Graph neural network for 3D classification of ambiguities and optical crosstalk in scintillator-based neutrino detectors

Deep-learning tools are being used extensively in high energy physics and becoming central the reconstruction of neutrino interactions particle detectors. In this work, we report on performance a graph neural network assisting with set event reconstruction. The three-dimensional tracks produced can be subject to ambiguities due multiplicity signatures detector or leakage signal between neighboring active volumes. Graph networks potentially have capability identifying all these features boost performance. As an example case study, tested network, inspired by graphsage algorithm, novel 3D-granular plastic-scintillator detector, that will upgrade near T2K experiment. developed has been trained diverse interaction samples, showing very promising results: classification track voxels done efficiencies purities 94%--96% per most identified rejected, while robust against systematic effects.