“Our multi-year plan will close existing proliferation detection capability gaps and address the principal components of proliferation analytics.”
– Eddy Banks
LLNL’s national security mission includes developing scientific and technological solutions to address the evolving landscape of nuclear proliferation threats. This means monitoring and detecting weapons of mass destruction as well as preventing the spread and availability of related materials and infrastructure.
The multi-institutional Advanced Data Analytics for Proliferation Detection (ADAPD, pronounced “adapt”) project aims to make a tangible difference in this crucial mission space through early detection of low-profile proliferation activity that may be small, inaccessible, or buried in background activities. The NNSA’s Office of Defense Nuclear Nonproliferation funds ADAPD’s research and development efforts, which draw upon the fruits of another LLNL project. Launched in 2018 and led by LLNL, ADAPD brings together four other DOE laboratories: Los Alamos, Sandia, Oak Ridge, and Pacific Northwest.
Proliferation detection relies primarily on the interpretation of quantitative, physics-based observables and the single-modality analysis of these data streams. Although these direct observables can provide high-confidence indicators to characterize known proliferation activities, they cannot support the collection, analysis, and interpretation of observables at the requisite volume, scope of coverage, or time scales needed for early, low-profile detection.
Computing’s Jim Brase, who serves as ADAPD’s venture manager, explains, “We need to look at all potential observables, then put them together in one detection scheme. To do so requires innovative analysis of multimodal data’s cumulative effects. Data science will ultimately enable early detection of low-profile proliferation via techniques more powerful than currently available methods.”
Therefore, ADAPD combines direct and indirect observables—including the technical, business, and human processes that enable and support proliferation activities, such as uranium isotope separation—to deliver a global-scale, real-time capability to detect, locate, and characterize low-profile proliferation.
According to the project’s principal investigator, Eddy Banks, ADAPD’s roadmap centers on state-of-the-art predictive modeling and analytics techniques. He says, “Our multi-year plan will close existing proliferation detection capability gaps and address the principal components of proliferation analytics through three main research and development areas.”
These areas are (1) predictive models for multiple proliferation observables that integrate new types of data-driven models with traditional physics models and the knowledge of subject matter experts; (2) multi-phenomenology detection to infer the state of hidden proliferation processes; and (3) transfer of models and detection capabilities to new environments.
To implement this strategy, ADAPD relies on technical expertise in physics-informed machine learning, multimodal data integration, and nuclear weapons development processes. “New developments in data science, machine learning, and computation provide approaches and tools that are new to this domain,” states Banks. Brase adds, “We will also bring LLNL’s high-performance computing systems to bear on the large-scale machine learning and simulation tasks needed to solve this problem.”
Besides subject-matter expertise and computer power, ADAPD requires data. The amount of accessible data relevant to detecting and characterizing nuclear proliferation activities around the world has grown exponentially over the past decade. Banks explains, “No analytics program can make progress without access to relevant data with well-understood backgrounds and ground truth. Existing proliferation testbeds and experimental activities provide this foundation.”
ADAPD researchers will partner with programs across DOE, NNSA, and DNN to compile data sets that capture the complexity and scale of proliferation activities. The availability of multiple related testbeds provides an opportunity to validate models against different and unknown environments as well as to test validation methods designed to operate in the absence of ground truth.
The project team is Eddy Banks, Barry Chen, Stefan Hau-Riege, Goran Konjevod, Rafael Rivera Soto, Ted Scharlemann, Brian Van Essen, David Widemann, Christian Martinez-Nieves, Ted Stirm, Maggie Arno, Jamie Van Randwyk, Yana Feldman, Steven Samson, Gene Ichinose, and Carmen Carrano.