Predicting Offshore Freshened Groundwater via Machine Learning and Surrogate Modelling

Coastal regions worldwide face increasing water stress, making unconventional resources like Offshore Freshened Groundwater (OFG) critically important. However, characterizing these vast subterranean reservoirs is hindered by the scarcity of direct subsurface data, and current predictive methods are either too coarse for local assessment or qualitative in nature. This study introduces a novel quantitative methodology to predict OFG distribution using machine learning (ML) trained on a synthetic dataset derived from geologically realistic surrogate models. The workflow involves generating numerous surrogate models of continental shelves based on globally available geomorphological data. We then run numerical simulations of variable-density groundwater flow on these models, forced by glacial-interglacial sea-level cycles, to create a robust training dataset linking geological geometry to OFG system characteristics. This study details the parameterization of surrogate continental shelf models from 8 distinct global regions into numerical feature vectors suitable for ML. Initial results indicate that key geometric parameters, such as the offshore extent of the primary aquifer and the inland topographic gradient, are first-order controls on the volume and distribution of emplaced OFG. This proof-of-concept validates that the surrogate modelling framework can effectively capture the sensitivity of OFG systems to geological controls. Ultimately, this methodology highlights a potential pathway to overcoming the data-scarcity challenge, enabling the development of a predictive tool for rapid, quantitative assessment of OFG resources on continental margins worldwide.