Dynamic Mode Decomposition on pressure flow field analysis: Flow field reconstruction, accuracy, and practical significance

Abstract

This study applies the Dynamic Mode Decomposition (DMD) technique to a prototypical wind engineering problem of flow past a square prism at a Reynolds number of 22,000 to investigate the DMD’s accuracy and practical values in pressure flow field analysis. In reconstructing the original pressure field obtained by Large-Eddy-Simulations (LES), a full-order DMD model achieves a stellar accuracy within 0.1% mean error. The model also provides a computationally simplistic alternative in lieu of the strenuous POD reconstruction. Spatiotemporal analysis reveals two types of reconstruction errors: Spatial error arising from DMD’s inherent linear approximations on nonlinear phenomena, and temporal error arising from DMD’s assumption on temporal behaviors of flow mechanisms. Additionally, reduced-order DMD models achieve an acceptable accuracy within 0.9% mean error, and aptly capture macroscopic flow features, while reducing the size of the essential data needed for flow field construction by as much as 25 times. In wind engineering applications, where flow field datasets are extensive but microscopic flow features are unnecessary, the usage of reduced-order DMD models greatly alleviates the intense computational burden on flow field post-analysis. With an adequate balance between model size and reconstruction accuracy, DMD proves an accurate and practically beneficial technique for wind engineering applications.