High-resolution hydrodynamic information is essential for coastal and estuarine management, yet traditional numerical downscaling remains computationally expensive and limits its use in operational and predictive applications. This study investigates machine learning as a surrogate modelling approach to emulate high-resolution hydrodynamics in a semi-enclosed estuary, using Santander Bay as a case study.
Three techniques of increasing complexity are evaluated: K-nearest neighbours, Adaptive Boosting, and long short-term memory recurrent neural networks. The models are trained to reproduce sea surface height and surface currents from tidal, meteorological, and fluvial forcings, using outputs from a calibrated numerical model.
All approaches successfully reproduce the main hydrodynamic patterns, with performance improving with model complexity. Long short-term memory networks achieve the highest accuracy in tidally energetic regions, while Adaptive Boosting provides more stable results in low-energy and nearshore areas. All methods reduce computational cost by several orders of magnitude. Full-domain predictions require minutes to hours, compared to months of high-performance computing for numerical simulations.
The proposed framework enables rapid prediction of coastal circulation, supporting forecasting, transport modelling, and water quality assessment, and contributes to scalable, data-driven approaches for marine management.