Supervised learning of photoelectron counting in scintillator-based dark matter experiments

TL;DR

This paper demonstrates that a neural network can accurately count photoelectrons in scintillator detectors, outperforming traditional Bayesian methods, and improve event position reconstruction in dark matter experiments.

Contribution

It introduces a supervised learning approach using a multi-layer perceptron to count photoelectrons directly from raw pulse features, enhancing accuracy over existing methods.

Findings

Neural network outperforms Bayesian methods in PE counting.

Improved position reconstruction accuracy in MiniCLEAN.

Applicable to fast scintillators with overlapping PEs.

Abstract

Many scintillator based detectors employ a set of photomultiplier tubes (PMT) to observe the scintillation light from potential signal and background events. It is important to be able to count the number of photoelectrons (PE) in the pulses observed in the PMTs, because the position and energy reconstruction of the events is directly related to how well the spatial distribution of the PEs in the PMTs as well as their total number might be measured. This task is challenging for fast scintillators, since the PEs often overlap each other in time. Standard Bayesian statistics methods are often used and this has been the method employed in analyzing the data from liquid argon experiments such as MiniCLEAN and DEAP. In this work, we show that for the MiniCLEAN detector it is possible to use a multi-layer perceptron to learn the number of PEs using only raw pulse features with better accuracy and precision than existing methods. This can even help to perform position reconstruction with better accuracy and precision, at least in some generic cases.