Adapting a machine learning method for the source discrimination of plutonium samples mixed from multiple reactor types

Abstract

Plutonium (Pu) samples that are sourced from multiple, different reactor types present a challenge to nuclear forensics investigations. Previous studies have developed a nuclear forensics methodology capable of identifying a Pu sample’s reactor type of origin using isotope ratios as features in machine learning classification models. However, the models could only attribute Pu sourced from a single reactor type. The methodology was adapted to discriminate between Pu produced from the six original single reactor type classes and twelve new classes comprised of binary mixtures of the six original reactor type classes. This adaptation was a success, with a support vector machine (SVM) identified as the most suitable model type for the task. The model’s sensitivity to different groups of features was examined and the model was also validated with data from experimentally produced Pu samples, in both single reactor type and mixed reactor type cases.