MAR2PROTECT - Data driven models for contrasting seawater intrusion via managed aquifer recharge

Since seawater intrusion has become a major threat to the security of aquifer resources, the MAR2PROTECT project aims to improve the current manage aquifer recharge (MAR) strategy to restrain seawater intrusion by injecting freshwater or treated wastewater. However, global change (GC) and climate change (CC) induce complexity and uncertainty in coastal aquifers. Our team is developing a robust framework to quantify uncertainty within the aquifer system, estimate the salinity intrusion process and evaluate MAR's expected performance. Given time series data of a different nature, we establish a data-driven model including ARMA and Gaussian processes to study the relationship between stationary or non-stationary processes like piezometer levels, sea level, temperature, and salinity. These models incorporate spatial and temporal analysis and predictions for the future. To quantify the impact of GC and CC, we utilize the data assimilation method, considering the related parameters as random variables or processes, computing the stochastic ODE/PDE model, obtaining posterior distribution for unknown parameters and states of interest, and making predictions for future months. On the other hand, MAR involves the pushback of seawater by freshwater, the seawater intrusion process, and even chemical reactions. In order to study the coupled, complicated system, we use physics-informed neural networks (PINNs), a grid-free method. PINNs insert physical equations with boundary and initial conditions into the neural network and improve the robustness of numerical results. Also, to study the impact of GC and CC on the MAR system, we insert random parameters into coupled systems and combine the PINNs with the Bayesian method to obtain the posterior distribution of MAR performance.