GemsDiff

Official implementation of GemsDiff. However, we realised after publication that our proposed model is more efficient so we provided the newest version without lattice diffusion. This repo contains dataset and checkpoints to retrain and generate materials.

All scripts can run on CPU and CUDA GPU but cuda is unsed by default. Please follow documentation to install pytorch and torch-geometric on CPU only.

Installation on a virtual environement

Create and activate the environement

python3 -m venv gemsdiff
source gemsdiff/bin/activate

Installing pytorch and torch geometric (see documentation: pytorch and torch geometric)

pip3 install torch
pip3 install torch_geometric
pip3 install torch_scatter

Install Crystallographic graph Crystallographic graph

pip3 install git+https://gitlab.com/chem-test/crystallographic-graph.git 

Install other dependancies

pip3 install torch-ema pandas tqdm matplotlib h5py pymatgen ase tensorboard

Sampling a specific composition

Sampling LiFeO2 from a checkpoint (OQMD)

python3 sampling.py LiFeO2 -c runs/without_cell_diffusion/training_2024_02_23_16_13_55_oqmd_604 -o LiFeO2.cif

Sampling structures of a given system from checkpoint

Sampling structure from the Li-Fe-O system from a checkpoint (Materials Project)

python3 sampling_system.py Li-Fe-O -c runs/without_cell_diffusion/training_2024_02_23_01_40_14_mp_110 -o Li-Fe-O.cif

How to cite

@article{klipfel_diff_aaai_2024,
    author={Astrid Klipfel and Ya{\"{e}}l Fr{\'{e}}gier and Adlane Sayede and Zied Bouraoui},
    title={Vector Field Oriented Diffusion Model for Crystal Material Generation},
    year={2024},
    journal={Proceedings of the AAAI Conference on Artificial Intelligence}
}