The Complexity Science for Nuclear Security team is currently using multi-source data analytics for nuclear nonproliferation applications. In the absence of direct detection of highly specific nuclear materials, indirect evidence must be used to understand the activities at both declared and undeclared nuclear facilities. Distributed wireless networked sensors provide important information to inform such assessments.
Pictured is one of our sensors, colloquially referred to as “canaries”
Our goal is to collect multimodal data and develop algorithms for integration of disparate data types to identify “useful” patterns related to the nuclear fuel cycle. We are currently utilizing recently popularized machine learning tools to classify the operational status of nuclear facilities as well as to identify patterns related and relevant to that operational status. Such patterns include indirect life cycles such as traffic patterns, movement of large vehicles, and human presence in and around facilities of interest. We are currently testing various mathematically motivated techniques, such as neural networks, for our classification efforts.
Here we have a live visualization of the data collected by our devices, to be used for multi-source data analytics and algorithm development.
Another one of our goals is to create models that are transferable between nuclear facilities with different purposes. While have been collecting data from more than one type of facility, we will continue to update the sensor hardware to include new modalities and expand our data collection campaigns to various types of facilities.
In the past 30 months, our team has:
• Developed a software framework capable of representing millions of nodes and tens of billions of links, with clamping, sub-network inclusion or exclusion, and dynamic specification of mathematical characteristics within and between networks.
• Informational Sensing for Nonproliferationet alQuantifying Correlations between International Relations and Nuclear Proliferation Status et al
• Begun construction of the Correlates of Proliferation database, modeled on the widely-used Correlates of War site maintained at UC Davis; the database will provide well-organized historical data about conflict, nuclear cooperation agreements, trade, and other relevant data that could influence the decision to build nuclear weapons.
• Presented on our algorithm development efforts at various conferences, including the INMM Central Region Chapter Workshop (October 2018) and the INMM Novel Technologies, Techniques, and Methods for Safeguards and Arms Control Verification Conference (August 2017)
• Created and automated a test that distinguishes the optimal model and set of hyperparameters for each dataset that we have, getting us a step closer to creating models that are transferrable between datasets from various types of nuclear facilities. Models trained and tested on our current dataset all have an accuracy, precision, and recall of over 99%.