A CFD-Based Surrogate Framework for Rapid 3D Flow-Field Reconstruction of Variable-Structure Tundishes

It is important to rapidly evaluate the flow behavior in a tundish during the design and optimization of flow control devices. However, repeated mesh generation and flow-field convergence for different structural schemes are time-consuming. To provide the optimized plan quickly, a CFD-based deep-learning surrogate framework is proposed for rapid three-dimensional (3D) flow-field reconstruction in variable-structure tundishes. In this framework, an Attention U-Net model is combined with a slice-wise dimensionality reduction strategy and spatial Euclidean distance encoding, and the original 3D reconstruction task is decomposed into batched 2D mapping subtasks. Furthermore, a physically constrained dynamic weighted loss function is introduced to improve the reconstruction of high-gradient flow features in the inlet and outlet regions. The results show that the Root Mean Square Error (RMSE) values of all reconstructed velocity components remain on the order of 10^-3 m/s at an inlet velocity of 0.5 m/s. For the <i>ω</i>=640 mm case, the relative error in the mean residence time between the OpenFOAM and AUNet velocity fields is only 0.88%. After offline training, the trained model can reconstruct a 3D flow field within 1 s. Thus, the proposed framework can serve as a CFD-based surrogate tool for rapid flow-field reconstruction and metallurgy-oriented screening of tundish configurations.