Optimizing neural networks to detect replaced spent fuel pins using the partial defect tester

Plutonium in spent nuclear fuel represents the majority of the nuclear material placed under safeguards today. The large radiation emission and complex isotopic composition make spent fuel a challenging material to be verified. Non-Destructive Assays (NDA) are generally chosen for the safeguards verification of spent fuel using instruments such as the Fork Detector, the Digital Cherenkov Viewing Device and the Passive Gamma Emission Tomography. Research is on-going both on the development of additional NDA instruments and on the data analysis of measurement results in order to enhance the identification accuracy and reduce the inspection time. The Partial Defect Tester (PDET) is among the NDA instruments being developed specifically for the detection of replaced or missing fuel pins in a spent Pressurized Water Reactor (PWR) fuel assembly. The PDET instrument consists of a set of small neutron and gamma-ray detectors that are inserted simultaneously in the guide tubes of the assembly. A large set of Monte Carlo simulations including complete fuel assemblies as well as fuel assemblies with replaced pins has been developed in the recent years at the Belgian Nuclear Research Centre (SCK CEN). Given the challenges of spent fuel verification and the characteristics of the PDET measurement results, machine learning was chosen for the data analysis. In this contribution, an Artificial Neural Network (ANN) algorithm has been trained to recognize, via the PDET neutron detector responses for a spent fuel assembly, if and how many pins are replaced with respect to the declared configuration. The selection of hyper-parameters in the ANN, namely the number of neurons and hidden layers, is investigated to optimize the output accuracy. Results and an outlook on the future research are discussed.