Campos-Caba et al. (2026) Storm surge dynamics in the northern Adriatic Sea: comparing AI emulators with high-resolution numerical simulations

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-08-07
• Authors: Rodrigo Campos-Caba, Paula Camus, Andrea Mazzino, Michalis Ioannis Vousdoukas, Massimo Tondello, Iván Federico, Salvatore Causio, Lorenzo Mentaschi
• DOI: 10.5194/nhess-26-3683-2026

Research Groups

• Department of Physics and Astronomy (DIFA), University of Bologna
• Interdepartmental Research Centre for Environmental Sciences (CIRSA), University of Bologna
• Departamento de Matemática Aplicada y Ciencias de la Computación, University of Cantabria
• Department of Civil, Chemical and Environmental Engineering, University of Genoa
• Istituto Nazionale di Fisica Nucleare, Sezione di Genova
• Department of Marine Sciences, University of the Aegean
• MV Coastal and Climate Research Ltd.
• HS Marine SrL
• CMCC Foundation – Euro-Mediterranean Center on Climate Change

Short Summary

This study compares the performance of machine learning (ML) emulators with a high-resolution numerical simulation for storm surge prediction in the northern Adriatic Sea. The results show that even simple models like Multivariate Linear Regression (MLR), when trained with the corrected Mean Absolute Deviation squared (MADc2) loss function, can achieve comparable performance to advanced neural networks while remaining orders of magnitude faster.

Objective

• To assess the ability of ML emulators to reproduce the statistical and physical characteristics of storm surge in the northern Adriatic Sea.
• To compare the performance of different ML architectures, including Multivariate Linear Regression (MLR), Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Hybrid Recurrent Neural Network (RNNh), Long Short-Term Memory (LSTM), and Hybrid Long Short-Term Memory (LSTMh).

Study Configuration

• Spatial Scale: The study focuses on the northern Adriatic Sea, with a spatial domain encompassing a broad portion of the region.
• Temporal Scale: The analysis spans the period 1987–2020, with hourly output from the SHYFEM-MPI hydrodynamic model.

Methodology and Data

• Models used: Multivariate Linear Regression (MLR), Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Hybrid Recurrent Neural Network (RNNh), Long Short-Term Memory (LSTM), and Hybrid Long Short-Term Memory (LSTMh).
• Data sources: The study uses observed data from tide gauges, atmospheric forcing from ERA5, and oceanographic variables from Med-MFC reanalysis.

Main Results

• The results show that even simple models like MLR can achieve comparable performance to advanced neural networks when trained with the MADc2 loss function.
• The TBPTT formulation consistently outperforms the sliding window approach in terms of bias and overall error characteristics.
• The adoption of MADc2 as a loss function enables smoother convergence in gradient-based optimization algorithms.

Contributions

• This study contributes to the understanding of ML emulators for storm surge prediction, highlighting the importance of selecting appropriate architectures and loss functions.
• The results demonstrate that simple models like MLR can be effective when trained with the correct loss function, offering a practical balance between computational efficiency and accuracy.

Funding

• This research was funded by the CMCC Foundation – Euro-Mediterranean Center on Climate Change.

Citation

@article{CamposCaba2026Storm,
  author = {Campos-Caba, Rodrigo and Camus, Paula and Mazzino, Andrea and Vousdoukas, Michalis Ioannis and Tondello, Massimo and Federico, Iván and Causio, Salvatore and Mentaschi, Lorenzo},
  title = {Storm surge dynamics in the northern Adriatic Sea: comparing AI emulators with high-resolution numerical simulations},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-3683-2026},
  url = {https://doi.org/10.5194/nhess-26-3683-2026}
}