Developments in machine learning downscaling for storm surge in the northern Adriatic sea

Accurate storm surge prediction is essential for coastal resilience and hazard mitigation, particularly as climate change increases the frequency and intensity of extreme events. While Machine Learning (ML) models have shown promise for downscaling storm surge predictions, they are often under-evaluated against high-resolution dynamical models and rarely tested on extreme conditions. This study addresses these limitations by integrating advanced dynamical modeling with ML approaches to evaluate storm surge prediction in the Northern Adriatic Sea. High-resolution simulations were developed using the SHYFEM hydrodynamic model with optimized configurations and high-quality forcing datasets. This provided a robust benchmark for assessing the performance of ML models ranging from simple Multivariate Linear Regression (MLR) to more complex Long Short-Term Memory (LSTM) networks. To improve model evaluation, a novel corrected mean absolute deviation (MADc) metric and a custom loss function (MADc2) were introduced, targeting improved performance on extreme events. Results show that while MLR offers computational efficiency, it lacks the ability to capture non-linear dynamics and extremes. In contrast, LSTM networks performed significantly better, particularly when trained using the MADc2 loss function. Training ML models on the output of the dynamical model led to strong consistency with observations, while direct training on tide gauge data at key locations (e.g., Punta della Salute and Trieste) revealed that some ML models could outperform the dynamical model in critical metrics. These findings highlight the potential of ML models, especially LSTM networks, as efficient and accurate alternatives to traditional numerical approaches. Given their lower computational demands and strong performance, ML techniques hold significant promise for operational storm surge forecasting, particularly in data-rich contexts or where computational resources are limited.