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

Area of Science:

• Nuclear Instrumentation and Measurement
• Applied Machine Learning
• Radiation Detection Physics

Background:

• Characterizing neutron radiation fields is challenging due to detector sensitivity and secondary particle generation in mixed environments.
• Pulse shape discrimination (PSD) in liquid scintillators is crucial for identifying neutrons amidst other radiation types.

Purpose of the Study:

• To develop and evaluate machine learning architectures for enhanced pulse shape discrimination in liquid scintillators.
• To investigate the impact of digitizer sampling parameters on machine learning algorithm performance for neutron detection.

Main Methods:

• Utilized EJ-301 liquid scintillator and a CAEN DT-5743 digitizer (3.2 GHz, 12-bit resolution).
• Employed 238Pu9Be and 241Am9Be neutron sources and 24Na, 60Co, 137Cs photon sources for data generation.
• Tested various machine learning architectures, including a modified GoogLeNet, with varying digitizer sampling rates and bit depths.

Main Results:

• A modified GoogLeNet architecture achieved the highest performance, with a 69.17% true positive rate for neutron identification.
• Achieved exceptional photon rejection rate of 99.9999%.
• Performance varied with neutron energy, ranging from 1.30% at >3 MeVee to 89.96% between 340-1000 keVee, with 21.48% identification below 200 keVee.

Conclusions:

• Machine learning, particularly the optimized GoogLeNet, significantly enhances neutron detection capabilities in mixed radiation fields.
• Full digitizer sampling rate and bit depth are critical for achieving optimal performance in machine learning-based PSD.
• The developed method shows promise for accurate neutron field characterization, especially in lower energy ranges.