Abstract: Computational Fluid Dynamics (CFD) uses numerical simulation methods to solve the governing equations of flow, which can be subject to high computational cost and low efficiency. The U-net network, with its symmetrical encoder-decoder structure and skip connection, can capture both local features and global information in the input data, showing an excellent performance in flow field prediction. Based on this, this paper reviews the hot topics of U-net networks in convolutional neural network research, explains the principles of the U-net network architecture, and summarizes improvements of U-net model in fluid simulation applications from the perspectives of incorporating physical constraints, modeling temporal dependency, hybrid resolution modeling, introducing neural operators, and integrating residual network. Future development directions for U-net are also discussed. The results show that the introduction of physical constraints enables the model to improve prediction accuracy by optimizing the control equations even without labeled data. The temporal dependency modeling captures the sequential evolution of flow fields through latent space learning, and hybrid-resolution modeling can effectively handle complex flow fields, especially turbulence. The neural operators introduce Fourier transforms to enhance the model′s ability to capture global characteristics, and residual network integration improves model stability and training speed.