Comparison of machine learning models for the detection of partial defects in spent nuclear fuel

Abstract

Within the framework of safeguards verifications spent nuclear fuel is a concern because it contains nuclear material. Non-destructive assays (NDA) are amongst the safeguards measures for spent fuel verification. In this work machine learning using simulated data is investigated for the detection of fuel pin diversion. Three NDA techniques (Fork, SINRD, and PDET) and two machine learning approaches (decision trees and k-nearest neighbors) are considered to classify the assemblies according to the percentage of replaced pins. These NDA techniques combine different types of neutron and gamma-ray detectors. This study found that the classification accuracies using SINRD and PDET are higher compared to Fork. In addition, k-nearest neighbors models reached higher classification accuracies compared to decision tree models, and for the considered NDA techniques the gamma-ray detectors were the most sensitive to the fuel pin diversion.