数据驱动的升力体飞行器表面热流快速预测

Data-driven rapid prediction of surface heat flux for lifting-body aircraft

  • 摘要: 为实现高速升力体气动热的快速预测、提升预测模型的泛化能力,本文针对一类三维升力体外形,基于参数化建模方法构建了包含15种外形、36种飞行条件,合计540个算例的升力体外形气动热数据集,并提出了一种基于切片和插值策略的数据预处理方式,将飞行器的外形及物面信息转化为神经网络能够识别的矩阵形式。然后,结合Transformer、多层感知机与UNet设计了一种编码器-解码器架构的网络模型D-TMU,并使用深度过参数化卷积替代了网络中的传统卷积。该模型能够在给定飞行器的几何形状、压强分布及飞行条件的基础上,直接预测其表面热流分布,避免了复杂的迭代计算。结果表明:D-TMU模型在测试集中的总体误差为1.21%,高热流区域误差为1.19%,关键点误差为0.97%,单个算例的平均预测时间仅为0.03 s,加速比可达6个数量级。这表明本文模型能够有效捕捉飞行器外形的全局特征,并充分学习局部特征之间的相关性,具有较高的预测精度和速度。同时,在应用于未参与训练的一般升力体外形时,模型仍具备一定的泛化能力。

     

    Abstract: This study developed a rapid method for aerodynamic heating prediction in high-speed lifting bodies. A dataset comprising 540 simulation cases was constructed through parametric modeling, encompassing 15 three-dimensional configurations and 36 flight conditions. To facilitate efficient neural network training, a geometry-aware preprocessing method based on surface slicing and interpolation was proposed, converting complex surface and geometry information into matrix representations. Subsequently, a D-TMU model based on deep learning was introduced. The model employed an encoder–decoder architecture integrating transformer modules, multilayer perceptrons, and UNet structures, with depthwise over-parameterized convolutions replacing standard convolution layers. D-TMU directly predicted surface heat flux distributions without iterative computations from input geometry, pressure distribution, and flight conditions. Validation results demonstrated the predictive performance of the model, with mean prediction errors of 1.21% on the test set, 1.19% in high heat flux regions, and 0.97% at key points. The average inference time per case was 0.03 seconds, representing computational acceleration a speedup of approximately six orders of magnitude compared over traditional CFD methods. These results indicated that D-TMU effectively captured global geometric characteristics and local feature interactions, maintaining achieving high predictive accuracy and computational efficiency. In addition, the model exhibited promising generalization capability for lifting body configurations beyond the training set.

     

/

返回文章
返回