Adaptive probabilistic modelling to support decision-making in the event of accidental atmospheric releases

In the event of accidental or malevolent atmospheric releases, decision-makers have to swiftly implement mitigating measures. Decisions are often based on the determination of danger zones and safe zones in which the concentration levels of substances emitted into the air are respectively above or below a given hazardous threshold. However, the maps representing the danger zones are established from atmospheric dispersion models whose input data on meteorology and the source term are uncertain. In addition, these maps are drawn from a limited number of simulations of atmospheric dispersion. Thus, if we consider confidence or credible intervals on low probabilities of exceeding concentration threshold, the grey zonein which no decision is possible can extend considerably. In this paper, we deal with this issue by developing a methodology to accurately estimate the probability of exceeding a concentration threshold of a substance adversely released in the atmosphere. Confidence or credible intervals associated with the probability of exceeding a given concentration are determined by taking into account the spatial correlation of the concentration field modelled by Gaussian processes. This methodology proves its effectiveness in lowering the significance limit of the probability estimates and allows for a more accurate estimate associated with a lower risk, especially in low probability areas. Moreover, it is applicable to various situations in terms of concentration threshold, accepted estimation risk and number of simulations. Finally, it appears promising for building maps of danger zone actually useful for decision-makers and will be implemented in a numerical decision-support tool following this work.

Automated summary

In the event of accidental or malevolent atmospheric releases, decision-makers need to take swift mitigating measures. However, the determination of danger zones and safe zones based on atmospheric dispersion models is uncertain due to unreliable input data. This paper proposes a methodology to accurately estimate the probability of exceeding a concentration threshold, taking into account spatial correlation and providing more accurate risk assessments in low probability areas. This methodology shows promise in creating useful maps for decision-makers and can be implemented in a decision-support tool.