Invited Talk:    Creating a Reference Library of Uranium Electron Microscopy Images Suitable for Deep Learning and Image Segmentation

Luther McDonald, University of Utah, USA

Abstract.

Nuclear forensics analysis routinely measures isotope ratios, impurities, and chemical composition to identify unknown materials. While all of these methodologies are well-established, a new signature that is more rapid and able to identify the processing history of unknown nuclear materials would be beneficial. Particle morphology has shown this promise. Historically, the microstructural properties of nuclear materials have been used to verify nuclear fuel properties, and they can indicate the chemical and physical processes used to make that material. While this has been known for decades, the lack of reference libraries and a detailed understanding of what created the unique particle morphologies limited the utility of this signature in nuclear forensics. To overcome these limitations, we have performed numerous synthetic campaigns to synthesize and subsequently catalog the morphology of uranium oxides from uranium ore concentrates. Through collaboration with computer and data scientists, we learned many variables that impact the morphology, how the morphology is quantified, and the importance of capturing enough data for statistical relevance. This talk will focus on the benefit of establishing collaboration between experimentalists and computer and data scientists when developing methodologies to quantify particle features.