EXPERT: AI-driven Analytics for Nonproliferation Monitoring

Usable and explainable models for global-scale, cross-lingual proliferation expertise identification and forecasting


Composite Image by Shannon Colson  |  Pacific Northwest National Laboratory

Detecting, anticipating, and reasoning potential proliferation expertise and capabilities using domain knowledge extracted from publicly available data is a highly desired task that supports the nuclear nonproliferation mission. Existing research efforts rely on graph analytics and are reactive in nature; they primarily focus on co-citation network analysis of scientific literature in English. In comparison, our effort supports moving away from reactive analyses to take a proactive posture by:

  • Fusing a variety of multilingual heterogenous public data sources and converting unstructured data (e.g., scientific publication content) into structured domain knowledge—about global proliferation expertise—jointly from content and context representations.
  • Using dynamically evolving knowledge representations to enable predictive and prescriptive inferences to achieve real-time domain understanding and contextual reasoning about global proliferation expertise and capability development.
EXPERT diagram

Technical Approach

EXPERT is developing and deploying a series of usable artificial intelligence-driven analytics to enable descriptive, predictive, and prescriptive inferences used to transform nuclear nonproliferation monitoring globally.

  1. Descriptive analytics: Natural language processing models are used to extract structured knowledge—global and local content graphs combined with context graphs—from a variety of unstructured multilingual datasets with high velocity, volume, and complexity (e.g., linguistic annotations, entity and relation extraction, topic modeling, embedding learning, and importance attribution in scientific texts).
  2. Predictive analytics: To anticipate future activity and expertise evolution globally and answer operational questions, like “in what venue will a given country publish next?” (venues); “what topics will a given country publish on?” (tags); “which institutions will publish from a given country?” (institution), in-house developed deep learning models are extended with recently emerged developments from geometric deep learning, temporal graph networks, and autoregressive approaches applied to temporal knowledge graphs.
  3. Prescriptive analytics: To explain and potentially intervene into the developments of global proliferation expertise evolution and answer questions like, “what should a country do to acquire a specific capability?” (prescriptive) or “could a country have acquired a capability if an alternate event had happened?” (counterfactual reasoning), in-house developed ensemble models are used for causal discovery and treatment effect estimation approaches, combined with agent-based simulations of proliferation expertise and technology development evolution globally.

Our transformational, artificial intelligence-driven approach aims not only to provide deeper understanding of how publicly available data could be used to detect, monitor, forecast, and potentially prevent proliferation, but also to discover real-world examples of patterns and behavior to facilitate the investigation of potentially illicit proliferation activity (e.g., before and after the Joint Comprehensive Plan of Action).

This project is a collaboration with the University of Washington and Lawrence Berkeley National Laboratory.