Multi-object aerodynamic design optimization using deep reinforcement learning

Xinyu Hui,Fei Qin,Wenqiang Li,Junqiang Bai,Hui Wang,Guoqiang He

doi:10.1063/5.0058088

Abstract

Aerodynamic design optimization is a key aspect in aircraft design. The further evolution of advanced aircraft derivatives requires a powerful optimization toolbox. Reinforcement learning (RL) is a powerful optimization tool but has rarely been utilized in the aerodynamic design. It can potentially obtain results similar to those of a human designer, by accumulating experience from training. In this work, a popular RL method called proximal policy optimization (PPO) is proposed to investigate multi-object aerodynamic design optimization. By observing the aerodynamic performances of different airfoils, the PPO updates a reasonable policy to generate the optimal airfoils in a single step. In a Pareto optimization problem with constraints, the PPO requires only 15% of the computational time of the non-dominated sorted genetic algorithm (II) to achieve the same accuracy. The results from testing show that the agent learns a policy that can achieve ∼4.3%–10.1% improvements of the aerodynamic performance compared with the results of baseline.

Highlights

In the field of aerodynamic design optimization, the requirements for advanced aircraft are rapidly increasing
The training includes updating the parameters of the policy and value network, and the network possesses the ability to provide the correct value of the current action at
It is shown that the Reinforcement learning (RL) can produce an optimal airfoil in a single step

Summary

Introduction

In the field of aerodynamic design optimization, the requirements for advanced aircraft are rapidly increasing. Gradient-free methods, such as genetic algorithms, are known to be effective at searching for global optimum solutions They face the curse of dimensionality as their computational cost is usually intolerable in high-dimensional problems, limiting the number of design variables.. Adjoint methods have proven to be accurate and efficient in computing gradients They can fall into local optima; they are highly sensitive to the starting point, and their efficiency is challenging in situations where the objective functions exhibit discontinuities or are strongly nonlinear. For both gradient-free and gradient-based methods, a surrogate model can be used to replace computational fluid dynamics (CFD) computations. These include radial basis functions, kriging, and artificial neural networks.

Methods

Results

Conclusion