Track Finding with Graph Neural Network classification in the MEG II Experiment

We present the implementation and performance of a Graph Neural Network (GNN) hit classifier applied to boost the performances of the Track Finder algorithm of the MEG II experiment, improving positron tracking capabilities at high beam intensities. The algorithm classifies detector hits using a heterogeneous graph neural network architecture that processes hits from both the cylindrical drift chamber (CDCH) and the pixelated timing counter (pTC), the subdetectors composing the MEG II spectrometer. This novel track finder algorithm achieves a higher tracking efficiency at all beam intensities, as well as a better resolution on kinematic observables, allowing to improve the experimental sensitivity on the µ+ → e+ γ search.