Swin-FlowNet: Flow field oriented optimization aided by a CNN and Swin-Transformer based model

Abstract

This paper presents a novel aerodynamic design optimization method for airfoils. The existing optimization method is transformed from being aerodynamic coefficient-oriented to flow field-oriented. By rapidly obtaining the flow fields of airfoils within the optimization domain and integrating them with aerodynamic theory, preoptimized airfoils can be obtained. To achieve this, we propose a hybrid network called Swin-FlowNet, which combines convolutional neural network (CNN) and Transformer. This network establishes a mapping between parameterized airfoils and their corresponding flow fields. The Transformer component captures long-range dependencies, while the CNN component extracts features from local information to characterize the flow field. We introduce a simple integration of convolution and attention mechanisms, resulting in the Swin-Conv Transformer Block (SCTB) that enhances local attention to improve channel and spatial attention. In our model, the multilayer perceptron (MLP) in the Transformer block is replaced with deep convolution, reducing the number of model parameters and enhancing feature extraction capability. Experimental results demonstrate that Swin-FlowNet achieves accurate and general performance in reconstructing and predicting the flow field around small variations of airfoils. Furthermore, Swin-FlowNet outperforms CNN-based and Transformer-based architectures in terms of accuracy and efficiency trade-offs.