Data-Driven Aerodynamic Performance Prediction Architecture with Self-Attention Mechanism

The computational fluid dynamics (CFD) and the wind tunnel test for obtaining aerodynamic performance is expensive and time-consuming. This paper proposed a data-driven aerodynamic performance prediction architecture, which can consider both aerodynamic shapes and flow conditions. The architecture is based on computer vision, which takes the grayscale image of the airfoil and flow conditions as input, learns an end-to-end mapping among the shapes, the flow conditions, and the aerodynamic forces. The flow conditions are considered by an adaptive layer norm and a scale parameter. To explain what the model has learned, attention matrices of different airfoils under different flow conditions are analyzed. Results show that the mean squared errors (MSE) of validation dataset and training dataset both converge to Formula: see text orders of magnitude, and in the test dataset, the relative errors between the predicted and the target aerodynamic forces are within 2.5%, which reflects that the model has good accuracy and generalization performance.