Transformative Data Analytics Capabilities for Nuclear Forensics and Safeguards

Novel data analytics methods and associated applications are currently being developed at Oak Ridge National Laboratory. The rigorous statistical formulation and introspective features of these methods enable transformative innovation in nuclear forensic and safeguards applications by facilitating analyst interpretation of the implicitly identified signatures. As an example, we present methods to infer the average properties of a nuclear reactor core using environmental samples. Inferring the properties of the overall core using a sample drawn at random is problematic: Analyzing multiple specimens requires knowledge of the sampling distribution, and analyzing a single specimen requires estimation of a potentially high-dimensional, nonlinear relation between the composition of multiple nuclides and the properties of interest. Using a novel multivariate regression algorithm and data from high-fidelity 3-D depletion calculations, we have determined the core-averaged burnup of a reactor consistent with plutonium production to within 4.5%. This model analyzed 90 nuclides and identified 10 nuclides as important for determining burnup independent of position. After model calibration, only these nuclides were used to make burnup determinations. We have also used this method to analyze data with fewer nuclides to study limited availability of nuclide assay data. After demonstrating our method’s strong dependence on the temperature of the reactor and the degradation of results when temperature is unknown and allowed to vary, we trained and tested models using data associated with multiple core-average temperatures and showed our approach’s ability to distinguish changes in nuclide composition associated with temperature from changes associated with burnup. Inspired by these results, we developed and demonstrate an approach to train models on distributions of data to produce models that are robust to measurement noise. Although most applications we have considered to date have been in nuclear forensics, our analytics capabilities will also advance the state of the art for safeguards applications. Our rigorous approach yields scientifically defensible results and valuable information to inspectors and other parties upstream of data analysts. Information produced using this approach will support prioritization of valuable data streams and indicate problematic data collection that should be repeated by on-site inspectors.