Fast coastal wave climate prediction using autoencoders-based models

Accurate forecasting of nearshore wave climate is crucial for mitigating coastal risks and ensuring navigation safety. However, conventional numerical models remain too computationally intensive for real-time use in complex environments. This study introduces a surrogate modelling framework that combines the Maximum Dissimilarity Algorithm (MDA), K-means clustering, autoencoders, and feed-forward Multilayer Perceptron (MLP) networks to efficiently predict spatial wave fields. The framework uses MDA to optimally reduce the training dataset from offshore wave–wind reanalysis. K-means clustering segments the nearshore domain into homogeneous regions. Within each area, autoencoders learn compact representations of SWAN-simulated significant wave height, which MLPs predict from offshore forcing and decode to reconstruct spatial fields. The model is applied to the port of Augusta (Italy), known for complex wave transformation. Results show that the model reproduces SWAN outputs with low errors (RMSE generally below 0.1–0.15 m for the significant wave height and normalized errors within 5%–15%, consistent with typical wave model uncertainty), while achieving a computational speed-up exceeding a factor of 60 compared to full numerical simulations. Clustering further boosts reconstruction skill and spatial consistency. Note that the current implementation is trained on numerical model outputs, which may limit generalization to measurement-based datasets. Overall, the proposed methodology provides a scalable, computationally efficient alternative to traditional physics-based models, with strong potential for integration into operational coastal forecasting systems, supporting engineering decision-making and risk management through rapid, reliable wave predictions.