Machine Learning Approaches For Nuclear Material Accounting Data From Irradiation And Reprocessing

Year
2020
Author(s)
Mark B. Adams - Oak Ridge National Laboratory
Adam Drescher - Oak Ridge National Laboratory
Scott L. Stewart - Oak Ridge National Laboratory
Ken J. Dayman - Oak Ridge National Laboratory
Louise G. Worrall - Oak Ridge National Laboratory
Abstract

Data analysis methods are currently being explored for their ability to strengthen the synthesis and evaluation of information generated within domestic and international safeguards regimes. Safeguards data typically includes rich sets of heterogenous data amenable to advanced data analytics methods such as machine learning. We have converted transactional data from a domestic nuclear material control and accountability system into a directed graph, where nodes are inventory items and edges are transactions. One advantage of a graph representation for this type of data is the potential to disambiguate an underlying nuclear process (in our case irradiation and reprocessing) from nuclear material control and accountability system transactional data. This allows for accountable items to be followed throughout their lifetime in a nuclear process, which facilitates outlier detection, classification, and process monitoring. This work will outline how we have reduced the dimensionality of the data using both global and local techniques to study local and global structure and identify salient features of the data. After transforming the flat transactional data to a structured graph representation, we used tools such as graph neural networks to produce graph embeddings usable for graph classification, node classification, clustering, or link prediction. Moreover, we have adapted statistical hypothesis testing methods originally developed with conventional principal component analysis to the graph data to identify data that is dissimilar from normal process data, which can be useful for outlier detection.