Comparison of ensembles of atmospheric dispersion simulations: Lessons learnt from the confidence project about uncertainty quantification

Work Package one (WP1) of the EU-funded project CONFIDENCE (COping with uNcertainties For Improved modelling and DEcision making in Nuclear emergenCiEs) is dedicated to uncertainties during the early phase of a radiological accident. More specifically, it consists of propagating input uncertainties through atmospheric dispersion models to generate a results ensemble in which radiological endpoints can be analysed, for example by threshold exceedance of dose reference levels.

The first step of any uncertainty propagation study consists of identifying and quantifying input uncertainties. Meteorological data (e.g. wind, rain fields’ forecasts) and source term (i.e. released rate of radionuclides as a function of time) are the key uncertainties during a nuclear crisis. The former was dealt with by using meteorological ensembles. For the latter, several scenarios were designed, from the most simple (a short release with crude perturbations on quantities, height and beginning time) to an ensemble of source terms, designed with the severe accident code ASTEC, including uncertainties. Finally, a significant literature review was undertaken to identify and characterise uncertainties linked to atmospheric dispersion models. Guidelines for ranking uncertainties in atmospheric dispersion were produced (Mathieu et al. 2017).

The second step was an uncertainty propagation exercise through atmospheric dispersion and radiological models, for both historical events and hypothetical scenarios. During this stage, participants from eight countries (Denmark, France, Germany, Greece, Hungary, the Netherlands, Norway and the UK) used the meteorological ensembles and release scenarios to propagate the uncertainties through their operational tools (Korsakissok et al. 2019). The level of uncertainties taken into account depends on the participant; some only propagated the meteorological ensemble, others used Monte Carlo methods to take into account all the identified uncertainties.

This exercise led to a tremendous amount of data: fields of atmospheric concentrations and deposition as a function of time, and associated doses, for a large number of simulations. Some lessons learnt relate to dealing with high-dimensional inputs (meteorological ensembles, source terms) and outputs, from very practical issues to more theoretical ones. This abstract aims at presenting a synthesis of the exercise, with a focus on issues related to the analysis and visualization of uncertainties, including statistical and graphical indicators to compare ensemble results.