首页|期刊导航|南京航空航天大学学报(英文版)|基于混合CNN与Chebyshev-Kolmogorov-Arnold网络的高精度翼型压力分布预测模型

基于混合CNN与Chebyshev-Kolmogorov-Arnold网络的高精度翼型压力分布预测模型OA

A High-Precision Airfoil Pressure Distribution Prediction Model Based on Hybrid CNN and Chebyshev-Kolmogorov-Arnold Network

中文摘要英文摘要

在翼型气动设计与优化中,传统计算流体力学(Computational fluid dynamics,CFD)求解Navier-Stokes方程虽能获得高精度压力分布,但网格离散求解流程计算开销大,开展大批量参数扫描时耗时极其突出.现有深度学习气动代理模型普遍存在前缘/后缘高梯度流场捕捉能力不足、训练收敛慢、容易出现数值振荡、跨翼型泛化性能有限等缺陷.针对上述问题,提出一种卷积神经网络(Convolutional neural network,CNN)与切比雪夫-柯尔莫哥洛夫-阿诺尔德网络(Chebyshev-Kolmogorov-Arnold network,Cheby-KAN)相融合的混合深度学习框架,实现二维翼型表面及流场压力分布的高精度快速预测.该框架采用残差CNN作为几何编码器,直接对翼型符号距离场(Signed distance field,SDF)网格进行特征学习,自动将复杂翼型几何转化为16维低维几何潜向量,不再依赖CST、NURBS等传统人工几何参数化手段.回归部分对KAN网络进行改进,采用平移切比雪夫多项式替代原始B样条作为边函数的基函数;借助平移切比雪夫多项式的正交性与等振荡逼近特性,有效抑制高次插值带来的Runge振荡现象,提升翼型前缘驻点、尾迹区等高梯度压力区域的数值稳定性.基于ANSYS Fluent结合SST k-ω湍流模型构建包含8 900万矢量样本点的大规模CFD数据集,并严格按算例层级划分数据集以杜绝数据泄露.对比实验结果表明:相较于同等任务下的传统多层感知机(Multi-layer perceptron,MLP),Cheby-KAN模块参数量下降98%,训练耗时减少60%,训练均方误差(Mean square error,MSE)降低92%.在完全未参与训练的翼型样本上开展零样本泛化测试,全场压力预测平均绝对误差(Mean absolute error,MAE)仅为2.1×10-3;由预测压力积分得到的升力系数与 CFD基准结果误差控制在 10%以内,具备良好的物理一致性.该模型能够快速输出流场压力系数场,可为翼型多目标气动优化提供高保真、低成本的数据驱动代理工具.

To address the high computational costs of traditional computational fluid dynamics(CFD)in aerodynamic optimization and the limitations of conventional deep learning models,such as slow convergence and insufficient accuracy in capturing complex flow physics,a novel framework integrating convolutional neural network(CNN)and Chebyshev-Kolmogorov-Arnold network(Cheby-KAN)is proposed for 2D airfoil pressure distribution prediction.The framework employs a CNN-based geometry encoder to automatically extract high-dimensional features from signed distance field(SDF)and map them into low-dimensional latent vectors,thereby bypassing traditional manual parameterization.In the regression module,shifted Chebyshev polynomials are introduced as basis functions within the KAN architecture(Cheby-KAN).Leveraging their orthogonality and superior approximation properties,Cheby-KAN effectively mitigates the Runge phenomenon and enhances numerical stability when resolving high-gradient pressure patterns.Experimental results on a dataset comprising 89 million data points demonstrate that Cheby-KAN achieves a 60%reduction in training time compared to conventional multi-layer perceptron(MLP),with a 92%decrease in training mean square error(MSE).Furthermore,the model exhibits exceptional generalization in"zero-shot"tests on unseen airfoils,maintaining a mean absolute error(MAE)of 2.1×10-3,which underscores its robustness as a high-fidelity surrogate model for advanced aerodynamic design.

陈耀鸿;张路驰;邓一菊;于彦泽;李响;焦仁山

中国航空工业空气动力研究院哈尔滨分院,哈尔滨 150001,中国中国燃气涡轮研究院,成都 610599,中国中国航空研究院研究生院,扬州 225111,中国中国航空研究院研究生院,扬州 225111,中国中国航空工业空气动力研究院沈阳分院,沈阳 110034,中国中国航空工业空气动力研究院哈尔滨分院,哈尔滨 150001,中国

航空航天

翼型压力分布深度学习卷积神经网络切比雪夫多项式柯尔莫哥洛夫-阿诺德网络气动预测

airfoil pressure distributiondeep learningconvolutional neural network (CNN)Chebyshev polynomialsKolmogorov-Arnold network (KAN)aerodynamic prediction

《南京航空航天大学学报(英文版)》 2026 (4)

592-613,22

This work was supported by the the Yangzhou Science and Technology Plan Project(No.YZ2025232).

10.16356/j.1005-1120.2026.04.008

评论