Aerodynamic performance prediction of 3D aircraft based on point cloud deep learning

Abstract

Abstract Accurate prediction of aerodynamic performance is crucial for the design and optimization of aircraft. To address the high costs of aerodynamics software for solving the aerodynamic responses of 3D aircraft and the limitations of traditional parametrization and 3D array shape representations, this paper proposes a PointConvAttn model based on point cloud data. The model is capable of predicting surface pressure and aerodynamic coefficients. It utilizes point cloud data to accurately describe shape features and to flexibly handle data at various resolutions. The model incorporates a convolutional and attention mechanism specifically tailored for point cloud data, which effectively balances the extraction of local and global features, reducing information loss and noise interference during sampling. Additionally, a transfer learning training method is employed in the prediction of lift and drag coefficients, which reduces training complexity and enhances prediction accuracy. In the experimental section, a specific unmanned aerial vehicle model is selected as the baseline configuration. Latin hypercube sampling is employed within the design space to generate multiple sets of shape parameters, thereby creating a data set that includes a variety of aerodynamic shapes. The corresponding aerodynamic response outputs are then calculated to serve as the experimental data. Results show that PointConvAttn delivers high-accuracy predictions with a mean error of 3.24% for pressure coefficients and 3.28% for lift coefficients, and drag coefficients across flight regimes. The trained model is capable of obtaining aerodynamic response results that are close to the true values within 2–3 s, offering a rapid, accurate, and cost-effective alternative for aircraft design optimization.

Topics & Concepts