Optimization Of A K-nearest Neighbors Regression Algorithm For Improved Pulse Shape Discrimination Of Gamma Rays And Neutrons In Organic Scintillators

Through the process of pulse shape discrimination (PSD), certain organic scintillators, such as EJ-299 and stilbene, have the ability to effectively distinguish gamma rays and neutrons. To improve the PSD performance of multiple organic-scintillator-photomultiplier combinations, including two SiPMs, one PMT, and the aforementioned scintillators, a K-Nearest Neighbors (KNN) regression algorithm has been implemented on datasets of mixed gamma ray and neutron pulses. The KNN algorithm works by regressing on a conventionally calculated pulse shape parameter (PSP), leading to more effective discrimination of gamma rays and neutrons. The KNN regression approach to machine-learning PSD has the distinct advantage of being directly comparable to conventional PSD methods by way of a figure-of-merit (FOM), and is largely void of the bias introduced by pre-labeling pulses as gamma rays or neutrons during the training phases, which is present in classification-based machine-learning PSD. This work aims to optimize this KNN PSD algorithm by investigating the process of converting raw pulses into input features suitable for training. Preliminary results for the Hamamatsu SiPM-EJ-299 combination using 10 equally spaced charge integrated regions of the pulse as the input features have led to a FOM improvement of over 200% in the 200-700 keVee light output range. Other input features investigated in this work include charge integrals of different sizes and quantities, various ratios of pulse widths to maxima, as well as features extracted from Fourier space. Additionally, multiple PSPs beyond the conventional tail to total integral ratio will be investigated as the regression parameter. This optimization in both input features and regression parameters will be performed on multiple detector-light sensor combinations with varying levels of innate PSD performance. Optimization of this approach will allow for an algorithm that improves the PSD FOM and lowers the light output threshold at which particles can be still effectively distinguished.