Application of deep learning based multi-fidelity surrogate model to robust aerodynamic design optimization

Abstract

In the present work, a multi-fidelity surrogate-based optimization framework is proposed, and then applied to the robust optimizations for airfoil and wing under uncertainty of Mach number. DBN (deep belief network) is employed as the low-fidelity model, and the k-step contrastive divergence algorithm is used for training the network. By virtue of the well trained DBN model and high-fidelity data, a linear regression multi-fidelity surrogate model is established. Verification results indicate that the multi-fidelity surrogate model obtains more accurate predictions than the DBN model and is highly reliable as a prediction model. The multi-fidelity surrogate model is embedded into an improved PSO (particle swarm optimization) algorithm framework, and is updated in each iteration of the robust optimization processes for both airfoil and wing. Comparisons between multi-fidelity surrogate predictions and CFD results indicate that, the multi-fidelity surrogate predictions tend to approach the CFD results as the iteration number increases. The robust optimization results of airfoil and wing demonstrate that, the multi-fidelity surrogate model performs very well as a prediction model, and improves the optimization efficiency obviously.