Improving ATLAS Muon Segment Regressor with Deep Learning

As part of the CERN Summer Student Programme 2025, I worked on improving the ATLAS muon segment regressor using deep learning. The High Luminosity Large Hadron Collider will increase the pile-up from 60 to 200, requiring upgrades to the ATLAS detector and the Trigger and Data Acquisition system. The muon segment reconstruction is one of the steps in the Event
Filter, and it faces several challenges, such as background noise and ambiguities. To deal with these challenges, we used graph neural networks for reconstructing the muon segments. During the programme, I developed and evaluated a one-segment regressor and a two-segment regressor, and worked on the optimization of the hyperparameters of the model. As a result, the regressor models predicted the true segment parameters with high accuracy, with all the errors mostly less than 2%.