Fast Reconstruction Using Convolutional Neural Networks for Neutrino Oscillation on IceCube

The IceCube Neutrino Observatory, a gigaton ice Cherenkov detector located at the South Pole, detects a high rate of atmospheric neutrinos. The DeepCore array extends IceCube’s detection of atmospheric neutrinos down to GeV-scales, which is the range necessary to measure neutrino oscillations. With the high statistics atmospheric neutrino sample, the reconstruction of GeV-scale IceCube neutrinos must be both accurate and computationally fast. Thus, Convolution Neural Networks (CNNs) were developed and optimized to reconstruct GeV-scale neutrino properties in the IceCube DeepCore detector such as energy, direction of travel, and interaction vertex. Additionally, CNNs were trained to distinguish particle topologies in the detector, to create a full suite of reconstruction parameters needed for a neutrino oscillation analysis on IceCube. Their performance has similar accuracy to leading likelihood-based reconstruction methods with the CNNs providing 103x speed improvement.