Enhancing Neutrino Event Reconstruction with AI-based Photon Separation

17th International Conference on Nuclear Structure Properties (NSP2025), Sivas, Türkiye, 25 - 27 Temmuz 2025, ss.103-107, (Tam Metin Bildiri)

ABSTRACT

In neutrino physics, the classification of Cherenkov and scintillation photons in water-based liquid

scintillators (WbLS) plays a critical role. Accurate separation of these photon types is essential for reliable

energy reconstruction, particle identification, and the effective discrimination of signal from background

noise in neutrino interactions. While traditional approaches relying on predefined kinematic thresholds offer

a basic means of photon discrimination, recent advancements in machine learning (ML) techniques have

demonstrated significantly superior classification performance.

In this study, we conducted a comprehensive evaluation of multiple ML-based classification algorithms and

benchmarked them against conventional rule-based methods. The analysis utilized a range of input features,

including photon arrival times, energy levels, and the spatial coordinates of photomultiplier tubes (PMTs)

within the detector geometry. Thirteen high-performing machine learning algorithms, chosen for their

relevance and demonstrated effectiveness in related fields, were trained and evaluated. Among these,

Random Forest, XGBoost, and LightGBM emerged as the most accurate models. These top-performing models

were subsequently fine-tuned through hyperparameter optimization to further improve classification

accuracy. Additionally, a stacked ensemble strategy combining the outputs of the three models yielded the

highest overall performance, highlighting the potential of ensemble learning in photon-type discrimination

tasks in neutrino experiments.

Keywords: Neutrino physics; Water-based liquid scintillator (WbLS); Photon classification; Machine learning;

Ensemble learning; Cherenkov photons; Scintillation photons.