Surrogate CFD modelling for simulating wind velocity over coastal dune terrain

Abstract

Dunes play a crucial role in protecting coastal areas from flooding, erosion and supporting their fragile ecosystems. Coastal dune management is important to protect these coastal areas. Dunes are formed by various factors, among which the wind magnitude and direction are critical. Traditionally, Computational Fluid Dynamics (CFD) methods are used to model the wind flow over coastal dune terrains. However, these methods are computationally expensive, which limits their application for large and complex aeolian transfer models. This research proposes the implementation of Convolutional Neural Networks (CNNs) for CFD surrogate modelling to predict the wind velocity vectors over coastal dune terrain. This approach aims to reduce the computational cost while trading off some accuracy. Various CNN architectures and backbones are evaluated. The research found that the combination of the Feature Pyramid Network (FPN) architecture and densenet121 backbone provided the best performance, significantly reducing the prediction time compared to traditional CFD simulations. While the model shows some consistent errors in certain upwind and downwind regions, the results show the potential of CNN surrogate modelling to enhance coastal management by offering a faster alternative to CFD simulations. Further research should focus on expanding the dataset to assess the model’s generalizability and on exploring backbones and ensemble methods to further improve the model’s robustness and accuracy.