Multi-fidelity deep neural network surrogate model for aerodynamic shape prediction based on multi-task learning

With the increasing demands of refined aerodynamic shape design for modern aircraft, it is very essential for aerodynamic shape optimization to obtain more accurate aerodynamic data. The widely used high-fidelity simulation is usually accurate but extremely time-consuming. Therefore, we propose an innovative multi-fidelity model based on multi-task learning for aerodynamic shape prediction. This model consists of Bézier-Auxiliary Classifier GAN (Bézier-ACGAN), Multi-gate Mixture-of-Experts and Multi-Fidelity (MMoE-MF). Firstly, Bézier-ACGAN is used to construct the subsonic and transonic datasets and is used as an intelligent parameterization method. Secondly, The MMOE-MF model is coupled with the parameters of Bezier-ACGAN to predict different-fidelity of aerodynamic data. The results show that the predicted results of the optimal airfoils agree with the results of high-fidelity simulation well. This method is a promising approach that can convert from low-fidelity data to high-fidelity data in a few seconds.