Performance-Oriented Supercritical Airfoil Sampling Method for Deep Learning and Geometry Predictions

In the design of supercritical airfoils for civil transport jets (e.g., long-range wide-body aircraft), it is essential to achieve a high lift-to-drag ratio ([Formula: see text]), favorable drag-divergence characteristics, and sufficient buffet margin. Deep learning technologies have been applied in the airfoil design process to predict optimized airfoil geometries or pressure distributions. To enhance the aerodynamic performance of supercritical airfoils, this study focuses on a deep-learning-based sampling method for supercritical airfoils. Unlike the conventional uniform geometric sampling approach for airfoils, the novel sampling method intends to achieve uniform distribution of the buffet-onset lift coefficient ([Formula: see text]) and drag-divergence characteristics is proposed via surrogate-model-based optimization. Additionally, the signed distance function method is used to transform the geometry and pressure distribution of samples into a more suitable format for deep learning training. Following training with the VGG16 convolutional neural network, the training results of supercritical airfoils obtained via the proposed uniform performance sampling method and the conventional uniform geometric sampling method are compared. The results indicate that samples generated by the uniform performance sampling method exhibit faster convergence rates and improved accuracy in predicting supercritical airfoil geometries from specified pressure distributions. Meanwhile, deep learning training was further performed on samples generated via the proposed uniform performance sampling method to inversely predict supercritical airfoil geometries from buffet characteristics, yielding promising results.