Probabilistic safety analysis for urgent situations following the accidental release of a pollutant in the atmosphere

This paper is an original contribution to uncertainty quantification in atmospheric transport & dispersion (AT&D) at the local scale (1–10 km). It is proposed to account for the imprecise knowledge of the meteorological and release conditions in the case of an accidental hazardous atmospheric emission. The aim is to produce probabilistic risk maps instead of a deterministic toxic load map in order to help the stakeholders making their decisions. Due to the urge attached to such situations, the proposed methodology is able to produce such maps in a limited amount of time. It resorts to a Lagrangian particle dispersion model (LPDM) using wind fields interpolated from a pre-established database that collects the results from a computational fluid dynamics (CFD) model. This enables a decoupling of the CFD simulations from the dispersion analysis, thus a considerable saving of computational time. In order to make the Monte-Carlo-sampling-based estimation of the probability field even faster, it is also proposed to recourse to the use of a vector Gaussian process surrogate model together with high performance computing (HPC) resources. The Gaussian process (GP) surrogate modelling technique is coupled with a probabilistic principal component analysis (PCA) for reducing the number of GP predictors to fit, store and predict. The design of experiments (DOE) from which the surrogate model is built, is run over a cluster of PCs for making the total production time as short as possible. The use of GP predictors is validated by comparing the results produced by this technique with those obtained by crude Monte Carlo sampling. Accidental toxic release in a built environment – risk map (at 1.5 m above the ground level) emphasizing the safe (risk is less than 2.5%), dangerous (risk is greater than 97.5%) and uncertain zones (risk is comprised between 2.5% and 97.5%), estimated using Monte Carlo sampling with 10,000 runs of the vector GP surrogate model. The orange line corresponds to the median critical toxic load contour level while the black line corresponds to that obtained from the deterministic risk assessment study. [Display omitted] •We quantify uncertainty in case of accidental atmospheric dispersion at local scale.•We account for imprecise knowledge of the meteorological and release conditions.•We resort to Lagrangian dispersion and to a vector Gaussian process surrogate model.•We validate the Gaussian process predictors by comparison with Monte Carlo sampling.•W