Uncertainty Quantification in Emission Quantitative Imaging

holdup measurement;quantitative imaging;uncertainty quantification;Nuclear Engineering and Radiological Sciences;Engineering;Nuclear Engineering & Radiological Sciences

Bevill, AaronKiedrowski, Brian ;

University of Michigan

关键词: holdup measurement;    quantitative imaging;    uncertainty quantification;    Nuclear Engineering and Radiological Sciences;    Engineering;    Nuclear Engineering & Radiological Sciences;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/136929/ambevill_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Imaging detectors have potential to improve the reliability of plutonium holdup measurements. Holdup measurement is a significant challenge for nuclear safeguards and criticality safety. To infer holdup mass today, inspectors must combine data from counting (non-imaging) detectors with spatial measurements, process knowledge, and survey estimates. This process results in limited certainty about the holdup mass. Imaging detectors provide more information about the spatial distribution of the source, increasing certainty.In this dissertation we focus on the emission quantitative imaging problem using a fast-neutron coded aperture detector. We seek a reliable way to infer the total intensity of a neutron source with an unknown spatial distribution. The source intensity can be combined with other measurements to infer the holdup mass.To do this we first create and validate a model of the imager. This model solves the forward problem of estimating data given a known source distribution. We use cross-validation to show that the model reliably predicts new measurements (with predictable residuals).We then demonstrate a non-Bayesian approach to process new imager data. The approach solves the inverse problem of inferring source intensity, given various sources of information (imager data, physical constraints) and uncertainty (measurement noise, modeling error, absence of information, etc). Bayesian approaches are also considered, but preliminary findings indicate the need for advanced Markov chain algorithms beyond the scope of this dissertation. The non-Bayesian results reliably provide confidence intervals for medium-scale problems, as demonstrated using a blind-inspector measurement. However, the confidence interval is quite large, due chiefly to modeling error.