FIRE-GNN – Force-Informed, Relaxed Equivariance Graph Neural Network

FIRE-GNN is a force-informed, relaxed equivariant graph neural network for predicting surface work functions and cleavage energies from slab structures. By incorporating surface-normal symmetry breaking and machine learning interatomic potential–derived force information, the approach achieves state-of-the-art accuracy and enables rapid, generalizable screening critical for the discovery of materials with tuned surface properties across the periodic table.

Please, cite the following paper if you use the architecture/model in your research:

@article{Hsu2025Dec,
	author = {Hsu, Circe and Schlesinger, Claire and Mudaliar, Karan and Leung, Jordan and Walters, Robin and Schindler, Peter},
	title = {{FIRE-GNN: Force-informed, Relaxed Equivariance Graph Neural Network for Rapid and Accurate Prediction of Surface Properties}},
	journal = {Advanced Intelligent Discovery},
	volume = {0},
	number = {0},
	pages = {0},
	year = {2025},
	month = dec,
	publisher = {Wiley},
	doi = {10.1002/aidi.202500162}
}

Installation

To install via pip, you can create your own pip environment or use your global environment and install the requirements.txt file.

cd FIRE
python -m venv firegnn
source firegnn/bin/activate
pip install -r requirements.txt

To install via micromamba (or any conda)

cd FIRE
micromamba env create -f environment.yml
micromamba activate firegnn

Usage

In order to run FIRE-GNN, set up your slabs as a cif file, folder of cif files, or json list of cif strings. Run

python run_custom.py FIRE-GNN-Model_structureid_split configs/struct_test_forces.json slabs.json