Math Modeling and Classification Techniques for non proliferation problems

The International Atomic Energy Agency (IAEA) was established in 1967 with a mission to develop and deploy safeguards technologies and techniques that are required to independently verify the compliance of Non-Proliferation Treaty signatories (Cohen 1993). With this in mind our focus is on developing and evaluating algorithms for nuclear non-proliferation assessments - in particular, algorithms to detect undeclared activities at nuclear facilities using both 1) computational models of the nuclear fuel cycle, and 2) sensor observations of multiple characteristics from multiple sensors. We used a nuclear simulation software called Cyclus to produce test data for several scenarios that would answer questions like: Can disparate observations from different sensors be combined to create data products with superior information content? and Is a facility making material of interest? If so, how much? With this work we were able to successfully answer the above questions. Our team formulated a forward model that provided a high-level representation of facilities with inputs and outputs (e.g. resources, facility on/off, waste material, production), along with observations of those facilities. The facilities were constructed in a way that distinct combinations of the facilities could be combined to represent different scenarios. For instance - diversion of materials from one of the facilities was readily demonstrated in this setting. This forward model fed into a machine learning classification algorithm. We utilized both random forest and convolutional neural networks. These methods achieved a level of success in addressing the above science question. With the Convolutional Neural Networks algorithms, we obtain above 90% accuracy on our classification with the machine learning methods on the simulated data.