Title: SURFACE MORPHOLOGY ANALYSIS OF URANIUM OXIDES FOR IMPROVING NUCLEAR FORENSICS IMAGE CLASSIFICATION MODELS
Committee:
Luther McDonald
Elizabeth Jurrus
Tara Mastren
Glenn Sjoden
Tolga Tasdizen
Abstract:
Surface morphology analysis has emerged as a valuable tool in identifying the origins of interdicted illicit nuclear or radioactive materials in nuclear forensics investigations. Surface morphology signatures of uranium oxides have been connected to specific parameters relating to a specimen’s synthesis or precipitation, thermal processing, elemental impurities, and aging in various atmospheric conditions; when paired with more traditional nuclear forensics techniques a more complete picture into the unknown specimen’s process history can be seen. In addition to qualitatively describing morphology features and quantitatively measuring particles by segmentation of images acquired by scanning electron microscopy (SEM), classification of micrographs by convolutional neural networks (CNN) might allow for faster characterization of surface morphology while eliminating human biases. However, previously published CNN implementations for uranium oxide morphology have considered only test datasets split from the training set images, leaving questions into the models’ abilities to generalize to different distributions covering a wider gamut of what one might expect to see for uranium oxide SEM data.
The work presented is four-fold, starting with an overview of computational image analysis techniques, and how they might be used for actinide chemistry and nuclear forensics analysis. Ore concentrates (UOC) are then synthesized by sulfuric acid leaching U-rich ores, purification by two different commercial solvent extraction (SX) processes, precipitation, then calcination to a uranium oxide. The UOCs produced by either the Amex or Dapex SX routes were similar but statistically unique and identifiable be a trained CNN classifier; impurity concentrations for each were within the range one would expect for commercial mining and milling processes. Next, uranium dioxide (UO2) was synthesized from uranyl fluoride (UO2F2) solutions by two different synthesis routes used in the deconversion processes after the enrichment steps of the nuclear fuel cycle. The materials from each route were converted to UO2 under two different atmospheric conditions and the surface morphology and elemental fluorine impurities were characterized for each.
Finally, uncertainty quantification (UQ) by variational inference methods and the ability of models to generalize to out-of-distribution (OOD) image data were evaluated. Fully supervised CNN classifiers and unsupervised vector quantized variational autoencoders (VQ-VAE) paired with supervised classifiers were trained on a set of images taken of pure, unaged uranium oxides from various synthesis routes that were acquired under the same SEM operating parameters. Results showed which changes to uranium process history or data collection parameters lead to poor predictions by the classifiers. These findings will be used to further develop the uranium oxide database for nuclear forensics and how microscopy data is applied for machine learning applications.
