cPaiNN

A new GNN model based on the PaiNN architecture and inspired by CHGNet, which we call charge-PaiNN (cPaiNN), that is capable of describing both the magnetic moments derived from Mulliken analysis and atomic charges obtained through Bader charge analysis based on the charge density.

To install cPaiNN use the commando, when in the repository;

pip install .

To install CHGNet use the commands:

pip install chgnet

To install M3GNet use the commands:

pip install m3gnet

To install Mace-MP-0 one need to use the commando:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Installing the universal machine learning potentials is only nessesary if you want to use them. They can simply be out comment in the simulation scripts

To train a cPaiNN model the path to the dataset needs to be added in the "config.toml" file. The dataset needs to be XYZ format, .traj format or any other format which can be read by the ase.io.read function. Tunning other parameters for the cPaiNN training are also done in the "config.toml" file. One can also parse the parameters directly in the training scripts as arguments. This will overwrite the parameters to in the "config.toml" file. To run the cPaiNN training use either the "run.sh" script or the commando:

python train.py --cfg config.toml

To train with atom charges represented using Bader charge analysis. One needs to save the bader charges into the Atom obects "array" dictionary. This can be done by writing "atom.array["bader_charge"] = bader_charges", where the bader charges can be extracted from the ACF.file using either 'ASE' or the function "attach_bader_charges" in extract_data/utils.py. An example is provided in the jupyter notebook "attach_bader_charge.ipynb"

The pre-trained cPaiNN model on the polyanaion sodium cathode dataset can be found along with the test dataset. The different name corresponds to the different properties it is trained on and the different hidden nodes and interaction layer used for the training.

cPaiNN can be used for different simualtion alone or along with other universal machine learning potentials (MLP). Adding new MLPs can easily be done as long as they have an ASE calcultator. Just add them in the "cPaiNN/relax.py" class object "ML_Relaxer.get_get_calc()" in the same way as the other MLPs are added. Four examples are given in this repository: One-hot calculation of atomic structures:

python One_hot_calculation.py

Structure optimization of atomic structures:

python Structure_relax.py

Molecular dynamic (MD) simulation of an atomic structure:

python MD_simulation.py

Nugded eleastic band (NEB) calculation of an ionic movement in an atomic structure:

python NEB.py

Along with the cPaiNN model we also included an active learning workflow called "autocPaiNN", where one can efficient and autonomously train a cPaiNN model for specific simualtion tasks. The active learning part is based on Curator and some additional scripts are included in the cPaiNN module to accomidate the use of active learning with cPaiNN. The github repository can be found here: https://github.com/dtu-energy/autocPaiNN

When using the code cite:

The code is modfied from older version of PaiNN, so also cite: https://doi.org/10.1038/s41524-022-00863-y and https://doi.org/10.26434/chemrxiv-2024-p5t3l