| |
Groundwater flow models deliver predictions of the impacts of human actions or natural phenomena on groundwater resources based on an underlying set of assumptions regarding the physical characteristics of the hydrogeologic environment. In the most general sense, we assume that models can reliably predict future impacts if they reasonably simulate current conditions. The degree to which they do this is evaluated through the model calibration process where good calibration to existing conditions indicates that the underlying model assumptions are reasonable and the predictions can be regarded as reasonably reliable. But, what defines “good calibration” or even “existing conditions?”
Here we demonstrate through an analysis of regional-scale groundwater flow models that “good calibration” can be misleading, that by expanding the tests of “existing conditions” a substantially different picture of model reliability can be revealed, and by adapting the underlying assumptions, a substantially better match to a more reasonable and robust definition of existing conditions can be achieved. We show that a simple calibration to heads through the evaluation of average absolute residuals is insufficient to provide a meaningful test of model reliability. For the evaluation presented, we show that despite being described as well-calibrated on the basis of only average absolute residual, the model displayed a spatial pattern of large residuals outside of the acceptable range, failed to reasonably simulate observed cones-of-depression, failed to reasonably simulate mapable springshed boundaries, and failed to reasonably conform to sub-watershed water budgets defined on the basis of measured discharge. The result was a model for which more than 40% of the simulated extractions could be sourced to model boundary assignments. By comparison, we show that by changing the underlying assumptions and adhering to a more rigorous definition of calibration, all of these conditions can be adequately simulated.
|