航空学报 > 2025, Vol. 46 Issue (10): 631183-631183   doi: 10.7527/S1000-6893.2024.31183

飞行器设计生成式模型专栏

基于门控扩散模型的飞行器气动性能预测与反设计

楼锦华1,2, 陈荣钱1(), 柳家齐1, 鲍越1, 吴昊1, 尤延铖1   

  1. 1.厦门大学 航空航天学院,厦门 361005
    2.厦门大学 人工智能研究院,厦门 361005
  • 收稿日期:2024-09-10 修回日期:2024-10-12 接受日期:2024-11-15 出版日期:2024-12-02 发布日期:2024-11-29
  • 通讯作者: 陈荣钱 E-mail:rqchen@xmu.edu.cn

Aircraft aerodynamic performance prediction and inverse design based on a gated diffusion model

Jinhua LOU1,2, Rongqian CHEN1(), Jiaqi LIU1, Yue BAO1, Hao WU1, Yancheng YOU1   

  1. 1.School of Aerospace Engineering,Xiamen University,Xiamen 361005,China
    2.Institute of Artificial Intelligence,Xiamen University,Xiamen 361005,China
  • Received:2024-09-10 Revised:2024-10-12 Accepted:2024-11-15 Online:2024-12-02 Published:2024-11-29
  • Contact: Rongqian CHEN E-mail:rqchen@xmu.edu.cn

摘要:

针对传统深度学习模型在三维飞行器超声速条件下反设计中的收敛困难,提出了一种基于先验预测模型引导的生成式门控去噪扩散概率模型(DDPM)。该模型通过门控机制,集成了气动性能预测与飞行器外形反设计的功能,旨在提高预测精度、反设计效率,同时克服传统模型在处理高度非线性问题时的训练挑战。通过构建深度神经网络(DNN)作为先验预测模型,以获取初步的气动性能预测、几何外形估计。基于这些初步结果,门控DDPM结合DNN的预测信息,利用扩散、逆扩散过程来生成设计结果。扩散过程逐步添加高斯噪声,逆扩散过程则恢复数据分布,这一机制提升了模型在复杂气动设计任务中的收敛性、预测准确性。基于轴对称飞行器数据集,验证了门控DDPM在气动性能预测、反设计中的有效性。相比先验模型,门控DDPM在反设计任务中能够将大多数气动参数的相对误差控制在0.75%内,显著优于DNN模型。结果表明,所提方法有效提高了飞行器设计中的气动性能预测、外形反设计的精度。

关键词: 飞行器反设计, 去噪扩散概率模型, 气动性能预测, 门控机制, 生成式模型

Abstract:

To address the convergence challenges faced by traditional deep learning models in the inverse design of three-dimensional aircraft under supersonic conditions, a generative gated Denoising Diffusion Probabilistic Model (DDPM) based on a priori prediction model guidance is proposed. This model integrates aerodynamic performance prediction and aircraft shape inverse design through the gating mechanism, aiming to enhance both prediction accuracy and design efficiency while overcoming the training difficulties encountered in handling highly nonlinear problems with traditional models. A Deep Neural Network (DNN) is constructed to serve as the a priori prediction model to obtain preliminary aerodynamic performance predictions and geometric shape estimates. Based on these initial results, the gated DDPM uses the predictive information from the DNN and employs diffusion and reverse diffusion processes to generate design outcomes. In this process, Gaussian noise is gradually added during the diffusion phase, and data distribution is restored in the reverse diffusion phase. This mechanism enhances both convergence and predictive accuracy in complex aerodynamic design tasks. The effectiveness of the gated DDPM is validated using an axisymmetric aircraft dataset. Compared to the prior model, the gated DDPM can control the relative error of most aerodynamic parameters within 0.75% in inverse design tasks, significantly outperforming the DNN model. The results demonstrate that the proposed method can effectively enhance the accuracy of aerodynamic performance prediction and shape inverse design in aircraft design.

Key words: aircraft inverse design, denoising diffusion probabilistic model, aerodynamic performance prediction, gated mechanism, generative model

中图分类号: