Use of artificial neural networks in measuring characteristics of shielded plutonium for arms control

Abstract

Artificial neutral networks were developed for use as a potential ‘information barrier’ technology in the verification of arms control treaty accountable items. They were used to identify and measure specific attributes from γ-ray spectra. These attributes included the presence or absence of plutonium, the plutonium Pu-239/Pu-240 isotopic ratio or ²³⁹Pu content and the material age. A set of over 400 training spectra were generated using a spectral simulation software package and various methods for the selection of input data were tested. An input data set which discounted low energy regions susceptible to shielding effects was found to be most effective. Once trained, the network correctly identified the presence or absence of plutonium from real γ-ray spectra. Accurate results were also achieved for estimating the content of ²³⁹Pu. In simulated test spectra a root mean squared error (RMSE) of less than 0.1 was found when using the optimum number of inputs. The network was also able to distinguish between spectra from plutonium samples of different ages. Further work is planned to investigate the estimation of a confidence level for whether a specific threshold of ²³⁹Pu content is exceeded. An improved training set is anticipated to improve accuracy in determining the material age, which was not achieved accurately.