Artificial neural networks for the evaluation of uranium hexafluoride cylinders

The United States Department of Energy (DOE) currently manages approximately 50,000 carbon steel cylinders containing more than 1 billion pounds of depleted uranium hexafluoride (UF₆), a byproduct of uranium enrichment activities. These cylinders are currently located at outdoor storage yards, exposed to the atmosphere. Knowledge about these cylinders’ minimum wall thickness will help decision-makers in deciding the most feasible method to manage and/or dispose of the cylinders. In this project, neural networks were trained to predict the cylinders’ minimum wall thickness and to categorize them according to their wall thickness. In addition, a genetic algorithm/neural network (GA/NN) model was trained to do the same task. The genetic algorithm was used to search for combinations of variables that would provide best neural network training results. Prediction results obtained from neural network training and the GA/NN model were compared to those from multiple linear regression. The data set used in the network training was from the UF₆ Cylinder Location, Inspection, and Measurement System (UCLIM) database. The neural network and GA/NN training results showed that the UCLIM database information was not sufficient to predict the cylinder wall thickness or categorize the cylinders according to their minimum wall thickness.