A novel deep learning-based emulator for predicting storm-driven nearshore submerged coastal morphodynamics

Rapid and accurate forecasting of storm-driven nearshore morphodynamic change is essential for effective coastal management and early warning systems. This study presents a deep learning-based emulator designed to predict two-dimensional, time-dependent submerged sediment erosion and deposition patterns under storm forcing. Trained and validated on data from high-resolution hydro-morphodynamic simulations using the process-based model XBeach. The emulator incorporates a convolutional encoder with spatial and channel attention mechanisms, ConvLSTM for temporal sequence modeling, and metadata-driven hydrodynamic forcing inputs (water level, significant wave height, peak wave period, and main wave direction), thereby enabling generalization across a wide range of storm conditions. Across synthetic test scenarios, the model achieves a coefficient of determination (R²) of up to 0.94, along with low Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) for both instantaneous and cumulative bed level changes. When evaluated against a real storm event, the emulator maintains robust performance, achieving an RMSE of 0.084 m at the final time step. Compared to traditional numerical simulations, the emulator offers approximately 100 × computational speedup while preserving spatial and temporal accuracy. These results highlight the potential of deep learning-based emulators as efficient surrogates for high-resolution process-based nearshore morphodynamic simulations and as enabling tools for rapid multi-scenario “what-if” analyses within Digital Twin–based coastal management frameworks.