A deep reinforcement learning optimization framework for supercritical airfoil aerodynamic shape design

In the context of traditional aerodynamic shape optimization design methods, the necessity to re-execute the complete optimization process when the initial shape changes poses significant challenges in engineering applications. These challenges encompass problems like data wastage and restricted ability for experience learning. We propose a policy learning-based optimization method that can automatically learn optimization experience through interactions with the environment. This optimization framework is based on deep reinforcement learning and consists of the policy learning process and the policy execution process. The action network, trained during the policy learning process, serves as a black box model of optimization experience and can directly and efficiently participate in guiding the actual optimization process. The optimization framework is validated through two-dimensional Rosenbrock function optimization, demonstrating its exceptional performance in achieving high-precision optimal solutions. Then, the effectiveness of this optimization method is demonstrated in the multi-point optimization design of supercritical airfoils, which aims to improve the buffet onset lift within predefined design constraints while maintaining the cruise lift-drag ratio. With the datum-coupled state format, the optimization experience can be tailored to the optimization requirements of different initial states during the learning process, leading to an optimization success rate in the optimization space that can exceed 90%.