Airfoil Inverse Design Based on Sequential CST-cGANs

Airfoil inverse design is a method to determine the geometric shape of an airfoil based on known target aerodynamic characteristics. With the continuous development of deep learning, scholars have attempted to utilize Conditional Generative Adversarial Networks (cGANs) to assist or guide the inverse design process. This paper proposes a sequential CST-cGANs approach for the airfoil inverse design to enhance the design accuracy of traditional CST-cGANs methods. Traditional CSTcGANs, which are based on modeling a large amount of diverse airfoil data, exhibit good generalization capabilities but suffer from poor accuracy in specific airfoil inverse designs, making it difficult to meet geometric tolerances. Sequential CST-cGANs improve upon traditional methods by correcting the design results. The correction approach involves using the results of the traditional method as initial values for sampling, establishing a new cGANs model with a small sample size, and then conducting inverse design again. If the design result does not meet the requirements, the aforementioned sampling and modeling process is repeated, thereby forming a sequential modeling process. Experimental results indicate that as the number of sequences increases, the Mean Absolute Error (MAE) of the inverse design results gradually decreases, and geometric tolerances exhibit a trend of decreasing first and then fluctuating within a smaller range. The overall training time of the new method increases minimally, as subsequent sequences after the first one require only a small number of samples for training.