Machine learning for digital pulse shape discrimination

Accurate discrimination of neutrons and gamma rays is critical for organic scintillation detectors, especially for detection systems where minimal false-alarm rates are paramount (nuclear non-proliferation). Poor pulse shape discrimination (PSD) necessitates long measurement times, and may also cause inaccurate characterization of emitted neutrons, leading to source misidentification. Digital, data-acquisition, measurement systems using a charge-integration PSD method are commonly used for particle classification. A 2-D, charge-integration PSD method tends to be reasonably accurate, although the separation is typically poor at lower energies (below ∼ 500-keV neutron energy deposited). The charge-integration method originated in analog systems; however, with digital measurement systems there is no need to restrict to only two features (for instance, tail and total integrals) of the pulse. Instead, a classifier may be a much more complex function of the measured pulse. In this work, we apply a machine-learning methodology; namely, the support vector machine (SVM), to determine a PSD classifier. We show that the SVM method leads to improved detection performance with respect to the charge-integration method. We also apply a recently developed methodology that gives more accurate performance estimates by accounting for the fact that the training data needed for the SVM are 'contaminated'.