Bidirectional prediction between wake velocity and surface pressure using deep learning techniques

Abstract

The surface pressure and flow field of rectangular cylinders are of great importance in aerodynamic analyses of the cylinders. In general, it is easy to obtain one side of the information, either the surface pressure or the flow field, in reality. Deep learning (DL) techniques provide a new perspective to infer one side of the information from the other. Novel DL algorithms, specifically Dense Neuron Networks (DNN) and Graphic Attention Networks (GAT), are incorporated into the proposed high accuracy bidirectional prediction models in order to tackle the practical problems above. DNN employs a sequential compression architecture with a residual connection, and GAT applies an attention mechanism to update node value by connection edges defined by the relative position. The results demonstrate that in predicting surface pressure using wake velocity, GAT exhibits a 50% lower mean square error and more stable training progress than the DNN model. Predicting wake velocity using surface pressure yields accurate results for both DNN and GAT models. Specifically, the GAT structure shows better performance in capturing the vortex information near the trailing edge of the cylinder. Comparison of two models suggests that the GAT capability of rationally defining the interconnection of nodes through edges is advantageous in solving flow problems involving a spatially generalized physical mechanism.