Optimization of supersonic inlet configuration using reinforcement learning coupled with aerodynamic performance surrogate model

The neural network-based surrogate model (NNSM) can establish the relationship between the geometric and aerodynamic parameters of an aircraft or its components, thus enabling rapid aerodynamic performance evaluation. In this study, this framework is employed for the aerodynamic performance evaluation and configuration optimization of curved compression inlet. The inlet is parametrically represented, where only three parameters can control its configuration including the external compression surface, the internal compression surface and the cowl. A neural network model is established and trained on dataset generated by numerical simulation, which takes the above geometric parameters as inputs and the pressure rise coefficient, mass flow rate coefficient and total pressure recovery coefficient at the inlet throat as outputs. For all the samples in the validation set, the relative errors upper bounds of the well-trained model for the mass flow rate coefficient and total pressure recovery coefficient of inlet are <0.01, while for the pressure rise coefficient it is <0.025. The model can accurately distinguish between critical and subcritical states by means of the mass flow rate coefficient. Moreover, the effects of wave systems, separation bubbles in the flowfield and even the weak compression waves generated by the curved cowl on the pressure rise coefficient and total pressure recovery coefficient can be accurately captured. With NNSM as the environment, deep reinforcement learning based on Advantage Actor-Critic algorithm is applied to optimize the inlet configuration. Taking the maximization of the total pressure recovery coefficient as the objective, a hyperparameter scheme with better convergence is determined. Then, under the constraints of a series of specific pressure rise coefficients, the task of maximizing the total pressure recovery coefficient is performed. This optimization framework is validated effective, with favorable convergence and on-demand inlet geometry generation.