During nuclear search operations, the localization of radioactive sources can be a time-consuming process that requires mapping relative radiation intensity in a large area to determine the position of a source. This article introduces the use of machine learning, specifically a temporal convolutional network (TCN), to estimate the direction between a detector array and a static 137 Cs source. This application of machine learning provides a directional vector in $4\pi $ with a 90% confidence of 5.6° and a 99% confidence within 11.2°. With the use of low-cost NaI(Tl) detectors, the effects of self-shielding within the array creates gamma-ray shadows depending on the orientation to the source. Using the convolved detector array response function, we apply supervised machine learning with a neural network to predict a unit vector that points toward the observed source. The directional vector is expected to reduce search times once implemented in future work.

Historically, nuclear science and radiation detection fields of research used Pulse Shape Discrimination (PSD) to label gamma-ray and neutron interactions. However, PSD’s effectiveness relies greatly on the existence of distinguishable differences in an interaction’s measured pulse shape. In the fields of machine learning and data analytics, clustering algorithms provide ways to group samples with similar features without the need for labels. Clustering gamma-ray and neutron interactions may mitigate PSD’s pitfalls, since clustering methods view the total waveform rather than just the area under the tail and the total area under the pulse. However, traditional clustering methods, such as the k-means clustering algorithm, suffer from poor performance on high dimensional data. This study explores unsupervised machine learning methods using Deep Neural Networks (DNN) to cluster gamma-ray and neutron interaction measurements collected with an organic scintillation detector, in order to perform binary labeling of gamma-rays and neutrons. Using various network architectures, this research demonstrates the effectiveness of using autoencoder-based neural networks to cluster gamma-ray and neutron interactions when compared to shallow clustering algorithms. The results reveal the effectiveness of autoencoders on high energy gamma-ray and neutron pulses with an energy deposit greater than 0.80 MeVee whilst greatly outperforming k-means comparatively in all cases.

In the nuclear sciences and radiation detection fields, the differentiation between gamma-ray and neutron interactions inside a detector volume continues to be an area of active research. Historically, the primary mechanism for conducting particle identification has been pulse shape discrimination (PSD). However, almost all variations of this technique rely on only two factors: the area of the tail and the total area of the pulse. In the last decade, the emergence of advanced machine learning techniques, most specifically artificial neural networks (ANNs), offers a unique opportunity to capitalize on the entirety of the waveform. But such techniques appear highly reliant on the quality of datasets used for training. Our research addresses this challenge to quantify the relative performances of networks trained on a variety of datasets and subjected to the same test. Furthermore, we offer an analysis of the portability of a network trained on one detector to a similar detector.

Radiation imaging is one potential method to verify nuclear weapons dismantlement. We present a method to discriminate between non-multiplying and multiplying neutron sources using a neutron coded aperture imaging system. This method applies time-correlated pulse-height (TCPH) analysis to identify neutrons in fission chain reactions and recreate images using only these events. This analysis was applied to measurements of weapons-grade plutonium (WGPu) metal and 252Cf performed at the Nevada National Security Site (NNSS) Device Assembly Facility in July 2015. The results demonstrate it is possible to eliminate the non-fissile 252Cf source from the image while preserving the fissile WGPu source. TCPH analysis was also applied to additional scenes in which the WGPu and 252Cf sources were measured individually. The results of these separate measurements further demonstrate the ability to remove the non-fissile 252Cf source and retain the fissile WGPu source. Simulations performed using MCNPX-PoliMi indicate that in a one hour measurement, hollow WGPu spheres are retained at a 1σ level for neutron multiplications M≃2.7 and above. This work has potential application to dismantlement verification where it may be valuable to verify that all neutron sources in an image are multiplying.