Using machine learning to improve neutron identification in water Cherenkov detectors

Area of Science:

• Particle Physics
• Experimental Physics
• Neutrino Physics

Background:

• Gadolinium is added to Water Cherenkov detectors like Hyper-Kamiokande to improve neutron detection.
• Neutron detection aids in distinguishing neutrino and anti-neutrino interactions and reduces backgrounds for proton decay searches.
• Neutron signals are subtle and can be mistaken for background noise from muon spallation.

Purpose of the Study:

• To optimize neutron capture detection in the Hyper-Kamiokande intermediate water Cherenkov detector (IWCD) using machine learning.
• To benchmark machine learning models against traditional statistical methods for improved classification accuracy.

Main Methods:

• Development and application of boosted decision tree (XGBoost), graph convolutional network (GCN), and dynamic graph convolutional neural network (DGCNN) models.
• Benchmarking machine learning models against a statistical likelihood-based approach.
• Feature engineering and analysis using SHAP (SHapley Additive exPlanations) to understand model decision-making.

Main Results:

• Machine learning models achieved up to a 10% increase in classification accuracy compared to the statistical approach.
• SHAP analysis provided insights into key features driving event type classification.
• The study utilized a dataset of approximately 1.6 million simulated particle gun events.

Conclusions:

• Machine learning techniques offer a significant improvement in neutron capture detection for neutrino experiments.
• Further development with more realistic datasets is necessary for real-world data analysis.
• Consideration of class imbalance techniques may be required for future analyses of experimental data.