Abstract:Computational Fluid Dynamics (CFD) simulation by the numerical solution of the Navier-Stokes equations is an essential tool in a wide range of applications from engineering design to climate modeling. However, the computational cost and memory demand required by CFD codes may become very high for flows of practical interest, such as in aerodynamic shape optimization. This expense is associated with the complexity of the fluid flow governing equations, which include non-linear partial derivative terms that are of difficult solution, leading to long computational times and limiting the number of hypotheses that can be tested during the process of iterative design. Therefore, we propose DeepCFD: a convolutional neural network (CNN) based model that efficiently approximates solutions for the problem of non-uniform steady laminar flows. The proposed model is able to learn complete solutions of the Navier-Stokes equations, for both velocity and pressure fields, directly from ground-truth data generated using a state-of-the-art CFD code. Using DeepCFD, we found a speedup of up to 3 orders of magnitude compared to the standard CFD approach at a cost of low error rates.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the high computational cost and large memory requirements in computational fluid dynamics (CFD) simulations. Specifically, traditional CFD methods simulate fluid flow by numerically solving the Navier - Stokes equations. However, in practical applications, such as aerodynamic shape optimization, the computational time and resource consumption of this method are very high. This is mainly because the Navier - Stokes equations contain non - linear partial derivative terms, which lead to a complex and time - consuming solution process and limit the number of hypotheses that can be tested during the design iteration process.
To solve this problem, the authors propose DeepCFD, a model based on convolutional neural networks (CNN) that can efficiently approximate solutions for non - uniform steady laminar flow. DeepCFD can directly learn the complete solutions of the velocity field and pressure field from real - data generated using state - of - the - art CFD codes. Experimental results show that, compared with standard CFD methods, DeepCFD can achieve a speed - up of up to three orders of magnitude while maintaining a low error rate.
### Main contributions of the paper:
1. **Propose a CNN - based CFD surrogate model**: This model is applicable to 2D non - uniform steady laminar flow and can provide solutions that are three orders of magnitude faster than traditional CFD methods with a low error rate.
2. **Expand previous research**: Compared with the work of Guo et al. [8], DeepCFD provides complete solutions for all components of the velocity field and pressure field, which is crucial for product development in engineering design.
3. **Release the code and data set**: The authors provide the code and data set used for training and validating the model to promote further research and development in this field.
### Method overview:
- **Traditional CFD method**: Use the simpleFoam solver in the OpenFOAM framework to solve the Navier - Stokes equations by numerical methods.
- **Machine learning method**: Use convolutional neural networks (CNN) to build a surrogate model. The inputs include geometric information (such as obstacle shapes) and boundary conditions, and the outputs are the velocity field and pressure field.
### Experimental setup:
- **Data generation**: Use OpenFOAM to generate a data set of more than 1,000 2D steady laminar flows around randomly - shaped obstacles.
- **Model training**: Use the AdamW optimizer, with a batch size of 64 and a weight decay of 0.005, and train on an Nvidia Tesla V100 SXM2 GPU.
- **Loss function**: Use the mean squared error (MSE) for velocity components and the mean absolute error (MAE) for pressure.
### Results analysis:
- **Model optimization**: Through hyperparameter search, the best model configuration, including the learning rate, convolution kernel size, and number of convolution layers, is selected.
- **Performance comparison**: The DeepCFD model outperforms the baseline model in terms of the test - set mean squared error (MSE) for all variables (velocity components Ux, Uy, and pressure p).
- **Prediction time**: Compared with traditional CFD methods, the prediction time of DeepCFD on CPUs and GPUs is improved by 1 to 5 orders of magnitude respectively.
### Conclusion:
DeepCFD provides an efficient and accurate method to approximate CFD simulation results, especially suitable for engineering applications that require a large number of iterative designs. Future work can further explore how to combine this data - driven method with physics - driven techniques to improve the generalization ability and robustness of the model.