversion results showed a viability in detector-level deployment, despite the loss in performance. This study serves as a foundation for ML-based triggering on displaced muon signatures, showing the potential for integration in future ATLAS upgrades, if the necessary optimisation standards are met.