Hironobu-Kawaguchi/GNNforPhysics

Graph Neural Networks for Physics

GNNforPhysics

Graph Neural Networks for Physics

fork from https://github.com/deepmind/deepmind-research/tree/master/learning_to_simulate

# Download dataset (e.g. WaterRamps):
bash ./learning_to_simulate/download_dataset.sh WaterDrop ./tmp/datasets

{DATASET_NAME} one of the datasets following the naming used in the paper:

1. WaterDrop
2. Water
3. Sand
4. Goop
5. MultiMaterial
6. RandomFloor
7. WaterRamps
8. SandRamps
9. FluidShake
10. FluidShakeBox
11. Continuous
12. WaterDrop-XL
13. Water-3D
14. Sand-3D
15. Goop-3D

# run jupyter in docker:
docker-compose up

# Train a model:
# Generate some trajectory rollouts on the test set:
# Plot a trajectory:

# edit dcx/docker-compose.yaml
cd dc0
docker-compose up

# Check model log by tensorboard:
tensorboard --logdir tmp/models

tensorboard image

• WaterRamps(1000000step)
  • Example
  • GRAPHS
  • SCALARS loss