Departmental Colloquia: Kristin P. Lennox

KRISTIN P. LENNOX kristin lennox

 

Director of Statistical Consulting
Lawrence Livermore National Laboratory

 

“Probabilistic Implicit Labeling for Fast Neutron Detection”

 

ABSTRACT

 

Fast neutron detectors are important tools for both nuclear nonproliferation and scientific applications. A key challenge in fast neutron detection is the gamma/neutron discrimination problem, where both gamma rays and neutrons are detected and must be distinguished from one another using subtle pulse features. While there are many classification algorithms available in both the machine learning and statistical literature that might contribute to this problem, most require labeled neutron and gamma pulses, which are not readily available. Alternatives for unlabeled pulse data rely on expert knowledge and tuning of classifiers, and frequently discard the majority of observed pulses due to classification ambiguity.

We present an automatic and accurate method for training a gamma/neutron discrimination algorithm without labeled data. Our method is distinct from an unsupervised, or clustering, approach in that we directly leverage physical information to assess the differences in typical pulse shapes between gamma rays and neutrons, but we do not require explicit labeling of any individual pulse. Our “implicit labeling” approach relies on shielding materials with differential absorption of gamma rays and neutrons combined with a fairly simple experimental procedure. The classification algorithm combines statistical approaches to dimension reduction and density estimation with a physics-based attenuation model to produce highly accurate neutron labels from completely unlabeled starting data. The use of statistical methods eliminates the need for human intervention and increases efficiency by allowing identification of low energy neutrons that would previously have been discarded.