Developments toward a novel methodology for spent nuclear fuel verification

Licentiate thesis, 2022

One of the tasks in nuclear safeguards is to regularly inspect spent nuclear fuel discharged from nuclear power reactors and verify the integrity of it, so that illegal removal and diversion of radioactive material can be promptly discovered. In the current project, which is a collaboration between Chalmers University of Technology and SCK CEN, a novel methodology for non-intrusive inspection of spent nuclear fuel is under development. The methodology consists of two main steps: 1) neutron flux and its gradient are measured inside spent nuclear fuel assemblies using small neutron detectors; and 2) the measurements are processed using an Artificial Neural Network (ANN) algorithm to identify the number and location of possible fuel pins that have been removed from the fuel assemblies and/or replaced with dummies. The use of small neutron detectors simplifies the inspection procedure since the fuel assemblies are not moved from their storage position. In addition, the neutron flux gradient measurements and its processing with the ANN algorithm have the potential for more detailed results. Different aspects have been investigated for the development of the methodology.
For the first step of the methodology, the concept of a new neutron detector has been studied via Monte Carlo simulations and it relies on the use of optical fiber-mounted neutron scintillators. The outcome of the computational study shows that the selected detector design is a viable option since it has a suitable size to be introduced inside a fuel assembly and can measure neutron flux gradients. Then, experimental work has been carried out to test and characterize two optical fiber-based neutron scintillators that can be used to build the detector, with respect to detection of thermal neutrons and sensitivity to gamma radiation.
For the second step of the methodology, a machine learning algorithm based on ANN is studied. At this initial stage, a simpler problem has been considered, i.e., an ANN has been prepared, trained and tested using a dataset of synthetic neutron flux measurements for the classification of PWR nuclear fuel assemblies according to the total amount of missing fuel, without including neutron flux gradient measurements and without localizing the anomalies. From the comparison with other machine learning methods such as decision trees and k-nearest neighbors, the ANN shows promising performance.