Conceptual model uncertainty in groundwater modeling: Combining generalized likelihood uncertainty estimation and Bayesian model averaging

Uncertainty assessments in groundwater modeling applications typically attribute all sources of uncertainty to errors in parameters and inputs, neglecting what may be the primary source of uncertainty, namely, errors in the conceptualization of the system. Confining the set of plausible system representations to a single model leads to underdispersive and prone-to-bias predictions. In this work, we present a general and flexible approach that combines generalized likelihood uncertainty estimation ( GLUE) and Bayesian model averaging ( BMA) to assess uncertainty in model predictions that arise from errors in model structure, inputs, and parameters. In a prior analysis, a set of plausible models is selected, and the joint prior input and parameter space is sampled to form potential simulators of the system. For each model, the likelihood measures of acceptable simulators, assigned to them based on their ability to reproduce observed system behavior, are integrated over the joint input and parameter space to obtain the integrated model likelihood. The latter is used to weight the predictions of the respective model in the BMA ensemble predictions. For illustrative purposes, we applied the methodology to a three-dimensional hypothetical setup. Results showed that predictions of groundwater budget terms varied considerably among competing models; despite this, a set of 16 head observations used for conditioning did not allow differentiating between the models. BMA provided average predictions that were more conservative than individual predictions obtained for individual models. Conceptual model uncertainty contributed up to 30% of the total uncertainty. The results clearly indicate the need to consider alternative conceptualizations to account for model uncertainty.