Conditional U-Net을 이용한 다중 유동 조건에서의 효율적인 다점 형상 최적 설계 기법

Abstract

In this study, Conditional U-Net model was used to predict the transonic flow of various shapes and flow conditions which shows large non-linearity, and compared with the POD-GPR model, a linear order reduction technique. In addition, multi-point shape optimization was performed based on the validated model. Both models showed an average error of around 1% for test data, but the accuracy of the Conditional U-Net prevailed for data with large non-linearity. It has also been demonstrated that applying gradient difference loss function is effective in improving accuracy of Conditional U-Net. When multi-point shape optimization was performed using the SL U-Net, the drag reduction at M=0.74 was 47.3%, a relatively small amount compared to the single-point shape optimization. But the performance of airfoil improved in most Mach number ranges. In terms of computational cost, the time required to predict one flow field using Conditional U-Net was measured to be about 0.05% of the FOM CFD. Therefore, it was found that the AI-based model developed in this study has sufficient accuracy and computational efficiency, and also is effectively applicable to multi-point optimization problems.