AI-based Emulation for Assessing the Impact of Nature-based Solutions on Waves and Currents for Coastal Protection and Resilience

In recent years, several studies have investigated the effectiveness of Nature-based Solutions in mitigating extreme coastal hazards such as waves, storm surges, and erosion, highlighting their significant role in coastal protection. Observational measurements have provided the basis for understanding ocean–vegetation interactions, enabling parameterizations that have been incorporated into numerical models. These models are widely used to simulate real conditions and explore what-if scenarios involving plant phenotypic traits, species composition, and the spatial distribution and organization of vegetated meadows.

However, such simulations are computationally demanding, limiting their applicability in operational and exploratory contexts. To overcome this limitation, we exploit machine learning and artificial intelligence to develop numerical emulators within a digital twin framework. Several models were tested, including Random Forest, LightGBM, SVM, and deep learning architectures. Among them, a U-Net-like architecture demonstrated the best performance, and the results are here shown.

The training dataset consists of one year of numerical simulations for 28 vegetation configurations, generated by varying shoot density, leaf length, and leaf width. Simulations were produced using the SHYFEM-MPI circulation model coupled with the WAVEWATCH III wave model, incorporating the vegetation formulation of Shirinov et al. (2025). The AI emulator estimates vegetation-induced impacts on multiple ocean variables, including significant wave height, mean wave period and direction, near-bottom orbital velocity, and currents.

Results show that the AI emulator accurately captures nonlinear wave–vegetation interactions, reproducing wave attenuation and current modulation at high spatial resolution across two regional pilot areas. The model generalizes well, providing reliable estimates for intermediate vegetation configurations not included in the training dataset. Low error levels across variables and temporal consistency of the results demonstrate the robustness and stability of this approach.

This work highlights the potential of integrating artificial intelligence into predictive coastal modeling as a science-based risk assessment tool for evaluating the effectiveness of Nature-based Solutions, significantly enhancing coastal protection strategies.