Identifying Nuclear Proliferation Using Inter-facility Material Transportation And Machine Learning Classification

In this paper we discuss a proof-of-concept statistical analysis of a nuclear fuel diversion scenario using only data representing material transportation between facilities. We represent these material transactions between facilities in units of shipping containers, which would be identifiable in the real world via satellite imagery. This approach is potentially advantageous because acquiring satellite data is minimally invasive and available even when IAEA inspectors are denied access to a site. We present a computational model that represents the nuclear fuel cycle for a clandestine nuclear pursuer based on a fuel diversion scenario. This model is built using Cyclus, an agent-based fuel cycle software tool that simulates material transactions based on facility requests. Previous work in Cyclus includes a study of complex fuel diversion pathways involving production of plutonium and highly enriched uranium and a study of machine learning classification for a simple diversion scheme. We expand on this previous work by introducing and limiting analysis to transportation between facilities. Using transportation data, we analyze the performance of various statistical methods in classifying the presence of nuclear material diversion, including simple and more advanced machine learning techniques. In our analysis, we address the following questions: Is it possible to reliably classify diversion using only shipping container unit data? What diversion scenarios are easier to hide relative to others based on this metric? In answering these questions, this study demonstrates the power of statistical analysis for nuclear nonproliferation problems.