Machine Learning Techniques for Gamma-Hadron Separation with IceCube - MSc Thesis

Sep 17, 2026

Uploaded by Levi

The IceCube Neutrino Observatory, built mainly to detect astrophysical neutrinos, also enables measurement of ultra-high-energy gamma rays using IceTop and in-ice detector information. IceTop reconstructs air-shower footprints, while the in-ice detector constrains high-energy muon content to suppress the dominant cosmic-ray background. The thesis applies machine learning for gamma-hadron separation, developing a custom IceTop CNN and an extended combined model with additional reconstructed parameters. Results show a large AUC improvement and strong proton rejection in the PeV range.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages98

Reading Time4 h 7 min

AccessFree to download

Introduction

Astroparticle Messengers

Cosmic Rays

Gamma Rays

IceCube Neutrino Observatory

Detection Principles

In-Ice Detector

IceTop Surface Array

Event Reconstruction and Simulation

IceTop Reconstruction

In-Ice Reconstruction

Simulated Datasets

Previous Gamma-Hadron Separation Methods

Machine Learning Techniques

Machine Learning Principles

Neural Networks

Convolutional Neural Networks

Optimising Neural Networks

Boosted Decision Trees

IceTop CNN Development

Data Preparation

Basic Network Architecture

Model Configurations

Alternative Architectures

Physics and Astronomy, School of Physical and Chemical Sciences, University of Canterbury, Christchurch, New Zealand Machine Learning Techniques for Gamma-Hadron Separation with IceCube Rhia Hewett MSc Thesis 2025 Supervisor: Prof. Jenni Adams Co-Supervisor: Dr. Agnieszka Leszczy´nska Abstract While primarily designed for the detection of astrophysical neutrinos, the IceCube Neutrino Observatory has unique instrumentation that can be used for the detection of ultra-highenergy gamma rays. The surface array component, IceTop, measures the footprint of air showers initiated when high-energy particles interact in the atmosphere. The detector array underneath the ice can provide a valuable indication of the high-energy muon content of these air shower events. Combining this information to distinguish gamma-ray induced events from the dominant cosmic-ray background remains a significant challenge. This thesis explores the application of machine learning techniques to enhance gamma-hadron separation using IceCube data. A custom convolutional neural network (CNN) was developed to process the charge, timing and lateral distance information from IceTop pulses. We conducted an extensive trialling procedure to tailor the architecture and configuration of the CNN to best extract features from IceTop pulses. This model was extended with a secondary neural network incorporating other reconstructed parameters. Relevant in-ice parameters were explored and incorporated into the secondary neural network to produce a combined in-ice model. We evaluated the performance of these models using standard machine learning metrics for classification tasks such as Area Under the Curve (AUC) scores, and also in terms of metrics more specific to the gamma-hadron separation task, such as passing fractions. The IceTop-only model achieved weighted area AUC scores of 0.795 (ROC) and 0.747 (PR), while the combined in-ice model demonstrated significantly improved performance, reaching AUC scores of 0.974 (ROC) and 0.966 (PR) . The inclusion of in-ice information about the TeV muon content of air showers proved critical for effective background suppression, reducing proton event passing fractions below 10 −2 in the PeV range while retaining 90% of gamma-ray events. This pilot study demonstrates the viability of machine learning for gamma-ray classification with IceCube and lays the groundwork for future gamma-ray source searches. With further optimisation and integration with other analysis methods, the machine learning approach developed here will enhance IceCube’s sensitivity to PeV gamma-ray sources. Contents 1 Introduction 1 2 Astroparticle Messengers 3 2.1 Cosmic Rays .................................... 3 2.1.1 Cosmic-Ray Accelerators ......................... 5 2.2 Gamma Rays ................................... 5 2.2.1 Extensive Air Showers .......................... 6 2.2.2 Gamma-Hadron Separation ....................... 8 2.2.3 Gamma-Ray Detectors .......................... 10 2.2.4 Gamma-Ray Studies with IceTop .................... 11 3 IceCube Neutrino Observatory 13 3.1 Detection Principles ............................... 13 3.1.1 Energy Losses in Matter ......................... 13 3.1.2 Event Signatures ............................. 14 3.2 In-Ice Detector .................................. 15 3.3 IceTop Surface Array .............................. 16 3.3.1 Local Coincidence and Filtering ..................... 16 3.3.2 Signal Calibration ............................. 18 4 Event Reconstruction and Simulation 19 4.1 IceTop Reconstruction .............................. 19 4.1.1 Core Position and Shower Front Curvature ............... 19 4.1.2 Lateral Distribution Function ...................... 20 4.1.3 Snow Correction ............................. 20 4.1.4 IceTop Fit Procedure ........................... 21 4.2 In-Ice Reconstruction .............................. 21 4.2.1 Muon Energ

FAQs

What detection challenge does the thesis target?

Distinguishing gamma-ray–induced events from the dominant cosmic-ray background in IceCube data remains challenging. The work addresses this through machine learning–based gamma-hadron separation.

What models were developed in the thesis?

A custom convolutional neural network (CNN) was built to process IceTop pulse charge, timing, and lateral distance information. A secondary neural network was then added to incorporate in-ice reconstructed parameters for a combined model.

How much improvement did the combined in-ice model achieve?

The combined in-ice model reached AUC scores of 0.974 (ROC) and 0.966 (PR), outperforming the IceTop-only model. Incorporating in-ice TeV muon content was critical for background suppression.

Sep 17, 2026

Uploaded by Levi