Optimization of machine learning models for the safeguards verification of spent fuel assemblies

Among safeguards verification activities, spent nuclear fuel is placed under inspections for the detection of possible diversions of the fissile material present, being of particular concern from the non-proliferation point of view. Several Non-Destructive Assay are utilized, or under investigation, for the verification of spent nuclear fuel. Among the latter, the Partial Defect Tester (PDET) is the one considered in this work. In this framework, some machine learning methods have been implemented in this thesis project, in order to detect the replacement of fuel pins in spent nuclear fuel assemblies. The input features of the models were based on different combinations of types and locations of detector responses, whose values were provided by a dataset of Monte Carlo simulations, based on the PDET prototype. The machine learning methods used were supervised regression models, namely k-nearest neighbors and neural network algorithms. The expected outcomes were the predictions of the number of replaced pins and their locations on the grid lattice. Within the different implemented configurations of the models, the results showed better performances when the responses of the gamma-ray detectors, for all the locations on the assembly, were involved in the input phase, compared to other models using other detector types. Comparing the different machine learning algorithms, larger accuracies were generally obtained with neural networks compared to k-nearest neighbors method, reaching 97% of correct predictions for the location of the replacements.