Solving computationally-demanding reliability-based design problems in hydrogeology.

The design of technologies used to manage groundwater quality or quantity typically involves the use of groundwater models and an optimisation algorithm. Further, decision makers increasingly request explicit consideration of the uncertainty of model parameters and predictions. Performing Monte Carlo simulations leads to a stack of multiple aquifer realisations, which must then be considered simultaneously during optimisation. The evaluation of the performance of a particular design in every single realisation of the stack provides an estimate of the expected reliability of the design, i.e. its chance to reach a given management target. However, reliability-based optimisation may become computationally impractical if a large number of realisations have to be considered. To substantially reduce the required number of realisations, we propose a new approach that works with small subsets of realisations, which are dynamically drawn from a much larger repository stack. The drawing is controlled by a so-called stack-ordering procedure that ranks the realisations with respect to criticalness. The results of an example application are promising, with computational savings of up to 98.5% and clear improvements as compared to random sampling.