With increasing demand and pressures on groundwater resources, accurate and reliable groundwater prediction models are essential for sustainable groundwater management. Groundwater models are merely approximations of reality, and we are unable to either fully characterize or mathematically describe the true complexity of the hydrologic system; therefore, inherent in all models are varying degree of uncertainty. A robust management policy should consider uncertainties in both the imprecise nature of conceptual/numerical models and their parameters. This study addresses the critical question of whether the use of multiple conceptual models to explicitly account for conceptual model uncertainty improves the ability of the models to assist in management decisions. Twelve unique conceptual models, characterized by three alternative geological interpretations, two recharge estimations, and two boundary condition implementations, were formulated to estimate sustainable extraction rates from Thailand’s Thaphra Area, where increasing groundwater withdrawals may result in water level declination and saline water upconing. The models were developed with MODFLOW and calibrated using PEST with the same set of observed hydraulic head data. All of the models were found to reasonably produce predictions of the available heads data. To select the best among the alternative models, multiple criteria have been defined and applied to evaluate the quality of individual models. It was found that models perform differently with respect to different evaluation criteria, and that it is unlikely that a single inter-model comparison criterion will ever be sufficient for general use. The chosen alternative models were applied both individually and jointly to quantify uncertainty in the groundwater management context. Different model-averaging methods were assessed in terms of their ability to assist in quantifying uncertainty in sustainable yield estimation. The twelve groundwater simulation models were additionally linked with optimization techniques to determine appropriate groundwater abstraction rates in the TPA Phu Thok aquifer. The management models aim to obtain maximal yields while protecting water level decline. Despite similar performances among the calibrated models, total sustainable yield estimates vary substantially depending on the conceptual model used and range widely, by a factor of 0.6 in total, and by as much as a factor of 4 in each management area. The comparison results demonstrate that simple averaging achieves a better performance than formal and sophisticated averaging methods such as Maximum Likelihood Bayesian Model Averaging, and produce a similar performance to GLUE and combined-multiple criteria averaging methods for both validation testing and management applications, but is much simpler to implement and use, and computationally much less demanding. The joint assessment of parameter and conceptual model uncertainty was performed by generating the multiple realizations of random parameters from the feasible space for each calibrated model using a simple Monte Carlo approach. The multi-model averaging methods produce a higher percentage of predictive coverage than do any individual models. Using model-averaging predictions, lower optimal rates were obtained to minimize head constraint violations, which do not ensue if a single best model is used with parameter uncertainty analysis. Although accounting for all sources of uncertainty is very important in predicting environmental and management problems, the available techniques used in the literature may be too computationally demanding and, in some cases, unnecessary complex, particularly in data-poor systems. The methods presented here to account for the main sources of uncertainty provide the required practical and comprehensive uncertainty analysis and can be applied to other case studies to provide reliable and accurate predictions for groundwater management applications.

The design of pump and treat systems for groundwater remediation is often aided by numerical groundwater modelling. Model predictions are uncertain, with this uncertainty resulting from unknown parameter values, model structure and future system forcings. Researchers have begun to suggest that uncertainty in groundwater model predictions is largely dominated by structural/conceptual model uncertainty and that multiple conceptual models be developed in order to characterize this uncertainty. As regulatory bodies begin to endorse the more expensive multiple conceptual model approach, it is useful to assess whether a multiple model approach provides a signi cant improvement over a conventional single model approach for pump and treat system design, supplemented with a factor of safety. To investigate this question, a case study located in Tacoma, Washington which was provided by Conestoga-Rovers & Associates (CRA) was used. Twelve conceptual models were developed to represent conceptual model uncertainty at the Tacoma, Washington site and a pump and treat system was optimally designed for each conceptual model. Each design was tested across all 12 conceptual models with no factor of safety applied, and a factor of safety of 1.5 and 2 applied. Adding a factor of safety of 1.5 decreased the risk of containment failure to 15 percent, compared to 21 percent with no factor of safety. Increasing the factor of safety from 1.5 to 2 further reduced the risk of containment failure to 9 percent, indicating that the application of a factor of safety reduces the risk of design failure at a cost directly proportional to the value of the factor of safety. To provide a relatively independent estimate of a factor of safety approach a single "best" model developed by CRA was compared against the multiple model approach. With a factor of safety of 1.5 or greater, adequate capture was demonstrated across all 12 conceptual models. This demonstrated that in this case using the single \best" model developed by CRA with a factor of safety would have been a reasonable surrogate for a multiple model approach. This is of practical importance to engineers as it demonstrates that the a conventional single model approach may be su cient. However, it is essential that the model used is a good model. Furthermore, a multiple model approach will likely be an excessive burden in cases such as pump and treat system design, where the cost of failure is low as the system can be adjusted during operation to respond to new data. This may not be the case for remedial systems with high capital costs such as permeable reactive barriers, which cannot be easily adjusted.

With increasing demand and pressures on groundwater resources, accurate and reliable groundwater prediction models are essential for sustainable groundwater management. Groundwater models are merely approximations of reality, and we are unable to either fully characterize or mathematically describe the true complexity of the hydrologic system; therefore, inherent in all models are varying degree of uncertainty. A robust management policy should consider uncertainties in both the imprecise nature of conceptual/numerical models and their parameters. This study addresses the critical question of whether the use of multiple conceptual models to explicitly account for conceptual model uncertainty improves the ability of the models to assist in management decisions. Twelve unique conceptual models, characterized by three alternative geological interpretations, two recharge estimations, and two boundary condition implementations, were formulated to estimate sustainable extraction rates from Thailand’s Thaphra Area, where increasing groundwater withdrawals may result in water level declination and saline water upconing. The models were developed with MODFLOW and calibrated using PEST with the same set of observed hydraulic head data. All of the models were found to reasonably produce predictions of the available heads data. To select the best among the alternative models, multiple criteria have been defined and applied to evaluate the quality of individual models. It was found that models perform differently with respect to different evaluation criteria, and that it is unlikely that a single inter-model comparison criterion will ever be sufficient for general use. The chosen alternative models were applied both individually and jointly to quantify uncertainty in the groundwater management context. Different model-averaging methods were assessed in terms of their ability to assist in quantifying uncertainty in sustainable yield estimation. The twelve groundwater simulation models were additionally linked with optimization techniques to determine appropriate groundwater abstraction rates in the TPA Phu Thok aquifer. The management models aim to obtain maximal yields while protecting water level decline. Despite similar performances among the calibrated models, total sustainable yield estimates vary substantially depending on the conceptual model used and range widely, by a factor of 0.6 in total, and by as much as a factor of 4 in each management area. The comparison results demonstrate that simple averaging achieves a better performance than formal and sophisticated averaging methods such as Maximum Likelihood Bayesian Model Averaging, and produce a similar performance to GLUE and combined-multiple criteria averaging methods for both validation testing and management applications, but is much simpler to implement and use, and computationally much less demanding. The joint assessment of parameter and conceptual model uncertainty was performed by generating the multiple realizations of random parameters from the feasible space for each calibrated model using a simple Monte Carlo approach. The multi-model averaging methods produce a higher percentage of predictive coverage than do any individual models. Using model-averaging predictions, lower optimal rates were obtained to minimize head constraint violations, which do not ensue if a single best model is used with parameter uncertainty analysis. Although accounting for all sources of uncertainty is very important in predicting environmental and management problems, the available techniques used in the literature may be too computationally demanding and, in some cases, unnecessary complex, particularly in data-poor systems. The methods presented here to account for the main sources of uncertainty provide the required practical and comprehensive uncertainty analysis and can be applied to other case studies to provide reliable and accurate predictions for groundwater management applications.