A Machine Learning Method for the Forensics Attribution of Separated Plutonium

Abstract

A nuclear forensics technique, based on the maximum likelihood method, for the attribution of reactor type, fuel burnup, and time since irradiation (TSI) of separated pure plutonium (Pu) samples was previously developed at Texas A&M University. The method utilized measured values of ten intra-elemental isotope ratios in the Pu sample and a large database consisting of the values for these ratios as a function of the three attributes: reactor type, fuel burnup, and TSI. However, this method failed for Pu samples with mixed attributes. Hence, a new technique based on machine learning methods was developed that matched the capabilities of the previous maximum likelihood method for pure Pu samples. This new methodology used support vector machines for reactor-type discrimination and Gaussian process regression for fuel burnup quantification. The TSI was calculated analytically using the predicted reactor type and fuel burnup. This new method holds great potential for the attribution of mixed Pu samples.