Displaced Muon Detection Using Machine Learning : ATLAS L0 Muon Trigger Upgrade

Sammanfattning

This work presents the development and evaluation of lightweight convolutional (CNN) and recurrent neural network (RNN) architectures for real-time muon triggering in the ATLAS Muon Spectrometer using Monitored Drift Tube (MDT) hit data. Addressing a critical need in modern particle physics, the models were specifically designed to efficiently detect not only prompt muons but also displaced muons, which are key signatures for new physics scenarios such as long-lived particles. The architectures were optimized to meet stringent hardware constraints, achieving high classification accuracy for both prompt and displaced muons, with true positive and truen egative rates exceeding 0.9 in some cases. Extensive training and validation on simulated data—including dark photon decays with varying parameters—as well as real background data, demonstrated robust generalisation and the ability to distinguish between different muon signatures. The neural networks were successfully converted to high-level synthesis (HLS) for FPGA implementation, with post-conversion validation confirming the viability of deployment despite anobserved reduction in accuracy. This study establishes a foundation for machine learning-based trigger systems in high-energy physics, highlighting the importance of displaced muon detection and the potential for further optimisation, integration with additional detector information, and expansion toward adaptive, physics-aware triggering strategies in future ATLAS upgrades.