Machine learning-powered surrogate models for optimising island groundwater management: striking a balance between cost, sustainability, and environmental impact

The strategy of coastal pumping optimisation is commonly used to address seawater intrusion (SWI), striking a balance between water demand and environmental protection. However, applying the popular tools, the simulation-optimisation (SO) method to solve the pumping optimisation problem formulated on a three-dimensional (3D) aquifer usually faces a computational burden, as 3D simulation models usually comprise numerous cells, causing each SWI simulation to be time-consuming. Even assisted by the surrogate-based SO (SSO) system where the SWI simulator is replaced by surrogate models during the optimisation, which has been validated more efficiently than the SO framework, solving a multi-objective groundwater management problem is still time-consuming when it comprises many decision variables. Therefore, this PhD research aims to develop a more efficient SSO framework for multi-objective groundwater management formulated on 3D coastal aquifers.
To achieve this aim, five objectives have been formed, including: 1) investigating the influence of pumping patterns on the SWI extent and groundwater supply cost; 2) exploring the efficient algorithm for offline training surrogate models; 3) exploring the efficient algorithm for online training surrogate models; 4) identifying the advantages, disadvantages of applying offline-trained and online-trained surrogate modes; 5) developing an efficient SSO framework for sustainable island groundwater management using a 3D island aquifer model based on the findings from the previous objectives. In this study, the SEAWAT model is used to simulate SWI under the pumping. The Gaussian Process (GP) modelling technique is adopted in this study to construct surrogate models, and the full enumeration method is applied to determine optimal solutions because of the inexpensiveness of the GP models.
Those research objectives have been tackled by conducting analyses through a two-objective groundwater management problem formulated on a simplified coastal aquifer created by hydrogeological conditions observed on San Salvador Island (Bahamas), minimizing groundwater supply cost and maximizing the volume of groundwater sent to the network. Key research findings indicate: 1) strengthening SWI constraints is prone to placing pumping wells closer to the shoreline or/and within deeper aquifers, leading to pumping saltier groundwater and thus enhancing the treatment cost; 2) introducing the iterative process can improve the efficiency in offline training GP models compared with the traditional offline training strategy, and the most appropriate algorithm is to select new points at each iteration by integrating information about their distances from known points and the gradients of estimates; 3) discretizing the objective space into equal sub-regions based on the obtained Pareto front and then selecting sampling points within these sub-regions can facilitate the convergence of online-trained GP models. 4) given limitations on SWI, online-trained GP models can produce more reliable optimal pumping schemes with higher efficiency, but offline-trained GP models can offer reliable predictions across the entire input space, adaptive to variations in SWI constraint conditions. Therefore, this study develops an efficient SSO framework relying on the offline-trained GP models to investigate two-objective island groundwater management in 3D aquifers under various constraint scenarios. Results indicate the proposed SSO framework can efficiently provide trustable optimal pumping schemes.