Hai V. Pham, Department of Civil and Environmental Engineering, Louisiana State University, 3525 Patrick F. Taylor Hall, Baton Rouge, LA 70803: firstname.lastname@example.org; Frank T.-C. Tsai, Department of Civil and Environmental Engineering, Louisiana State University, 3418G Patrick F. Taylor Hall, Baton Rouge, LA 70803: email@example.com
Groundwater systems are complex and subject to multiple interpretations and conceptualizations due to a lack of sufficient information. As a result, multiple conceptual models are often developed and their mean predictions are preferably used to avoid biased predictions from using a single conceptual model. Yet considering too many conceptual models may lead to high prediction uncertainty and may lose the purpose of model development. In order to reduce the number of models, an optimal observation network design is proposed based on maximizing the Kullback-Leibler (KL) information to discriminate competing models. The KL discrimination function derived by Box and Hill  for one additional observation datum at a time is expanded to account for multiple independent spatiotemporal observations. The Bayesian model averaging method is used to incorporate existing data and quantify future observation uncertainty arising from conceptual and parametric uncertainties in the discrimination function. The goal of the optimal network design is to find the number and location of observation wells and sampling rounds such that the highest posterior model probability of a model is larger than a desired probability criterion (e.g., 95%). The optimal design is implemented to a groundwater study in the Baton Rouge area, Louisiana to collect new groundwater heads from USGS wells. Results show that model conceptualization strongly affects to optimal observation network design. Total model variance varying over time and space should be considered in the design procedure to account for various sources of the future observation uncertainty.
Keywords: Conceptual model uncertainty, optimal network design, MODFLOW, model discrimination