Methodology and performance comparison of statistical learning pulse shape classifiers as demonstrated with organic liquid scintillator

In this paper, we present novel methods for automated pulse shape discrimination. The classifiers are trained using simple radionuclide sources and do not require ground truth labeling. We test their performance using labels derived from time of flight experiments and present the results in terms of energy-dependent Receiver Operating Characteristic (ROC) curves. In addition, we also train and test standard pulse shape discrimination methods on the same data for comparison. We find multiple methods that can yield similar false neutron and true neutron rates at 24keVee as tail-to-total or Gatti's optimal linear filter yield at 54keVee.