Meshgraphnet_pytorch PyTorch implementation of MeshGraphNet (GNN) with three benchmark studies:

Deformed Flag	CFD	Deformed Plate

Dataset Download

Navigate to the dataset directory

cd path/to/dataset/directory
mkdir -p ${DATA}
bash meshgraphnets/download_dataset.sh flag_simple ${DATA}

Run the following command to generate the .idx file:

python -m tfrecord.tools.tfrecord2idx <file>.tfrecord <file>.idx

Excute

python deformedflag.py
python cfd.py
python deformedplate.py

Hyperparameter tune

Timestep > hidden dimension == Layer > epoch > batch

-Timestep: flag:5e-3, cfd: 1e-4

-hidden dimension > 64

-layer > 20

-epoch < 100

-batch ..

Predict Multifields

Velocity/pressure	displacement/stress

Graph attention model (GAT)

Single layer attention layer + encoder-processor-decoder

Ground True vs Prediction Vs Error	Attention propogation

References

1. https://github.com/google-deepmind/deepmind-research
2. https://medium.com/stanford-cs224w/learning-mesh-based-flow-simulations-on-graph-networks-44983679cf2d
3. https://github.com/wwMark/meshgraphnets