Comparison Of Different Supervised Machine Learning Algorithms To Predict PWR Spent Fuel Parameters

Vaibhav Mishra - Uppsala Universitet
Erik Branger - Uppsala Universitet
Zsolt Elter - Uppsala Universitet
Sophie Grape - Uppsala Universitet
Peter Jansson - Uppsala Universitet
File Attachment

Nuclear safeguards verification of spent nuclear fuel (SNF) is imperative to ensure the peaceful use of nuclear material. To verify the correctness and completeness of operator declarations by conducting non-destructive assay (NDA) measurements of gamma and neutron radiation from the SNF plays a central role in safeguards. Verification of fuel based on such measurements is done routinely by safeguards inspectors as well as it is expected to be conducted prior to preparation of SNF for final disposal. Traditionally, SNF verification has been carried out by analyzing data from a single NDA instrument at a time whereas, in recent work, it has been shown that simultaneously analyzing multiple SNF signatures for the prediction of SNF parameters is not only feasible but also more successful.In this study, we investigate the performances of different machine learning algorithms in their ability to make meaningful predictions of SNF parameters such as fuel burnup (BU), initial enrichment (IE), and cooling time (CT). Predictions of this nature will be made based on measurables such as passive gamma-ray measurements of individual radionuclide activities. Other fuel signatures that will be used include derived parameters such as the Cherenkov light intensity (measured with the help of the Digital Cherenkov Viewing Device) and the parameterized differential die-away time (tau). To this end, we intend to train and evaluate multiple machine learning models on a set of simulated data containing 596,181 fuel samples (comprising of nuclide activities, tau and Cherenkov light intensity at various values of BU, CT, and IE) giving us a broad range of these three parameters to encompass the majority of the spent nuclear fuel in Sweden. It will also be of interest to evaluate the resilience of these machine learning algorithms on noisy data by the introduction of simulated systematic errors in the training features. It is foreseen that the results from the present study will also prove instrumental in deciding which algorithm performs best on experimentally measured data.