Efficient Approaches for Offline Training of Gaussian Process Models in Coastal Groundwater Management. Presented at 1st IACRR International Conference on Coastal Reservoirs and Sustainable Water Management, Nanjing, China

We focuson the use of mathematical optimization techniques to manage freshwater demand and control saltwater intrusion (SWI) in coastal aquifers. Traditional methods involve linking variable density flow models with optimization algorithms, but these can be computationally expensive. We propose a more efficient approach using data-driven surrogates within the optimization loop, specifically focusing on a novel iterative search algorithm to select training points for surrogate model development.

The study applies this approach to a 2D model of the San Salvador Island aquifer in the Bahamas, considering trade-offs between groundwater supply cost, the produced groundwater rate, and SWI. The optimization problem aims to minimize operation cost and maximize fresh groundwater supply while adhering to constraints on aquifer drawdown and salt mass increase caused by pumping.

Gauss Process (GP) regression is employed to train surrogates for various parameters in the decision variable space, which includes pumping depth, distance from shoreline, and groundwater abstraction rate. Three GP model training strategies are proposed: iterative methods based on maximum distance, maximum gradient, and a score function.

Results and discussion focus on the efficiency of the proposed GP training strategies, comparing their performance in terms of the average probability of Pareto-optimality for pumping schemes. We conclude that the iterative strategy using the score function outperforms other methods in terms of computational costs and Pareto-optimal solutions.