Nonproliferation Data Scientist Oak Ridge National Lab Knoxville, Tennessee, United States
The Analytical Discovery Venture (ADVenture) is addressing an emerging need for novel means of effective, trustworthy, and deployable discovery of low-profile signatures of new entities and activities in the nuclear threat landscape. The ADVenture workflow will rely on multiple nascent and still evolving research capabilities – identification of high-value data in a boundless corpus, methods for structuring data to effectively extract knowledge, and AI-informed reasoning that leverages that structured knowledge; this workflow is described in companion presentations at this meeting. One key enabling capability within ADVenture is encoding data to maximally extract information from sparse observations. To that end, we will present a conceptual workflow for physics-constrained data semantics and embeddings through fuel cycle modeling. Traditionally, fuel cycle modeling tools have been used to track the temporal evolution of the material and technology balance required to produce nuclear energy. We extend these tools by considering uncertainty in input data and real-world observations, as well as wholly incomplete data. We view these uncertainties and incomplete data as giving rise to three different analysis regimes that address separate questions: (1) What is the distribution of observables we expect to see?; (2) What is the distribution of an unknown parameter(s) given a set of observations?; (3) What are plausible fuel cycle configurations, potentially with diversion or acquisition pathways, that could give rise to a set of observations?
