Enhancing searches for astrophysical neutrino sources in IceCube with machine learning and improved spatial modeling

TL;DR

This paper improves the detection of astrophysical neutrino sources in IceCube by combining advanced spatial modeling, machine learning for PSF prediction, and analyzing specific cosmic ray source regions, enhancing sensitivity and accuracy.

Contribution

It introduces a machine learning approach using gradient-boosted decision trees with vMF likelihood loss for better PSF prediction in neutrino source searches.

Findings

Enhanced spatial modeling improves directional reconstruction.

Machine learning-based PSF prediction increases search sensitivity.

Analysis of specific gamma-ray sources and regions like Cygnus Cocoon.

Abstract

Searches for astrophysical neutrino sources in IceCube rely on an unbinned likelihood that consists of an energy and spatial component. Accurate modeling of the detector, ice, and spatial distributions leads to improved directional and energy reconstructions, resulting in increased sensitivity. In this work, we utilize our best knowledge of the detector ice properties and detector calibrations to reconstruct in-ice particle showers. The spatial component of the likelihood is parameterized either by a 2D Gaussian or a von Mises Fisher (vMF) distribution at small and large angular uncertainties, respectively. Here, we use a gradient-boosted decision tree with a vMF spatial likelihood loss function, reparameterized through two coordinate transformations, to predict per-event point spread functions (PSF). Additionally, we discuss the search for PeV cosmic ray sources using the IceCube Multi-Flavor Astrophysical Neutrino (ICEMAN) sample. Our search contains both an analysis of individual neutrino sources coincident with greater than 100 TeV gamma-ray sources and also a stacking analysis. We outline the prospects for extended neutrino emission originating from the Cygnus Cocoon region.