A data-driven model for predicting velocity field around a circular cylinder based on pressure on the cylinder

The flow field in bluff-body fluid dynamics involves high-dimensional spatial-temporal evolution laws. Though such problems have caused numerous concerns and researches, it is far from getting an analytical solution. However, numerous data from experiments and high-fidelity numerical simulations make it possible to employ deep learning to establish a data-driven quantitative relation model for bluff-body fluid dynamics. Among deep learning algorithms, convolutional neural networks (CNNs) are suitable for extracting abstract features from grid-like data. In this work, a deep CNN architecture consists of paths with and without a pooling layer is employed to establish the model for predicting velocity field around the circular cylinder based on pressure on the cylinder surface. The input of the CNN is the pressure fluctuations on the cylinder surface, which are transformed into a grid-like topology. The output of the CNN is the velocity field around the circular cylinder to be predicted. The CNN is trained by Adam algorithm. The predicted results over various Reynolds numbers indicate that the intrinsic relationship between the wake and the pressure on the cylinder surface is learned by the data-driven model.