Aerodynamic flow analysis using conditional convolutional autoencoder in various flow conditions and application to CFD-based design optimization

Abstract This study investigates the accuracy and efficiency of a convolutional autoencoder in predicting flow solutions of diverse characteristics, including strong local nonlinearity and unsteady wake vortices. Modifications to the standard U-net method were made suitable for non-Cartesian CFD mesh topology, enhancing solution accuracy. Additionally, conditions for predicting flows in unseen environments are integrated into a bottleneck layer between the encoder and decoder structures, guiding flow interpolation or extrapolation and parameter types. For direct comparison, this study uses a proper orthogonal decomposition (POD)-based ROM with linear reconstruction using dominant basis vectors from the flow solution space. Interpolation and extrapolation of generalized coordinates are performed using Gaussian process regression (GPR) and Long Short-Term Memory (LSTM) networks, respectively. The Conditional Unet (CUnet)’s accuracy is demonstrated through inviscid transonic airfoil flows, capturing shock waves effectively. Additionally, it can also be used for predicting the flow field of the three-dimensional shape of the Onera M6 wing. Vortex shedding flows around an Eppler airfoil at a 16-degree angle of attack in turbulent conditions were well-resolved, with root mean squared errors under 1% compared to full-order CFD results. Remarkably, the CUnet’s computational efficiency is highlighted as the wall clock CPU time for these 2D flows was less than one second. Finally, the ROM’s effectiveness is further validated through successful multi-point shape optimization, minimizing wave drag of RAE 2822 airfoils across subsonic to transonic conditions. This resulted in a maximum drag reduction of 37.38% at Mach 0.74 without performance degradation at off-design conditions.