资源环境科技发展态势分析平台

When applying environmental models, the choice of model complexity and the design of field campaigns depend on each other and on the modeling/prediction goal. We propose jointly optimizing model complexity and data collection (design) by maximizing the expected performance for the modeling goal. We use ensembles of highly resolved virtual realities and of less complex modeling variants that differ in their degrees of upscaling and simplified parameterization. For each design under consideration, we simulate hypothetical measurement data (subject to noise) with all realizations of all models. To mimic model calibration with hypothetical data, we identify pairs of best‐fitting realizations between virtual reality and each model variant for each design. Then, we emulate model choice by selecting (across the model variants, for each design and for each virtual reality) the pair that shows the best predictive match in the modeling goal. Finally, we identify the model/design combination that offers, on average over all virtual realities, the best predictive match. As a test application, we consider a heterogeneous, stratified aquifer, in which the stratification enhances hydraulic anisotropy on the macro‐scale. We define two different modeling goals: (a) estimating the hydraulic conductivity tensor upscaled to the full aquifer thickness and (b) predicting the pumping rate needed to dewater a construction pit. Our results indicate that jointly optimizing observation networks and model selection can reduce the prediction uncertainty of parameters at lower experimental costs. We also show that the involved trade‐offs between model complexity and required design depend on the target quantity.