Metamodels to Bridge the Gap Between Modeling and Decision Support

Insights from process-based models are a mainstay of many groundwater investigations; however, long runtimes often preclude their use in the decision-making process. Screening-level predictions are often needed in areas lacking time or funding for rigorous process-based modeling. The U.S. Geological Survey (USGS) Groundwater Resources and National Water Quality Assessment Programs are addressing these issues by evaluating the “metamodel” to bridge these gaps. A metamodel is a statistical model founded on a computationally expensive model. Although faster, the question remains: Can a statistical model provide similar insights to a numerical model with faster results? Metamodeling was developed to overcome long runtimes for sensitivity analysis (Blanning 1975); our focus is decision support applications. Two representative groundwater applications are: (1) the contribution of surface water to wells in shallow groundwater systems (e.g., Fienen and Plant 2014), and (2) unsaturated zone nitrate flux to groundwater (e.g., Nolan et al. 2012). The first step is to generate a representative sample of input/output combinations from the numerical model over a range of conditions. This variability is especially important when propagating uncertainty to predictions. Variability can be represented by many model runs using different input values or by few model runs with samples scattered in space/time experiencing the range of natural system variability. In the second step, a statistical learning technique is selected with which a predictive model can be “learned” from the data derived from the model. Techniques include