An efficient approach for predicting two-dimensional airfoil aerodynamic coefficients using progressive layered extraction and multimodal dynamic fusion

The rapid prediction of aerodynamic coefficients for two-dimensional airfoils is of great importance in airfoil design and optimization. Traditional computational fluid dynamics (CFD) methods, while highly accurate, are time-consuming and struggle to meet the real-time and high-accuracy requirements of iterative optimization in engineering applications. To address this challenge, this paper proposes a deep learning-based rapid prediction method specifically designed for predicting aerodynamic coefficients of two-dimensional multi-airfoils. The method takes airfoil images, physical characteristics, and operating condition parameters as inputs and outputs aerodynamic coefficients, including lift, drag, and pitching moment. By integrating multimodal dynamic fusion techniques with an improved Progressive Layered Extraction network architecture, an efficient deep learning model, Multitask Multimodal Fusion Network (MTMFNet), is constructed. Experimental results demonstrate that the proposed method reduces mean squared error loss by a factor of 10 compared to traditional deep learning methods. Furthermore, compared to CFD methods, MTMFNet reduces the computation time for predicting 20 operating conditions from several hours to just 1.2 s, which is approximately 10 500× faster. This significant speed improvement meets the real-time requirements of engineering applications while maintaining high prediction accuracy.