Skip to content

Code for performing adversarial attacks on atomistic systems using NN potentials

License

Notifications You must be signed in to change notification settings

learningmatter-mit/Atomistic-Adversarial-Attacks

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

33 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Atomistic Adversarial Attacks

DOI

Code for performing adversarial attacks on atomistic systems using NN potentials. The software was based on the paper "Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks", and implemented by Daniel Schwalbe-Koda and Aik Rui Tan.

The folder examples contains several Jupyter notebooks that illustrate the examples shown in the manuscript:

The folder data contains three datasets used in the paper: the DFT energies/forces of ammonia, OPLS energies/forces of alanine dipeptide, and zeolites occluded with neutral molecules, in the format readable by the Neural Force Field repo.

The full atomistic data is available at the Materials Cloud Archive on the link https://doi.org/10.24435/materialscloud:2w-6h.

Installation from source

This software was tested with PyTorch 1.4. The installation time highly depends on your internet connection and availability of a conda installation, but should not take more than an hour.

We recommend creating a conda environment to run the code. To do that, follow the setup instructions at the Neural Force Field repository.

conda upgrade conda
conda create -n nff python=3.7 scikit-learn pytorch=1.4.0 cudatoolkit=10.0 ase pandas pymatgen sympy rdkit hyperopt jq openbabel -c pytorch -c conda-forge -c rdkit -c openbabel

Then, install the remaining requirements using pip:

conda activate nff
pip install ipykernel nglview sigopt e3fp

To ensure that the nff environment is accessible through Jupyter, add the the nff display name:

python -m ipykernel install --user --name nff --display-name "nff"

Tutorials on how to use the NN potential

More tutorials are available on the Neural Force Field repository

Citing

The reference for the paper is the following:

@article{schwalbe2021differentiable,
  title={Differentiable sampling of molecular geometries with uncertainty-based adversarial attacks},
  author={Schwalbe-Koda, Daniel and Tan, Aik Rui and G{\'o}mez-Bombarelli, Rafael},
  journal={Nature Communications},
  volume={12},
  pages={5104},
  year={2021},
  publisher={Nature Publishing Group}
}

About

Code for performing adversarial attacks on atomistic systems using NN potentials

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages