Use of machine learning models for the detection of fuel pin replacement in spent fuel assemblies

Research output: Contribution to journalArticle

Institutes & Expert groups


The nuclear material contained in the spent fuel assemblies represents the majority of the material verified during the safeguards inspections, and the replacement of spent fuel pins from an assembly is one of the possible scenarios to divert nuclear material. Due to the high number of fuel pins contained in a fuel assembly (e.g. 264 pins in a PWR 17x17 geometry), a practically infinite number of diversion scenarios can be considered by a potential proliferator. In this framework, Monte Carlo simulations were used to model some of the possible diversion scenarios and to develop a database of detector responses corresponding to different nondestructive assay (NDA) techniques. In addition, the database contains the detector responses obtained with complete fuel assemblies with different initial enrichment, burnup, and cooling time. Given the large size of the database and the multiple detector responses resulting from the NDA techniques, the use of machine learning is proposed for the data analysis. In this work we focus on the classification problem with the aim of classifying the diversion scenarios based on the percentage of replaced pins. Several machine learning models were developed for this problem using decision trees, discriminant analysis, support vector machine, and nearest neighbors algorithms. The accuracy of the models was calculated as the number of correct classifications in the whole dataset. The results from the study show that the selection of the detector type used as input in the machine learning model has a strong impact on the accuracy of the developed model. In general the use of gamma-ray detectors leads to higher accuracies compared to the use of neutron detector responses. In addition, several machine learning models achieved a complete correct classification.


Original languageEnglish
Pages (from-to)22-35
Number of pages13
JournalEsarda Bulletin
Publication statusPublished - 1 Jun 2019


  • Machine learning, fuel diversion, Monte Carlo, spent fuel, non-destructive assays

ID: 5304234