NERSCPowering Scientific Discovery Since 1974

Deep Networks for Neutrino Telescopes

This page provides a description of an analyses on IceCube Neutrino Detector data using deep neural networks on Cori.array withSears amp DetailedLabels

Problem Description

 The IceCube Neutrino Detector is searching for neutrinos from distant astrophysical sources. Buried 1500m beneath the surface of the South Pole, the detector uses polar ice as it's detection medium. It searches for light created when muons and neutrinos speed through the ice. This is a high background experiment with background events out numbering signal events by a factor of 10 million. The goal is to use deep neural networks to increase the number of signal events while rejecting the maximum number of background events.

Input Data

This analysis will use hdf5 files containing information about each background and signal event. There are two sets of data, event-level and pulse-level. Event-level data contains global data like event number and event time. Pulse-level data contains data like the time and location of light detection (there can be many of these per event since IceCube has ~5160 sensors).

Method

This analysis will use TensorFlow.