Multi-fidelity convolutional neural network surrogate model for aerodynamic optimization based on transfer learning

Abstract

In aerodynamic shape optimization, a high-fidelity (HF) simulation is generally more accurate but more time-consuming than a low-fidelity (LF) simulation. To take advantage of both HF and LF simulations, a multi-fidelity convolutional neural network (CNN) surrogate model with transfer learning (MFCNN-TL) is proposed, which integrates different fidelity information through fine-tuning and adaptively learns their nonlinear mapping. The proposed surrogate model provides a new optimization framework, which maps the relation between shape parameters and aerodynamic performance. In the optimization framework, the HF model with a fine grid and the LF model with a coarse grid is used, respectively. In each optimization iteration, a multi-fidelity infilling strategy is adopted, and HF samples and LF samples are added to update the surrogate model. Finally, it is applied to the aerodynamic shape optimization of NACA0012 airfoil and RAE2822 airfoil. The optimization results show that the proposed MFCNN-TL surrogate model can significantly reduce the calculation cost and improve the optimization efficiency compared with the single-fidelity surrogate model.