Abstract
High-resolution hydrodynamic data are essential for coastal and estuarine management. However, traditional downscaling methods based on numerical modeling remain computationally expensive, limiting their applicability for long-term hindcasts and operational forecasting systems. This study evaluates the use of machine learning for the reconstruction of sea surface height and surface currents in a semi-enclosed estuary, using Santander Bay as a case study. Three techniques spanning increasing model complexity are analyzed: K-nearest neighbors, Adaptive Boosting, and long short-term memory networks. The models are trained to emulate high-resolution hydrodynamic-model outputs using a comprehensive set of tidal, meteorological, and fluvial forcings. Performance is assessed through spatial validation, cluster-based analysis, representative-point time series, and independent comparison against in-situ observations. Results show that all techniques successfully reproduce the main hydrodynamic patterns, with accuracy increasing with model complexity. Long short-term memory networks achieve the highest skill in tidally energetic regions, while Adaptive Boosting provides more stable performance in low-energy and shoreline areas. Computational cost analysis demonstrates that all machine-learning approaches achieve speedups of several orders of magnitude relative to numerical modelling, with inference costs that are negligible at both point and domain scales. These findings demonstrate the potential of machine learning as a computationally efficient approach for high-resolution modelling of coastal hydrodynamics, with important implications for operational forecasting and coastal management applications.