<p>Computational fluid dynamics (CFD) is a widely used method for calculating turbulent flow in continuous casting tundishes, but repeated CFD simulations are computationally expensive and cannot meet the demand for rapid screening of tundish structures. Thus, a physics-aware deep-learning surrogate, which integrates CFD with CNN- and U-Net-based architectures, is developed to obtain the single-phase flow field in a tundish with a dam under isothermal assumptions. The model has two input channels: a fluid region function and boundary distance function, and two output channels corresponding to the spatial distributions of two velocity components. The dataset is expanded by a logically extended integrated augmentation strategy. Comparative experiments were conducted among five architectures: CNN, UNet, MUNet, AUNet, and AMUNet. Quantitative results indicate that, while AMUNet achieves the lowest prediction error <i>u</i><sub>MAE</sub>&#xa0;=&#xa0;0.00208&#xa0;m/s and <i>v</i><sub>MAE</sub>&#xa0;=&#xa0;0.00220&#xa0;m/s, AUNet offers the optimal trade-off between accuracy and efficiency, achieving <i>u</i><sub>MAE</sub>&#xa0;=&#xa0;0.00244&#xa0;m/s and <i>v</i><sub>MAE</sub>&#xa0;=&#xa0;0.00219&#xa0;m/s with only 68&#xa0;pct of AMUNet’s training time (759.02&#xa0;seconds). The optimal configuration consists of a data scale of 200 augmented flow fields with random perturbation. The AUNet model reproduces the main vortex structures and jet stream calculated by OpenFOAM, and its inference time is less than 1&#xa0;seconds. Numerical results indicate that the combination of CFD and the deep-learning surrogate framework provides a feasible route for tundish-structure screening.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

AUNet-CFD: Rapid Surrogate Modeling for Isothermal Single-Phase Flow in a Tundish with a Dam

  • Yili Sun,
  • Hong Lei,
  • Changyou Ding,
  • Yuanxin Jiang,
  • Haoyu You,
  • Yan Zhao

摘要

Computational fluid dynamics (CFD) is a widely used method for calculating turbulent flow in continuous casting tundishes, but repeated CFD simulations are computationally expensive and cannot meet the demand for rapid screening of tundish structures. Thus, a physics-aware deep-learning surrogate, which integrates CFD with CNN- and U-Net-based architectures, is developed to obtain the single-phase flow field in a tundish with a dam under isothermal assumptions. The model has two input channels: a fluid region function and boundary distance function, and two output channels corresponding to the spatial distributions of two velocity components. The dataset is expanded by a logically extended integrated augmentation strategy. Comparative experiments were conducted among five architectures: CNN, UNet, MUNet, AUNet, and AMUNet. Quantitative results indicate that, while AMUNet achieves the lowest prediction error uMAE = 0.00208 m/s and vMAE = 0.00220 m/s, AUNet offers the optimal trade-off between accuracy and efficiency, achieving uMAE = 0.00244 m/s and vMAE = 0.00219 m/s with only 68 pct of AMUNet’s training time (759.02 seconds). The optimal configuration consists of a data scale of 200 augmented flow fields with random perturbation. The AUNet model reproduces the main vortex structures and jet stream calculated by OpenFOAM, and its inference time is less than 1 seconds. Numerical results indicate that the combination of CFD and the deep-learning surrogate framework provides a feasible route for tundish-structure screening.