Machine-learning methods for blind characterisation of nuclear fuel assemblies

The global prevalence of uranium as fissile fuel in nuclear reactors, paired with its transmutation to plutonium inside the core in an isotopic ratio corresponding to a direct-use material insignificant quantities, makes the handling of spent nuclear fuel an important and sensitive matter towards non-proliferation efforts. A fast and reliable method for characterising spent fuel is thus desirable for spent nuclear fuel reprocessing and storage facilities. We propose a non-destructive, blind and fast measure method of the quantity of spent nuclear fuel inside an assembly. By measuring photon fluency collectively for the complete assembly we determine the number of present fuel rods without the need to open the array and manually check. For this, we circle a detector set-up around the assembly and feed its measurements into a neural network fora prediction. Different specifically designed architectures based on dense and convolutional layers are trained on synthetically generated data using self-developed code on python. We arrive at the election of a convolutional network for optimal results. We achieve an exact prediction with over three sigmas of confidence (99.85% accuracy) thanks to the double detector set-up we introduce in this article, proving the prediction power of neural networks in this instance with a relatively simple measure configuration.