This is the codebase for the the paper SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models presented at ICLR 2025.

It is recommended to install each necessary package (with appropriate versions mentioned in the symmcd.yaml) file following this order: pytorch, pytorch-lightning, pyg, pyxtal, pymatgen, matminer, einops, hydra-core, symd, dotenv, wandb,p_tqdm,torch_scatter, torch_sparse, smact,chemparse.

1. Rename the .env.template file into .env and specify the following variables.

PROJECT_ROOT: the absolute path of this repo
HYDRA_JOBS: the absolute path to save hydra outputs
WABDB_DIR: the absolute path to save wandb outputs
WABDB_DIR: the absolute path to save wandb cache outputs

2. Create the anaconda environment with the symmcd.yml file (conda env create -f symmcd.yml: creates environment with name symmcd)
3. Install matminer with cd matminer && pip install -e .
4. Clone cdvae repo in the same directory level as conf and data and install it
5. You can now execute the training command below without errors. If it throws some package error, please install those (and create a PR).

SymmCD (should now look like this, and the environment should contain cdvae and matminer packages)
├── cdvae
├── conf
├── data 
├── symmcd
├── matminer
├── scripts
├── .env
├── .........

Before training, you should change the following in cdvae/cdvae/pl_modules/gnn.py (swish can no longer be imported from torch_geometric.nn.acts):

• Comment out the import: from torch_geometric.nn.acts import swish
• Add the following lines to the script before the class InteractionPPBlock:

def swish(x):
    return x * x.sigmoid()

python symmcd/run.py data=<dataset> model=discrete_diffusion_w_site_symm expname=<expname>

• The <dataset> tag can be selected from mp_20 and mpts_52.
• For multiple GPUs, please add train.pl_trainer.devices=2 to above commands (ensure 2 gpus on machine where script launches).

python scripts/generation.py --model_path <model_path> --dataset <dataset>
python scripts/compute_metrics --root_path <model_path> --tasks gen --gt_file data/<dataset>/test.csv

python scripts/sample.py --model_path <model_path> --save_path <save_path> --formula <formula> --num_evals <num_evals>

• Change/Add hyperparameters and their values in the hyperparam_sweep.yaml file.
• wandb sweep --project <project-name> -e <entity-name> hyperparam_sweep.yaml.
• wandb agent <above-agent-id>.

If you find this codebase useful, please cite the following paper:

@inproceedings{levysymmcd,
  title={SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models},
  author={Levy, Daniel and Panigrahi, Siba Smarak and Kaba, S{\'e}kou-Oumar and Zhu, Qiang and Lee, Kin Long Kelvin and Galkin, Mikhail and Miret, Santiago and Ravanbakhsh, Siamak},
  booktitle={The Thirteenth International Conference on Learning Representations}
}

This codebase was adapted from the DiffCSP repository. We thank the authors for their work and open-sourcing their code.