A CNN-PINN-DRL driven method for shape optimization of airfoils
Ying Liu, Jian Shen, Ping Yang, Xin Wen Yang
Abstract
Shape optimization of airfoils is crucial for enhancing the aerodynamic performance of large blades. Nowadays, the integration of computational fluid dynamics and intelligent optimization algorithm has become the dominant approach for airfoil shape optimization. However, this kind of method still faces the challenges of high-dimensional design space and high cost of performance evaluation. In this work, a novel approach was proposed to optimize the shape of airfoils and achieve a high lift-drag ratio. The approach integrates convolutional neural networks (CNNs), physics-informed neural networks (PINNs), and deep reinforcement learning (DRL) techniques. CNNs extract image features from airfoils and compress their shapes into six parameters. This significantly reduces the number of fitting parameters of airfoils and provides a low-dimensional design space. A PINN-based approach is utilized to evaluate the aerodynamic performance, addressing issues of collapse and non-convergence often encountered in the traditional Xfoil method. Deep reinforcement learning (DRL) is employed to integrate parameter dimensionality reduction and airfoil performance evaluation methods, identifying optimal solutions and facilitating algorithm transferability. The results demonstrate an enhanced lift-drag ratio for airfoils, and the proximal policy optimization (PPO) strategy improves the stability of the optimization algorithms.