A preliminary assessment of machine learning algorithms for predicting CFD-simulated wind flow patterns over idealised foredunes

Foredunes play an important role in protecting coastal communities and their assets. The use of Computational Fluid Dynamics (CFD) to simulate wind flow over foredunes has great potential because it can enable three-dimensional visualisations of the flow field, critical to predicting sediment pathways. Generalised conceptual models of foredune evolution and maintenance can then be built and revised over time as more evidence from the field becomes available. Obtaining field data is, however, time consuming, costly, and weather dependent. CFD is ideally suited to explore what happens when wind transitions across the beach and encounters the stoss face of the foredune. A simple dune shape is used in CFD simulations to tease out the influence of various dune parameters under varying wind conditions. However, it is computationally expensive to run CFD simulations for all combinations of parameters. Representative data were used to train machine learning algorithms, and the results were compared to predicted CFD simulations. The machine learning algorithms were able to identify the cases when recirculation vortices were present and to some extent their relative scales and locations, allowing the exploration and identification of key parameters related to wind flow and dune geomorphology that are associated with turbulent flow structures.

岸前沙丘（Foredunes）在保护沿海社区及其相关资产方面发挥着重要作用。采用计算流体动力学（Computational Fluid Dynamics，CFD）模拟岸前沙丘上方的气流具有巨大应用潜力，因其可实现流场的三维可视化，这对预测泥沙运移路径至关重要。随后，可随着野外实地证据的不断积累，逐步构建并完善岸前沙丘演化与维持的广义概念模型。然而，获取野外实地数据不仅耗时耗财，还受天气条件制约。CFD非常适合探究气流跨越海滩并接触岸前沙丘迎风面时的流动状态。在CFD模拟中，研究者采用简化的沙丘形态，以厘清不同风速条件下各类沙丘参数对流动的影响。然而，针对所有参数组合开展CFD模拟的计算成本极高。研究人员采用代表性数据集训练机器学习算法，并将训练结果与CFD模拟的预测结果进行对比。上述机器学习算法可准确识别存在回流涡的工况，并能在一定程度上预测回流涡的相对尺度与位置，从而助力探索并识别出与气流流动及沙丘地貌相关、且与湍流结构存在关联的关键参数。