Dynamic Data-Driven Event Reconstruction for Atmospheric Releases

The role of an event reconstruction capability in a case of an atmospheric release is to characterize the source by answering the critical questions--How much material was released? When? Where? and What are the potential consequences? Accurate estimation of the source term is essential to accurately predict plume dispersion, effectively manage the emergency response, and mitigate consequences in a case of an atmospheric release of hazardous material. We are developing a capability that seamlessly integrates observational data streams with predictive models in order to provide probabilistic estimates of unknown source term parameters consistent with both data and model predictions. Our approach utilizes Bayesian inference with stochastic sampling using Markov Chain and Sequential Monte Carlo methodology. The inverse dispersion problem is reformulated into a solution based on efficient sampling of an ensemble of predictive simulations, guided by statistical comparisons with data. We are developing a flexible and adaptable data-driven event-reconstruction capability for atmospheric releases that provides (1) quantitative probabilistic estimates of the principal source-term parameters (e.g., the time-varying release rate and location); (2) predictions of increasing fidelity as an event progresses and additional data become available; and (3) analysis tools for sensor network design and uncertainty studies. Our computational framework incorporates multiple stochastic algorithms, operates with a range and variety of atmospheric models, and runs on multiple computer platforms, from workstations to large-scale computing resources. Our final goal is a multi-resolution capability for both real-time operational response and high fidelity multi-scale applications