Speaker
Description
Linear alkyl benzene (LAB)-based liquid scintillator (LS) has been widely used as the target for neutrino detectors in recent decades due to its environmentally friendly and economic characteristics. The reconstruction of neutrino events is based on the scintillation light emitted from the LS; thus, understanding the LS response helps to understand the reconstructed neutrino event. It has been reported that the timing, light yield, and wavelength shift of the emitted scintillation light are influenced by the concentration of the fluors dissolved in the LS. However, the timing property and wavelength shift exhibit a non-linear relationship with the fluor concentration, making it difficult to distinguish the fluor concentration. In this study, we employed a convolutional neural network (CNN) to model the non-linear relationship between fluor concentration and LS properties. The network learned the featured characteristics of the scintillation events through observed waveforms and the relative ratio of the light yield below 425 nm to the total light yield (short-passed ratio) detected by a photomultiplier tube (PMT) at the different fluor concentration. The trained CNN was able to discriminate the scintillation events with different PPO concentrations concentration from the observed waveform. When the information from the short-passed ratio was combined with the waveform information, scintillation events were distinguished from samples with different PPO and bis-MSB concentrations. The classified scintillation events for each LS sample exhibited clear characteristics for the different LS concentrations, demonstrating the discrimination power of the trained CNN. This is the first demonstration of LS concentration discrimination using a machine-learning technique.
