Application Of Machine Learning Algorithms To Identification Of Spent Fuel Reactor Of Origin

Year
2020
Author(s)
Pavel Grechanuk - Oregon State University
Camille Palmer - Oregon State University
Todd S. Palmer - Oregon State University
Abstract

In this paper, we present the application of machine learning algorithms to the prediction of the reactor of origin of a sample of spent nuclear fuel, based on training data from The Spent Fuel Isotopic Composition (SFCOMPO) database. This database catalogues experimental data from 44 different reactors containing over 750 fuel samples, with more than 24,000 measurement entries . Our first objective was to predict the reactor of origin of a nuclear fuel sample based on a set of measured isotopic concentrations. The second object was to predict the isotopic concentrations of a sample of spent fuel from a specified set of data on reactor of origin, burnup, and initial enrichment. We employed ensembles of decision trees called Random Forests for their robust performance and ease of interpretation. To accomplish the first objective, we used the major actinide concentrations, ratios of the concentrations in the fuel, and the burnup as features for machine learning. For each of the fuel samples, the task was to predict which of the 12 possible reactors the sample came from. A Random Forest classifier was trained and a accuracy of 0.8465% was obtained. These are strong results for a first pass investigation considering we did not use any of the minor actinides or other fission products. We were less successful in meeting our second objective. Again, a Random Forest regression model was employed, and we measured success using the root mean squared log error (RMSLE) because the magnitudes of the concentrations varied significantly. We find that the error is very low for most isotopes, but quite large for U-238. We have demonstrated that machine learning algorithms can be applied to the SFCOMPO database to obtain novel insights, and that we can reliably predict the reactor of origin for an unknown nuclear fuel sample. We have also shown that the spent fuel isotopic compositions can be predicted when using the burnup, reactor design, and initial fuel concentrations as features.