MANet: A Deep Learning Framework for Flow Field Prediction

Traditional methods for predicting airfoil flow fields primarily rely on computational fluid dynamics (CFD) simulations and wind tunnel experiments. However, solving the N avier-Stokes (NS) equations typically requires significant memory and computational cost, while wind tunnel experiments demand substantial human and financial resources. This presents a bottleneck for large-scale engineering designs, such as aerodynamic shape optimization. Deep learning models offer a highly efficient solution for rapid flow field simulation by significantly reducing costs while maintaining high accuracy. A novel deep learning model, which integrates convolutional attention mechanisms and multiple decoders into a modified U-N et network, has been proposed to further enhance model precision and lightweight performance. This model reduces parameter size by 75%, improves feature extraction capability, and mitigates the mutual influence between multiple prediction variables. Comparative experiments show that this approach reduces parameter size by about 42 %, mean absolute error (MAE) by 20%, and mean squared error (MSE) by 12% compared to the traditional U-Net network.