Using Machine Learning to Separate Cherenkov and Scintillation Light in Hybrid Neutrino Detector

TL;DR

This work tackles the challenge of separating Cherenkov and scintillation light in a hybrid WbLS neutrino detector by using Rat-Pac/Geant4 simulations to generate labeled Cherenkov and scintillation data and applying XGBoost for binary classification. The study demonstrates a robust separation performance with a ROC of $0.96 \pm 1.2\times 10^{-4}$ and provides detailed feature-importance analysis via SHAP, revealing that the number of hits dominates the decision while mean hit time is less influential. Spatial concentration near detector walls degrades performance, which SHAP-guided coordinate cuts mitigate, yielding Cherenkov accuracy as high as $98.4\%$ and scintillation accuracy of $92.9\%$. These results validate the potential of AI-based separation for hybrid detectors and inform design choices for PMT layouts and light-yield configurations in future neutrino experiments.

Abstract

This research investigates the separation of Cherenkov and Scintillation light signals within a simulated Water-based Liquid Scintillator (WbLS) detector, utilizing the XGBoost machine learning algorithm. The simulation data were gathered using the Rat-Pac software, which was built on the Geant4 architecture. The use of the WbLS medium has the capability to generate both Scintillation and Cherenkov light inside a single detector. To show the separation power of these two physics events, we will use the supervised learning approach. The assessment utilized a confusion matrix, classification report, and ROC curve, with the ROC curve indicating a performance result of $0.96 \pm 1.2\times 10^{-4}$. The research also aimed to identify essential parameters for effectively distinguishing these physics events through machine learning. For this, the study also introduced the SHAP methodology, utilizing game theory to assess feature contributions. The findings demonstrated that the number of hits has a significant effect on the trained model, while the mean hit time has a somewhat smaller impact. This research advances the utilization of AI and simulation data for accurate Cherenkov and Scintillation light separation in neutrino detectors

Figures (7)

• Figure 1: Visualization of the detector and its inner PMTs using the RAT-PAC simulation framework. The detector is 3 meters in height and width. The WbLS medium contains 5 % LS and 95% water. 174 PMTs are placed on detector walls.
• Figure 2: The figure shows the outputs obtained as a result of the simulation. In the figures, the Scintillation dataset's outcomes are depicted in purple, whereas the Cherenkov dataset's results are represented in blue.
• Figure 3: The plots show the evolution of the classification algorithm through iteration based on the training and validation data.
• Figure 4: The figure shows XGBoost's confusion matrix. The top row predicts entire Scintillation events (negative case) and the bottom row Cherenkov events (positive case). The x-axis represents model training predictions, and the y-axis represents the original dataset's true cases. The top left is true negative, the top right is false negative, the bottom left is false positive, and the bottom right is true positive.
• Figure 5: The Receiver Operating Characteristic (ROC) curve displays the True Positive Rate (TPR) on the vertical axis, while the False Positive Rate (FPR) is shown on the horizontal axis of the curve. The Roc accuracy result was acquired using several test and training datasets. The mean ROC value is the average of these ten distinct ROC accuracy results.