REPRESENTING UNCERTAINTY USING DIVERSE MODEL ENSEMBLES: A TEST CASE IN AN ALPINE KARST SYSTEM

When disputes arise over groundwater resources, policy-makers increasingly rely on models to make decisions. However, different models often explain available data equally well. This project tests a model ensemble approach for decision-support in karst systems.

Karst aquifers are difficult to model because the discrete, heterogeneous nature of groundwater flow through conduits, rather than through distributed pore spaces, leads to high structural uncertainty. Existing models rely either on a detailed conduit map, or on effective flow parameters approximating a porous medium. Neither approach is adequate for most karst systems, where conduits are unmapped, yet flow patterns are fundamentally different from those in porous media.

Our approach links three components: 3D geologic modeling with GemPy, an open-source Python package; conduit network generation with the Stochastic Karst Simulator (SKS), a pseudo-genetic structural model; and pipe flow modeling with the EPA Storm Water Management Model (SWMM). We use pre-existing data from a long-term research site, the Gottesacker karst system in the German/Austrian Alps.

First, competing geologic models are built in GemPy, based on existing maps. Each geologic model is then fed to SKS, which generates many possible conduit networks. For each proposed network, the flow behavior is modelled in SWMM, which requires assigning conduit hydraulic parameters. This yields an ensemble of competing models, organized into a model tree recording the geologic structure, conduit network, and hydraulic parameters for each model.

The likelihood of each model will be assessed by comparing model-predicted spring discharge timeseries to observed data. The models that best reproduce discharge behavior can then be compared to the known conduit network, to assess the effectiveness of this approach.

The impact on stakeholders of the predicted discharge behavior for each model will also be assessed, to identify the most consequential uncertainties, as well as to guide additional data collection to reduce uncertainty. In a risk assessment context, uncertainty reduction targeting models of concern (those predicting highly undesirable behavior) focuses on the uncertainties that matter most to stakeholders, increasing the usefulness of model predictions for decision-making.