资源环境科技发展态势分析平台

Groundwater age is an important indicator of groundwater susceptibility to anthropogenic contamination and a key input to statistical models for forecasting water quality. Numerical models can provide estimates of groundwater age, enabling interpretation of measured age tracers. However, to extend to national-scale groundwater systems where numerical models are not routinely available, a more efficient metamodeling approach can provide a less precise but widely applicable estimate of groundwater age, trained to make forecasts based on predictor variables that can be measured independent of numerical models. We trained gradient-boosted regression tree statistical metamodels to MODFLOW/MODPATH-derived groundwater age estimates in five inset models in the Lake Michigan Basin, USA. Using high-throughput computing, we explored an exhaustive range of tuning parameters and tested metamodels through cross validation, a 20% holdout, and a round robin approach among the five inset models withholding each inset model from training and testing on the held-out inset model. Forecast skill-measured by Nash Sutcliffe efficiency-was high for age-related responses in the 20% hold-out case (ranging from 0.73 to 0.84). The round robin analysis provided the opportunity to explore extending to unmodeled areas and a greater range of skill indicated the need to evaluate when it is appropriate to apply a metamodel from one region to another. We further explored the ramifications of metamodel simplification achieved through removing predictor variables based on their estimated importance. We found that similar metamodel performance was achievable with a fraction of the candidate set of predictor variables with well construction variables being most important.

Plain Language Summary Constructing computer models of groundwater and making forecasts to help water resource managers can both be time consuming. Metamodeling takes advantage of recent advances in artificial intelligence to make simple, fast-running statistical models learned from paired inputs and outputs generated by a more complicated computer model. These metamodels are easier to construct, based on salient environmental characteristics, which can be obtained using maps spanning large scale. When metamodels adequately duplicate the underlying model behavior, they can be extended into previously unmodeled areas. In this work, we created metamodels to forecast groundwater age in pumping wells in the Lake Michigan Basin. Groundwater extracted from wells represents a mixture of water ranging in age from recent to thousands of years old. Water younger than 65years old is more likely to contain nitrate as this time period saw high use of inorganic nitrogen for agricultural fertilizer resulting in nitrate entering the groundwater. The ability to rapidly create a metamodel of a previously unmodeled area allows forecasting and mapping groundwater age over large regions. Our metamodels reproduce 73% to 84% of the underlying training model information. This prototype screening approach, with age indicating nitrate vulnerability, can be extended for planning new well locations throughout the glaciated northern United States.