An interactive multi-objective framework for groundwater inverse modeling

This work presents a novel interactive multi-objective framework to solve the groundwater inverse problem - focusing on finding the optimal conductivity fields given measurements of aquifer response (such as hydraulic heads). The framework is based on an interactive multi-objective genetic algorithm (IMOGA) that considers model calibration as a multi-objective optimization process with user preference (expressing the plausibility of parameter fields) as an additional objective along with quantitative calibration measures such as prediction error and regularization. Given this information, the IMOGA converges to a set of Pareto optimal solutions representing the best trade-off among all (qualitative as well as quantitative) objectives. Results for the IMOGA show incorporating the site expert's subjective knowledge about the hydro-geology of the modeled aquifer leads to more plausible and reliable calibration results.; Since the IMOGA is a population-based iterative search it requires the user to evaluate hundreds of solutions leading to the problem of 'user fatigue'. A two-step methodology is proposed to combat user fatigue. First the user is shown only a fraction of the total population in every generation by clustering potential solutions and only showing unique samples from distinct clusters. Next the unranked solutions are ranked using a surrogate model that 'learns' from the user preferences. Image processing tools are used to improve the clustering and learning algorithms to better mimic the human's visual evaluation of the parameter field. Such an approach is shown to reduce user fatigue by up to 50% without compromising the solution quality of the IMOGA.; An important part of groundwater inversion is assessing parameter uncertainty and its effect upon model predictions. This work uses a multi-level sampling approach to incorporate uncertainty in both large-scale trends and the small-scale stochastic variability. The large-scale uncertainty is addressed using model-averaging approaches such as GLUE and MLBMA. A geostatistical approach is adopted to create realizations of the small-scale variability. The prediction model is then run using the simulated fields to get the distribution of predictions.; These approaches are developed and tested on a hypothetical groundwater aquifer as well as a field-scale application based on the waste isolation pilot plant (WIPP) site.