A high-dimensional aerodynamic optimization framework for airfoils based on deep reinforcement learning

As the amount of airfoil optimization data and the complexity of models increase, the optimization intensity rises rapidly with the increase of optimization dimensions, imposing higher demands on optimization methods. Effectively finding the optimal solution in high-dimensional spaces has become a significant challenge. Therefore, this study innovatively establishes an optimization framework combining deep learning and reinforcement learning, effectively solving high dimensional optimization problems. Specifically, this framework proposes an intelligent optimization method that integrates Deep Neural Network (DNN) prediction model and Proximal Policy Optimization (PPO) algorithm, and tests it on the aerodynamic performance optimization problem of the airfoil NACA0012. It is compared with four traditional optimization algorithms: genetic algorithm, particle swarm optimization, ant colony algorithm, and simulated annealing, under the setting conditions of design parameter dimensions of 10, 50 and 100 respectively. The study shows that by comparing the optimization effects of different algorithms, the optimization framework based on deep reinforcement learning outperforms traditional optimization algorithms in airfoil examples from low to high dimensions, with optimization magnitudes of 16.02%, 50.08%, and 40% under the dimensions of 10, 50, and 100, respectively, and the impact of high dimensions is the smallest among the algorithms compared. The results indicate that the established deep reinforcement learning framework has good application prospects in high-dimensional optimization scenarios.