An implementation combining FermiNet with effective core potential (ecp). For paper, see https://arxiv.org/abs/2108.11661.

This repository directly depends on FermiNet (Many thanks to this awesome software and the team behind it!). Certain files are modified from the corresponding ones in FermiNet, and we added comments prefixed with "MODIFICATION FROM FERMINET" on the introduced changes.

pip install -e . will install all required dependencies.

Workflow of ferminet_ecp is similar to the original FermiNet, which uses the ConfigDict from ml_collections to configure the system. A few example scripts are included under ferminet_ecp/configs/ecp.

ferminet_ecp --config ferminet_ecp/configs/ecp/X.py:Ga,1 --config.batch_size 256 --config.pretrain.iterations 100

To use ECP for atoms, define ECP-related fields in the corresponding config file. For instance,

from pyscf import gto
from ferminet_ecp import base_config

def get_config(input_str):
    symbol, spin = input_str.split(',')
    cfg = base_config.default()
    mol = gto.Mole()
    # Set up molecule
    mol.build(
        atom=f'{symbol} 0 0 0',
        basis={symbol: 'ccecpccpvdz'},
        ecp={symbol: 'ccecp'},
        spin=int(spin))

    cfg.system.pyscf_mol = mol
    return cfg

Moreover, we want to mention that we remove local energy outlier in training phase via config.optim.rm_outlier flag, which violates variational principle and needs to be turned off in inference phase.

Our experiments were carried out with jax==0.2.12 and jaxlib==0.1.65+cuda102. We hit some cuda issues with cuda 11, especially when training with KFAC.

If you use this code in your work, please cite the associated paper.

@misc{li2021fermionic,
      title={Fermionic Neural Network with Effective Core Potential}, 
      author={Xiang Li and Cunwei Fan and Weiluo Ren and Ji Chen},
      year={2021},
      eprint={2108.11661},
      archivePrefix={arXiv},
      primaryClass={physics.chem-ph}
}