Prediction of three-dimensional flow characteristics for cylinders with wavy geometric disturbance using deep learning models

For the first time, this study extracts three-dimensional flow characteristics of geometric disturbances as passive flow control using deep learning models, which will contribute to innovatively advancing research in passive flow control. It addresses the limitations of predicting two-dimensional flow and the lack of methods for predicting flow characteristics under variable geometric disturbances. Specifically, the study establishes a convolutional neural network (CNN) and an encoder–decoder (ED) model to predict 3D flow characteristics for cylinders with wavy geometric disturbances. To generate the dataset for the deep learning models, large eddy simulation is utilized. The accuracy of the CNN model is validated by comparing the predicted and true force coefficients. The study also examines common features of mean 3D flow structures, reconstructed by the established ED model. The ED model constructs longer 3D spanwise vorticity structures, jet-like flows, and weaker turbulence kinetic energy, which are correlated with smaller force coefficients. Finally, parametric studies of the geometric disturbances are conducted using the established deep learning models. The ED model generates reasonable mean 3D flow fields, supporting the physical understanding of changes in force coefficients. Consequently, approaches presented in this study are reliable and efficient for conducting parameter studies on geometric disturbances. •The CNN and ED are established to predict 3D flow characteristics of the cylinders with wavy geometric disturbance.•The force coefficients and the mean 3D flow structures are predicted by established CNN and ED models.•The parametric studies are performed using the established models and provide the maps of force coefficients.