GAME-Net-UQ

This repository contains the Python code used to train and evaluate GAME-Net-UQ, a graph neural network with uncertainty quantification (UQ) for predicting the DFT energy of relaxed species and transition states adsorbed on monometallic transition metal surfaces.

Conda environment

We will soon provide a .yml file from which generate the conda environment needed for the code. Main dependencies are: Python 3.11, Pytorch, Pytorch Geometric, ASE.

DFT dataset

The DFT dataset fg.db (217 MB) used to train the GNN will be soon uploaded to Zenodo as ASE database including the DFT VASP relaxed geometries, simulation settings, and other metadata.

Graph dataset generation from ASE databases

The graph dataset (92 MB) can be automatically generated from the ASE database with the script gen_dataset.py. The same script can be used to generate your custom dataset from external ASE databases.

Model training and finetuning

To train the model, run the script train_mve.py. The input template file provides an explanation for each entry required in the training configuration file.

python train_mve.py -i input.toml -o output_dirname

Pretrained model

The final pretrained model is available within CARE (link).

License

The code is released under the MIT license.

Reference

• A Foundational Model for Reaction Networks on Metal Surfaces
  Authors: S. Morandi, O. Loveday, T. Renningholtz, S. Pablo-García, R. A. Vargas Hernáńdez, R. R. Seemakurthi, P. Sanz Berman, R. García-Muelas, A. Aspuru-Guzik, and N. López
  DOI: 10.26434/chemrxiv-2024-bfv3d