Super Resolution Generative Adversarial Networks for Multi-Fidelity Pressure Distribution Prediction

Abstract

Airfoil pressure distributions (����) present crucial insights into aerodynamic performance. Moreover, achieving airfoils with well-performed ���� is an essential aspect of aerodynamic design optimizations. However, calculating the ���� by computational fluid dynamics is timeconsuming, while surrogate prediction requires a substantial amount of high-fidelity training samples especially when shocks exist. In this work, we propose a novel multi-fidelity modeling architecture through the super resolution generative adversarial networks (SRGAN). Compared with conventional multi-fidelity surrogates through addition and multiplication operations on low-fidelity data, SRGAN consists of two deep neural network models, i.e., a generator and a discriminator. In particular, the generator reads low-fidelity ���� and produces super-resolution ����, while the discriminator differentiates the super-resolution ���� from a high-fidelity data set. Thus, training the generator minimizes the difference between the super-resolution prediction and corresponding high-fidelity data, and in the meantime maximizes the similarity of super resolution with the high-fidelity training data set. Simultaneously, training the discriminator forces the improvement of generator by differentiating super resolution from the high-fidelity training data set. We demonstrated the SRGAN performance for multi-fidelity ���� prediction within the Mach range of [0.7, 0.75] and target lift-coefficient range of [0.75, 0.8], where strong shocks are likely to occur. Results presented that SRGAN predictions outperformed low-fidelity simulations and direct deep neural network predictions to the extend of capturing shock locations and strengths. In particular, SRGAN predictions achieved 98.5% relative generalization accuracy on the testing data set. This novel multi-fidelity surrogate modeling architecture is of great generality for other engineering applications.