Application of machine learning algorithms to the safeguards verification of spent nuclear fuel

Several types of machine learning algorithms have been analyzed in this thesis in the framework of safeguards verifications, with the purpose of detecting the diversion of fuel pins from spent fuel assemblies. The goal of the thesis is to classify spent fuel assemblies according to the percentage of diverted pins. The term spent fuel corresponds to the assemblies that after a long period of irradiation are no longer able to sustain the chain reaction and for this reason are unloaded from the reactor core and transferred to the spent fuel pool. Spent fuel is a large radiation source, both in terms of neutrons and gamma-rays, as well as residual heat due to the presence of fission products and minor actinides. The presence of fissile material (235U, 239Pu) in amounts around 2% make spent fuel a very important material for safeguards, in order to verify that this material is used only for peaceful purpose. The detector responses obtained from Monte Carlo simulations for three type of detectors, namely the Fork detector, Self-interrogation neutron resonance densitometry (SINRD) and Partial defect tester (PDET), were organized in databases to be used in this thesis. The detector responses are used as input features inside 3 different type of machine learning algorithms to generate models that are able to classify the fuel assembly according to the percentage of diverted fuel pins. A machine learning algorithm is an algorithm that receives as input a database containing both input predictors and the associated output responses; in this case study the detectors responses represented the input predictors, and the class of the assembly based on the percentage of the missing pins indicated the output response. Once a machine learning model has been trained, it is able to predict the class of a new observation starting from the values of the detector responses. The machine learning models considered in this study were discriminant analysis, the decision tree, and the K-nearest neighbors (KNN). The results showed that the gamma-ray detector response is generally the most sensitive to the diversion of fuel pins, and that higher classification accuracy was obtained for PDET and SINRD compared to the Fork detector. Comparing the different machine learning models, the KNN approach usually led to larger classification accuracies, followed by decision trees and by discriminant analysis.