A feasibility study on distinguishing fluor concentrations in liquid scintillators from scintillation events observed by photomultiplier tubes using convolutional neural networks

Linear alkyl benzene-based liquid scintillators (LSs) have been extensively used as targets for neutrino detectors in recent decades owing to their environmentally friendly properties, high light yield, and cost efficiency. Neutrino events are typically reconstructed from scintillation events observed by photomultiplier tubes (PMTs) attached to the detector. A comprehensive understanding of the LS response is required for interpreting reconstructed neutrino events during detector operation. In this study, we investigate the properties of scintillation events such as light yield, waveform, and wavelength shift of the emitted scintillation light at various concentrations of fluor dissolved in the LS. The light yield, waveform, and wavelength shift exhibit a nonlinear relationship with fluor concentration, complicating the determination of fluor concentration from the observed characteristics of the scintillation events. We employ a convolutional neural network (CNN) to model this nonlinear relationship between fluor concentration and LS properties. The CNN learns the distinctive features of the scintillation events from observed waveforms and the relative ratio of the light yield below 425 nm to the total light yield detected by a PMT at different fluor concentrations. The trained CNN was able to distinguish the scintillation events with different 2,5-diphenyloxazole and 1,4-bis(2-methylstyryl)benzene concentrations according to the observed waveform and relative wavelength shift. The classified scintillation events for each LS sample exhibited clear features for the different LS concentrations, emphasizing the discriminative capability of the trained CNN. This research presents the first demonstration of LS fluor concentration discrimination using machine-learning techniques in PMT-based detectors.