This repository contains PyTorch implementations of meshgraphnets for flow around circular cylinder problem on the basic of PyG (pytorch geometric).

The original paper can be found as following:

Pfaff T, Fortunato M, Sanchez-Gonzalez A, et al. Learning mesh-based simulation with graph networks[J]. International Conference on Learning Representations (ICLR), 2021.

Some code of this repository refer to Differentiable Physics-informed Graph Networks.

• Jiang
• Zhang
• Chu
• Qian
• Li
• Wang

• h5py==3.6.0
• matplotlib==3.4.3
• numpy==1.21.1
• opencv_python==4.5.4.58
• Pillow==9.1.0
• torch==1.9.0+cu111
• torch_geometric==2.0.4
• torch_scatter==2.0.8
• tqdm==4.62.3

pip install -r requirements.txt

• Download cylinder_flow dataset using the script https://github.com/deepmind/deepmind-research/blob/master/meshgraphnets/download_dataset.sh.

• Parse the downloaded dataset into .h5 file using the tool parse_tfrecord.py

• Change the dataset_dir in train.py to your .h5 files.

• train the model by run python train.py.

• For test, run rollout.py, and the result pickle file will be saved at result folder, the you can run the render_results.py to generate result videos that can be saved at videos folder.

• Here are some examples, trained on cylinder_flow dataset.

• In addition, we use simulation software to generate new training data. The test results on our data are as following:

📧 [email protected]