Using machine learning for the detection of missing fuel pins in spent nuclear
fuel assemblies based on measurements of the gradient of the neutron flux

One of the main tasks in nuclear safeguards is the inspection of spent nuclear fuel (SNF) assemblies to detect possible diversions of special nuclear material such as 235U and 239Pu. In the inspection, measurements of relevant observable quantities are acquired from the assembly, e.g., neutrons emitted by the spent fuel, and used to verify whether they are consistent with the declared configuration of the assembly or not. The procedure requires a physical model that can estimate the response of the detectors for a given arrangement of fuel pins in the assembly, and an unfolding technique, based on the physical model, that can be applied to retrieve, from the detector responses, the parameters of the system configuration. In this work, the use of neutron flux gradient measurements for the identification and characterisation of diversions in a SNF assembly is investigated. The unfolding procedure relies on an artificial neural network (ANN), which has the advantage of generalizing in an efficient manner the mapping of the input (in this case, the measurements from the SNF assembly) to the output (i.e., the fuel pins that are intact or replaced with dummy pins in the assembly). The training and testing of the ANN makes use of a dataset generated using Monte Carlo simulations of a typical 17x17 PWR assembly with different patterns of missing fuel pins. The dataset is built of unique scenarios so that the ANN can be tested and assessed over scenarios that are not part of the learning phase. The study shows that information related to the neutron flux gradient can lead the ANN to be more accurate in identifying the replaced fuel pins. Although the developed ANN models cannot fully reconstruct any of the diversion patterns included in the dataset, they provide results close to the real assembly configurations in most cases.