Seawater intrusion is a growing concern for the safety of aquifer resources in coastal areas due to the combined effect of climate change and overexploitation. Hydrogeological uncertainty, mixed with the uncertainty associated with climate predictions, makes physical modeling unfeasible and/or unaffordable. As an alternative, we propose a data-driven method to analyze and predict the behavior of an unknown aquifer system. Our time-series data set is associated with the phreatic aquifer of the Emilia-Romagna Region (Italy). For each piezometer well, we develop a non-parameter statistic model tool (DRONE) by non-stationary Gaussian process (GP). The non-stationary kernel function can describe and predict the seawater intrusion process spatially and temporally. Then, we apply the Gaussian kernel function to merge the prediction model for different sites to generate the forecasting map for the entire regional coastal line. The general map presents the seasonal salinity distribution and supports risk assessment for the seawater intrusion process. To improve the predictive robustness and study the effect of climate change, we introduce a physically informed structure for our data-driven model. Since full-scale physical simulation is computationally onerous, our physical model is either an analytical or low-fidelity numerical scheme. Then, we treat the GP model as a model error to assimilate the physical model by observations. As such, the physically informed framework allows for long-term prediction and physical explanation to support the design of managed aquifer recharge (MAR) systems. Our study is in the context of the MAR2PROTECT project, funded by the EU, which aims to propose innovative MAR to prevent groundwater contamination.
Vittorio Di Federico (Wed,) studied this question.