Neutrino Detectors and Neural Networks for Nuclear Nonproliferation

The CHANDLER detector is planned to measure the neutrino spectrum of nuclear reactors for nonproliferation monitering. The use of neural networks to classify events in this detector by particle type and evaluate the strength of variables as classifiers is explored. Two types of perceptron network are trained on a computer-simulated monte carlo dataset that identifies the event as a neutron or a neutrino, then gives a list of data values for the event. The first network– which optimized a single hyperplane cut– achieved a significance of 100, and the second network– a more sophisticated model with four ReLu processing neurons– achieved a significance of 146, outperforming the decision tree used previously for the CHANDLER detector. Subsequently, separate iterations of these networks were trained on 1-gamma and 2-gamma events to allow them to isolate features individual to these classes of events. This procedure found that 1-gamma events were classified very poorly by both networks, so a new variable that measures the escape probability of the second gamma was introduced, achieving a small improvement at selecting for 1-gamma IBDs. An alternative method of separating 1-gamma IBDs from 2-gamma IBDs was also found through the analysis of 2-d histograms, which improved the classification rate further. Finally, a new reward function that optimized for significance directly was introduced to train the neural network, greatly reducing the amount of manual tuning and re-learning needed to train an effective network. The combination of all of these achieved a significance of 170, outperforming all prior classification methods.