Author(s): Zexia Zhang; Fotis Sotiropoulos; Ali Khosronejad

Linked Author(s):

Keywords: No Keywords

Abstract: To overcome the high computational cost of high-fidelity models, we developed autoencoder convolutional neural networks (CNNs) algorithms to reconstruct the first- and second-order statistics of turbulent flood flow in large-scale rivers. The training dataset for construction of the CNN models is obtained from large-eddy simulation (LES) results of the flood flow in a large-scale meandering river with several piers foundations. The developed CNNs are validated using separately done LESs of meandering rivers with different piers foundation configurations. The results show good agreement between the LES and CNNs predictions while the CNNs are several orders of magnitude computationally more efficient than LES.

DOI:

Year: 2022