FullereneNet

Extrapolating Beyond C60: Advancing Prediction of Fullerene Isomers with FullereneNet

Model

Feature

Dependency

pip install torch
pip install torch-geometric
pip install rdkit
pip install scikit-learn
pip install pandas matplotlib
pip install tqdm 

Feature Generation

Before training, generate the features by running:

python feature/generate_all_features.py

Data Split Strategy

We provide several data split strategies for training and evaluation:

1. Random split

   Specify the seed when running training:

   python train_and_test_FullereneNet.py --seed 0

2. Random split

   See split_method/cross_validation_split.py.

3. Leave-one-group-out split

   See split_method/leave_one_group_out_split.py.

4. Leave-one-cluster-out split

   See split_method/leave_one_cluster_out_split.py.

Training and Testing

FullereneNet

Train and test the model using:

python train_and_test_FullereneNet.py --with_edge --epochs 300 --target 'Eb' --lr 0.001 -seed 42

Target choices: 'homo', 'lumo', 'gap', 'dipole_total', 'G-water', 'G-dich', 'logP', 'Eb'. The explanation of each target can be found on paper.

--with_edge indicate using both node and edge feature to train the model.

Training and testing results will be printed to the console. All available arguments can be found inside the script.

Matformer

We also benchmark with Matformer, using bond distances as features. These features should be obtained from optimized structures and can be downloaded from Figsure, find feature_for_matformer.zip file, download it. Then move the file to feature/opt_and_unopt_for_matformer folder, unzip file.

Run training and testing with:

python train_and_test_FullereneNet.py --epochs 300 --target 'Eb' --lr 0.001 --seed 42 --use_optimized_structure

--use_optimized_structure indicate using DFT optimized structures to train the model.

Molecule Dynamic with GAP-20

We provide example LAMMPS input scripts using the GAP-20 potential to calculate binding energy and perform geometry optimization on GAP-20 folder.