A deep learning framework for aerodynamic pressure prediction on general three-dimensional configurations

In this paper, a deep learning framework is proposed for predicting aerodynamic pressure distributions in general three-dimensional configurations. Based on the PointNet++ structure, the proposed framework extracts shape features based on the geometric representation of point cloud, outputs pressure coefficients corresponding to locations, and is able to accept inputs of point clouds with different resolutions. By PointNet++, we mean that local and global features of three-dimensional configurations could be effectively extracted through a multi-level feature extraction structure. A parametric approach is utilized to generate 2000 different space shuttle three-dimensional shapes, and their flows at the hypersonic speed are solved by computational fluid dynamics (CFD) as a dataset to support the training of the deep learning. Within the dataset, accurate predictions of pressure and aerodynamic forces are demonstrated for 400 unseen testing shapes. Out of the dataset, geometries that are tested for generalizability include slender, waverider, spaceplane, Apollo capsule, lifting body, and rocket. Remarkably, the framework is capable of predicting pressure distributions and aerodynamic forces for the unseen, independently designed geometries of various types in near-real-time and near-CFD accuracy, proving its excellent applicability to general three-dimensional configurations.