Adsorbate Chemical Environment - Graph Convolution Network (ACE-GCN)

This repository contains necessary scripts to implement Adsorbate Chemical Environment - Graph Convolution Neural Network (ACE-GCN): A graph neural network based model that uses graph convolutions to extract/transform/learn high-coverage adsorbate surface configurations for arbitrary number of surface adsorbates, binding sites, surface facets.

How to cite:

Please cite the following work if you want to use ACE-GCN:

Background:

This model is inspired by implementation of graph convolutions to model bulk mateirals and small molecules properties:

Overall scheme:

The idea is to interface ACE-GCN model with a systematic enumeration strategy to create relevant guess structures for high coverage surface adsorbate configurations. To do so, we use SurfGraph package.

The training methodology for setting up the screening workflow is as proposed:

Prerequisites:

This package requires:

If you are new to Python, the easiest way of installing the prerequisites is via conda. After installing conda, run the following commands to create a new environment named ace_gcn and install all prerequisites:

Download the Python 3+ version of Miniconda (recommended) or Anaconda for your OS.

NOTE: It is better to install miniconda vs full-fleged anaconda distribution. You can get the link for installation here

Follow the official instructions and install Miniconda/Anaconda.
Start the terminal (Mac/Linux) or Anaconda Prompt (Windows).
Create a virtual environment for ace_gcn:

conda create --name ace_gcn python=3.6

Activate the virtual environment.

conda activate ace_gcn

a. Install the necessary Python libraries.

conda install --yes numpy scipy matplotlib pandas jupyter jupyterlab seaborn scikit-learn
conda install --channel conda-forge --yes pymatgen

#Pytorch (CPU ONLY):
pip install torch==1.7.1 torchvision==0.8.2 

#Pytorch scatter (CPU ONLY)
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cpu.html #Pytorch scatter

b. To aid graph object generation it is recommended to use dask for parallelization:

conda install dask dask-jobqueue -c conda-forge

If for any reason you find that your conda environment is corrupted and you need to start over, please run:

conda env remove --name ace_gcn

and simply start from step 4 above to redo your setup

To start

Take a look at the Minimal Example folder to get started

NOTE: conda might have issues with installing certain modules. If you run into Solving Environment error check the conda version on the system. ACE-GCN was run on system having conda version 4.6.14.

Dataset

Data used in the Pt3Sn/NO example can be found in the here in *tar.bz format.

License

ACE-GCN is released under the MIT License

License

pghaneka/ACE_GCN

About

Resources

License

Stars

Watchers

Forks

Releases

Languages