Speaker
Description
A significant fraction of the spent nuclear fuel used in nuclear power and research plants around the world is stored in dry storage casks. Since these casks are equipped with heavy shielding, monitoring of the condition of the spent nuclear fuel is often based on indirect methods, such as temperature or radiation checks. Due to their high penetration power and natural occurrence, the use of natural cosmic-ray muons presents a suitable way to directly image the fuel inside such storage casks. As a result, applying muon tomography to the monitoring of nuclear fuel casks has become an active field of research over the recent years. While these studies have shown promising results, the complex and heavily shielded cask geometry has only allowed up to now the reconstruction of only entire or partial fuel. The necessary goal of reconstructing single fuel rods has not been achieved so far.
This work presents the first clear demonstration of the feasibility for single fuel rod resolution within a nuclear storage cask. The study is based on a simplified, but realistic cask model of the CASTOR V/19 storage cask, the most common nuclear storage cask in Germany. The cask is imported to the GEANT4 simulation toolkit using the B2G4 workflow for three different scenarios: an empty cask without any fuel rod, a cask randomly filled with fuel rods at 90% capacity, and a cask filled with fuel rods at 100% capacity. The presented work utilizes a gradient descent-based implementation of the MLEM algorithm allowing a sparse and scalable, but high-resolution reconstruction of the different cask scenarios. In order to achieve single rod resolution, the resulting tomographic maps are denoised in the final step of the analysis by a residual U-Net deep learning architecture with an incorporated attention mechanism. When overlaid with a ground truth model denoting fuel rod positions a clear correspondence is shown, indicating that single fuel rods can be successfully identified.