High-throughput screening of the local environments of Single Atom Catalysts

This repository includes necessary scripts and models for reproducing "Active learning accelerated exploration of the single atom local environments in multimetallic systems for oxygen electrocatalysis". More information about the data, models and results can be found here. Dataset and trained model can be found in our Zenodo dataset here

Installation

1. Clone the repository

git clone https://github.com/learningmatter-mit/SingleAtom-LocalEnv.git

2. Conda environments setting and necessary package

• Conda environment

conda upgrade conda
conda env create -f environment.yml
conda activate SingeAtom

• Install NeuralForceField (nff)

Used for the dataset (graph) structure. Tested up to commit 72d1f32f43f202c1a466116beeed15845a6456e7 on the master branch.

git clone https://github.com/learningmatter-mit/NeuralForceField.git

Add the following to ~/.bashrc or equivalent with appropriate paths and then source ~/.bashrc.

export NFFDIR="/path/to/NeuralForceField"
export PYTHONPATH=$NFFDIR:$PYTHONPATH

4. Install Wandb

To monitor the training log more efficiently you can use wandb.

Create an account here and install the Python package:

pip install wandb
wandb login

5. Install persite_painn

# Go to SingleAtom-LocalEnv directory
pip install .
# For editable mode
pip install -e .

Usage

• Train the model:

  run main.py with config (example config files in examples/configs)

## In case you have converted dataset
python main.py --data_cache "path/to/dataset" --details "path/to/details" --savedir "path/to/savedir"
## In case you have raw dataset
python main.py --data_raw "path/to/dataset_raw" --data_cache "path/to/dataset" --details "path/to/details" --savedir "path/to/savedir"

• Active Learning Sampling

python active_learning.py --total_dataset "path/to/search_space" --model_path "path/to/trained_ensemble_models" --dataset "path/to/dataset" --multifidelity --uncertainty_type bayesian --save --plot

Examples

Some examples of ensemble inference and node embedding analysis are described in examples.