Code to paper: EGRAFFBENCH: EVALUATION OF EQUIVARIANT GRAPH NEURAL NETWORK FORCE FIELDS FOR ATOMISTIC SIMULATIONS

Setup environment

see : env

Download example data and model weights(~450 MB): folder example_mdbenchgnn

1. cd ../
2. pip install gdown (install gdown to download data from google drive)
3. gdown --folder --id 1PrrKaMBbjMyt3DrXM94XQ-X48nVrf-kY

create few symbolic-links (symbolic-links are great way to save space and avoid copying data)

1. cd mdbenchgnn/
2. ln -s ../example_mdbenchgnn example
3. ln -s path/to/output_dir output_dir_sl
4. ln -s path/to/data_dir data_ls

Models supported

1. Nequip

2. Allegro

3. Mace

4. Botnet

5. Equiformer

6. TorchMDnet

Try on example data: example/lips/data

1. to train the models go through: scripts/train
2. to run a MD simulation on trained model: scripts/md_simulation

Run on custom datasets:

see : preprocessing

Postprocess

1. Radial Distribution function scripts/struct_props

Acknowledgement

Our implementation is based on PyTorch, PyG, nequip, allegro, equiformer, mace, TorchMD-NET and MDsim