Multi-Viscosity Physics-Informed Neural Networks ($\mu$-PINNs)

This is a project about the article "Multi-Viscosity Physics-Informed Neural Networks for Generating Ultra High Resolution Flow Field Data".

The project provides code and associated datasets to generate ultra-high resolution flow field data. You can use the provided code to generate training points, train a model, and make flow field predictions. We also provide specific examples using OpenFOAM where viscosity values can be modified to obtain corresponding numerical simulation results.

File Structure

data/ folder: Contains the training points used during training and the data generated using OpenFOAM.
get_point_train/ folder: Contains code for generating training points.
model/ folder: Contains the trained model.
openfoam/ folder: Contains specific examples using OpenFOAM.

Usage

Prediction

Run the following command to make a prediction:

python mupinns_trad.py --state predict --scenario cylinder --mu 1e-2 --load_model_dir ./model/model_trad_cylinder.pth

Using OpenFOAM

After entering the OpenFOAM environment, modify the viscosity value in the transportProperties (or physicalProperties) file.
Run the Allrun script to obtain the numerical simulation results for the corresponding viscosity value, which will be saved in the VTK folder.
Use Paraview for visualization and to export the results as a CSV file.

Citation

If you find this project helpful, please cite our paper as follows:

@article{zhang2023multi,
  title={Multi-Viscosity Physics-Informed Neural Networks for Generating Ultra High Resolution Flow Field Data},
  author={Zhang, Sen and Guo, Xiao-Wei and Li, Chao and Zhao, Ran and Yang, Canqun and Wang, Wei and Zhong, Yanxu},
  journal={International Journal of Computational Fluid Dynamics},
  pages={1--19},
  year={2023},
  publisher={Taylor \& Francis}
}

Contact Information

If you have any questions or suggestions, feel free to contact the project authors.