A novel inverse design method for morphing airfoil based on deep reinforcement learning

Abstract

In this study, a reinforcement learning inverse design (RLID) framework is proposed and applied to the design of a morphing airfoil under variable operating conditions. The Latin hypercube sampling method is used to sample the design space of the airfoil parameters, and the corresponding aerodynamic parameters are obtained via CFD simulations. Then, a database is established to train the Kriging surrogate model, which is utilized to obtain the aerodynamic parameters of the airfoils during the design process. In response to the variable environments, the deep Q network (DQN) based on unsupervised learning is employed to morph the airfoil. The NACA 0012 airfoil is selected as the baseline airfoil, and the Hicks-Henne function method is adopted to allow the airfoil to morph autonomously in eight degrees of freedom. Finally, the strategy of the morphing airfoil which meets the requirements under different operating conditions is obtained via the DQN. Furthermore, the design results via the RLID framework are compared with those via the conditional generative adversarial network (CGAN) method. The results indicate that the proposed RLID framework obtains satisfactory strategy of morphing airfoil efficiently under variable operating conditions, and shows better performance than the CGAN method.