本研究建立一套基於深度學習之代理模型，用以預測二維雙機翼構型之流場分布與氣動性能。雙機翼構型可在有限結構空間內增加有效升力面積，因此適合應用於小型無人飛行載具與微型飛行器。然而，其氣動性能會受到前、後翼之間非線性交互作用影響，例如尾流效應、上洗與下洗流場，使得整體壓力分布、速度場、升力與阻力特性產生複雜變化。因此，若使用傳統計算流體力學方法評估大量雙機翼構型，將需要較高的計算成本。
本研究以 OpenFOAM 建立二維雙機翼 CFD 資料庫，包含 29 種 UIUC 資料庫翼型、自由流速度 U_∞=5∼25 m/s、攻角 α=-6°∼10°，以及前翼相對後翼之水平與垂直位置變化。翼弦長固定為 0.2 m，計算域為 2 m × 1 m。模擬所得之壓力場、x 方向速度場、y 方向速度場、升力係數與阻力係數，作為 U-Net 卷積神經網路之深度學習訓練資料；模型輸入包含雙機翼形狀及排列幾何、來流條件與相對位置參數，輸出則包含流場分布與氣動係數。
結果顯示，模型可重建雙機翼周圍之主要流場結構。在驗證資料集中，壓力場、x 方向速度場與 y 方向速度場之均方根誤差分別為 22.670、0.550 與 0.203，對應相對誤差為 0.01%、2.68% 與 0.92%；升力係數、阻力係數與升阻比之決定係數（Coefficient of Determination, R^2）分別為 0.954、0.950 與 0.949。低速、高攻角及特定下方排列之誤差較高。
整體而言，本研究所建構之 U-Net 模型可作為訓練參數域內二維雙機翼氣動分析與構型初篩之快速代理工具。然而，模型以穩態二維 RANS 資料為基礎，尚未涵蓋有限展弦比、翼尖渦、展向流動及非定常三維分離等效應；應用於實際飛行器設計前，仍須進行三維 CFD 或實驗驗證。

This study develops a U-Net-based surrogate for predicting two-dimensional tandem-airfoil flow fields and aerodynamic coefficients. A database of 3,387 steady RANS cases was generated in OpenFOAM using 29 airfoil profiles from the UIUC database. The freestream velocity is U_∞=5∼25 m/s, the angle of attack is α=-6°∼10°, the chord length of the tandem airfoils is fixed at 0.2 m, and the computational domain is 2 m × 1 m. Model inputs comprise airfoil geometries including airfoil profiles and tandem arrangements, freestream conditions, and relative-position parameters; outputs comprise pressure, x- and y-velocity fields, lift coefficient, and drag coefficient. On the validation set, the three flow fields achieved root-mean-square errors of 22.670, 0.550, and 0.203, with relative errors of 0.01%, 2.68%, and 0.92%. The coefficients of determination R2 for lift coefficient, drag coefficient, and lift-to-drag ratio were 0.954, 0.950, and 0.949. Fifteen independent cases showed that the model captured the principal effects of geometry and vertical arrangement, while errors increased in low-speed, high-angle-of-attack conditions. The model is therefore suitable for rapid preliminary screening within the training domain; its steady two-dimensional RANS basis requires additional three-dimensional validation before aircraft-level application.

[1] R. Jones, D. J. Cleaver, and I. Gursul, "Aerodynamics of biplane and tandem wings at low Reynolds numbers," Experiments in Fluids, vol. 56, no. 6, pp. 1-25, 2015.
[2] N. Hosseini, M. Tadjfar, and A. Abba, "Configuration optimization of two tandem airfoils at low Reynolds numbers," Applied Mathematical Modelling, vol. 102, pp. 828-846, 2022.
[3] Z. Cai, Z. Yu, and H. Niu, "Aerodynamic optimization design of tandem-wing configurations based on multi-fidelity deep neural networks," SSRN, Preprint, 2025.
[4] X. Hui, J. Bai, H. Wang, and Y. Zhang, "Fast pressure distribution prediction of airfoils using deep learning," Aerospace Science and Technology, vol. 105, p. 105949, 2020.
[5] C. Duru, H. Alemdar, and Ö. U. Baran, "CNNFOIL: convolutional encoder decoder modeling for pressure fields around airfoils," Neural Computing and Applications, vol. 33, pp. 6835-6849, 2021.
[6] O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional networks for biomedical image segmentation," Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, vol. 9351, pp. 234-241, 2015.
[7] N. Thuerey, K. Weissenow, L. Prantl, and X. Hu, "Deep learning methods for Reynolds-averaged Navier-Stokes simulations of airfoil flows," AIAA Journal, vol. 58, no. 1, pp. 25-36, 2020.
[8] T. M. Faure, L. Hétru, and O. Montagnier, “Aerodynamic features of a two-airfoil arrangement,” Experiments in Fluids, vol. 58, no. 10, Art. no. 146, 2017.
[9] M. Tripathi, "Facial image denoising using AutoEncoder and UNET," Heritage and Sustainable Development, vol. 3, no. 2, pp. 89-96, 2021.
[10] V. Dumoulin and F. Visin, “A guide to convolution arithmetic for deep learning,” arXiv preprint arXiv:1603.07285, 2016.
[11] S. Cai, Y. Wu, and G. Chen, "A novel elastomeric UNet for medical image segmentation," Frontiers in Aging Neuroscience, vol. 14, p. 841297, 2022.