FBFU-Net: Flow Branches Fusion U-Net for predicting flow fields around buildings

Urban wind effects, including wind-induced damage, can cause significant environmental disruptions, particularly in areas with high-rise buildings. Computational Fluid Dynamics (CFD) simulations offer a viable approach for assessing these impacts, though they face challenges in accuracy and computational efficiency. Recent advancements in machine learning provide opportunities to enhance these simulations. We propose a novel ML-based method, the Flow Branches Fusion U-Net (FBFU-Net), which builds upon the U-Net architecture utilizing Convolutional Neural Networks (CNNs). The network enhances predictions by leveraging distinct processing branches: one dedicated to velocity data and the other to pressure data. The integration of the Multimodal Transfer Module (MMTM), which fuses information from both branches, further improves the performance of FBFU-Net. The experimental results demonstrate the potential of the proposed method as an effective AI-driven tool for CFD-based wind analysis.