Improving IceCube Event Reconstruction Using A Graph Convolutional Network And Semantic Segmentation

Machine learning is a candidate for the next-generation reconstruction for neutrino experiments such as IceCube. IceCube is an ice-Cherenkov neutrino detector embedded in a cubic kilometer of glacial ice in Antarctica. The detector observes astrophysical and atmospheric neutrinos via the light emitted by charged particles produced in neutrino interactions with 5160 digital optical modules~(DOM). A typical $\nu_\mu$ interaction~(1TeV$\sim$100TeV) originating inside the detector, namely the ``starting track'', is dominated by deep inelastic scattering which produces a hadronic cascade near the interaction vertex and a muon track. Conventional reconstruction assumes continuous energy loss for ``starting track'' events in the hadronic cascades, leading to a generally underestimated energy reconstruction. Correct clustering the hadronic cascade and track is crucial for improving the energy reconstruction. This study presents a graph convolutional network for semantic segmentation to distinguish the DOM charges as cascade-like and track-like charges.