Past research efforts have shown that the factors affecting the surface morphology characteristics of uranium ore concentrates (UOCs) are numerous and complex. While driven largely by the route of precipitation and calcination product of the UOC, the solution precipitation conditions, the extent to which precipitates are washed, the presence of elemental impurities, and the conditions under which the materials have been aged will all impact the morphology of the UOC. The variety of factors that can impact the morphology results in large datasets often with overlapping variables making cause and effect correlations challenging to find. Nonetheless, convolutional neural networks (CNNs) have proven useful in classifying scanning electron microscopy (SEM) images of the UOCs. They have been used as a rapid means of processing large sets of images, classifying them into  fine-grained labels representing specific aging parameters, elemental impurities, or as pure, unaged “unperturbed” materials from different precipitation routes.<br/>This work ties together past UOC morphology studies to determine the ability of CNNs trained on image sets of unperturbed materials to generalize to impure, aged, partially converted, or otherwise altered UOCs that might have been imaged under varying SEM parameters and image scales. The results identify specific challenges that will need to be addressed as CNN techniques join the nuclear forensics analysis toolbox for characterizing the surface morphology of nuclear materials. Specifically, the CNNs performed well when classifying materials precipitated from different uranium solutions and images of unperturbed materials acquired at different scales, but failed to accurately predict aged, impure, and partially calcined materials. These results highlight the need for additional research into surface morphology as a nuclear forensics signature; principally, in developing SEM image databases that could be used to characterize the process history of nuclear materials.

Generalizability of convolutional neural networks to out-of-distribution uranium morphology image data

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.