ASGCNN

This repository contains an implementation of the ASGCNN (Adsorbate-Site Graph Convolutional Neural Network) that predicts the adsorption energies with the help of classification tasks for adsorbate types and adsorption sites of slab structures.

Requirment

In parentheses is a version that is compatible after testing, current potential conflicts are from dgl and torch(torchdata).

• torch (2.1.0)
• torchdata (0.7.0)
• dgl (2.2.1)
• igraph (0.11.8)
• networkx
• scikit-learn
• pymatgen
• matplotlib
• tqdm
• numpy
• pandas
• qmpy_rester
• hyperopt

Overview

• ASGCNN/Encoder.py: Generate graph structure from VASP structure file and encode node and edge features.
• ASGCNN/Model.py: Pytorch implementation of the ASGCNN model.
• ASGCNN/Traniner.py: A module that calls the GNN model for training and prediction.
• data: Stores graph structures and targets for network training. Graphs are stored as .bin files in the dgl package.
• figures: Pictures drawn in Python in the article. Some of the drawings require custom Jworkflow scripts. Some code cannot run directly due to data size limitations.
• pretrained: Pretrained models. There are five models learned in an ensemble method, and they predict together to provide the uncertainty of the prediction results.
• structures: VASP structure files for calculation and graph structure generation.

Tutorials

• Query data: Query Heusler alloy data from OQMD: Tutorials_query_data.ipynb
• Batch construction of adsorption models and analysis of VASP results: This part is done through custom Jworkflow scripts
• Use the pre-trained model or train a new model from scratch: Tutorials_model_training.ipynb

Citation

If you are interested in our work, you can read our literature, and cite us using

@article{ZHOU2024160519,
title = {Machine-learning-accelerated screening of Heusler alloys for nitrogen reduction reaction with graph neural network},
journal = {Applied Surface Science},
volume = {669},
pages = {160519},
year = {2024},
issn = {0169-4332},
doi = {https://doi.org/10.1016/j.apsusc.2024.160519},
url = {https://www.sciencedirect.com/science/article/pii/S0169433224012327},
author = {Jing Zhou and Xiayong Chen and Xiao Jiang and Zean Tian and Wangyu Hu and Bowen Huang and Dingwang Yuan}