Enabling High-throughput Atmospheric Simulations for Radionuclide Background and Source Estimation

Abstract:

Predicting source or background radionuclide (RN) emissions is limited by the effort needed to run gas/aerosol atmospheric transport models (ATMs). For example, source estimation for a single-isotope detection can require thousands-to-millions of ATM simulations to estimate basic characteristics such as emission location, time, and magnitude. Reduced-order surrogate modeling is a technique that applies physical, dynamical, and/or statistical simplifications to accelerate an ATM and enable the analysis of larger datasets and more complex emission scenarios.

This talk introduces an Atmospheric Transport Model Surrogate (ATaMS) developed for the Hysplit (NOAA) ATM to accelerate transport simulation through model reduction, code optimization, and improved scaling on high performance computing (HPC) systems. ATaMS uses statistical sampling techniques to reassemble RN particle trajectories from a reference dataset and efficiently predicts plume evolution for background concentrations or emission source characteristics. For model scales on the order of 7-10 days and 1000s of kilometers, ATaMS was able to accelerate transport simulation by factors of 10x-100x and reproduce downstream detections within an order of magnitude. This capability has been integrated into HPC workflows for probabilistic source prediction and the estimation of the RN atmospheric background. The ATaMS-supported algorithms were used to analyze concentration observations spread over two years from three radioxenon monitoring stations in Japan. Results are presented on emission source estimates which account for confounding signals contributing to the background at each station. Error estimates and computational savings over traditional models show ATaMS and surrogate modeling facilitate effective gas/aerosol signal analysis for treaty monitoring and similar applications.