Machine Learning Approaches To Determine Missing Material From Nuclear Fuel Assemblies

Confirming the presence of radioactive material within a sample is advantageous for various applications in nuclear safeguards and material accountancy. For example, non-destructive assay (NDA) may be performed on a spent fuel assembly to verify the declaration of nuclear material, and to determine material unaccounted for (MUF). Gamma ray spectroscopy is a common NDA technique, and utilizing multiple gamma ray detectors, one can attempt to verify that each component within an assembly is present. This work makes use of three machine learning algorithms (the k-nearest neighbors, random forest, and support vector machine) to investigate their utility in processing gamma ray spectroscopy data for the purpose of determining if there is MUF, and identifying the location within the array from which the material is missing. To test these algorithms, simple models of spent fuel assemblies are simulated in Monte Carlo N-Particle (MCNP) surrounded by multiple NaI detectors. Fuel rods are modeled as solid U₃O₈ to account for self-shielding, and are simulated to emit various gamma ray energies characteristic of spent nuclear fuel. Testing and training databases are assembled by running the MCNP models for various combinations of individual rods being present or missing. Algorithms trained off of simulated results have been validated against bench top laboratory measurements. Algorithms are assessed in terms of their accuracy in identifying MUF, predicting the number of missing rods, and in predicting the exact locations from within the assembly that fuel rods are missing. Additionally, the performance of the algorithms is evaluated as a function of fuel assembly size, percentage of missing material, and relative variance in gamma ray emissions from individual fuel rods. Preliminary results with a 3x3 fuel assembly show that the machine learning algorithms can predict the number of missing fuel rods with near 100% accuracy with a relative variance in fuel rod emissions of 10%, and can predict the exact location of those rods with 98% accuracy. This work will focus on determining the limits of these machine learning methods, and investigate the process of converting raw MCNP outputs into useful input features for the algorithms to learn with.