NeutralNet: Development and testing of a machine learning solution for pulse shape discrimination.

Accurate description of radiation fields containing neutrons continues to be a difficult task to complete. This difficulty arises because of the inherent sensitivity of neutron detectors to other types of radiation, and the ability of neutrons to generate secondary particles producing mixed field environments. This research looks at the development and performance of various machine learning architectures when applied to the task of pulse shape discrimination with liquid scintillators. This work was carried out with a neutron sensitive liquid scintillator, EJ-301, with signals digitized at 3.2 GHz with 12 bits of resolution utilizing a CAEN DT-5743 digitizer. Measurements were artificially reduced in sampling depth and frequency to investigate the importance of these parameters for performance of the machine learning algorithms. Two isotopic neutron source, PuBe and AmBe, and three photon sources: Na, Co, and Cs were used for generation of the training and validation sets as well as for energy calibration. The greatest performance was achieved with a lightly altered implementation of GoogLeNet and with the full sampling rate and bit depth afforded by the digitizer. A true positive rate of 69.17 % was achieved while correctly rejecting 99.9999 % of photon events. This performance ranges from 1.30 % at events greater than 3 MeVee to 89.96 % between 340-1000 keVee. For neutron events with energy below 200 keVee 21.48 % of the validation neutrons are still identified as neutrons.