QHNet

QHNet (Quantum Hamiltonian Net)

This is the official implement of QHNet Paper Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian.

Installation

• clone this repo

• create the env and install the requirements

  $ git clone https://github.com/divelab/AIRS.git
  $ cd AIRS/OpenDFT/QHNet
  $ source ./install.sh

Reproduce

Please download the model parameters from QHNet_models - Google Drive. Then set the model path and version in the configs accroding to the names of the saved models.

Training

Training the QHNet on the MD17 dataset. Note that $DATASET is the name of datasets from water, ethanol, malondialdehyde, uracil.

python train_wH.py dataset=$DATASET model=QHNet_w_bias model.version=QHNet_w_bias

Training the QHNet on the mixed MD17 dataset.

python train_mixed.py dataset=all model=QHNet_w_bias model.version=QHNet_w_bias

Test

Testing the QHNet on the MD17 dataset. Note that $PATH_TO_SAVED_MODEL is the path to the downloaded model parameters or saved model paramters. The $Selected_version$ are in ['QHNet_wo_bias', 'QHNet_w_bias'] according to the model version name in the model filenames.

python test_wH.py dataset=$DATASET model=QHNet_w_bias model.version=$Selected_version model_path=$PATH_TO_SAVED_MODEL

Testing the QHNet on the mixed MD17 dataset.

python test_mixed.py dataset=all model=QHNet_w_bias model.version=$Selected_version model_path=$PATH_TO_SAVED_MODEL

Clarification about the model version - QHNet_w_bias suggested

For the model version, the QHNet_w_bias applies self-interaction layer with bias in the \ell=0 by setting bias=True in self-interaction layer and normalize the path using weights from the dictionary. For the QHNet_wo_bias, it does not apply such bias and path_normalization operations.

When conducting experiments for the paper, the model is developing and have minior setting difference for different experiments. In order to exactly reproduce the reported results, we provide the model with different versions. After our exploration, we clean our model and provide the suggested QHNet_w_bias for following experiments. We applogize for the inconvience.

Benchmark

The results of QHNet in the main paper is based on float64, the it applies a linear schedule of total 200, 000 steps with warmup 1,000.

Dataset	Training strategies	MAE	$\epsilon$	$\psi$
water	LSW (1,000, 200,000)	10.79	33.76	99.99
Ethanol	LSW (1,000, 200,000)	20.91	81.03	99.99
Malondialdehyde	LSW (1,000, 200,000)	21.52	82.12	99.92
Uracil	LSW (1,000, 200,000)	20.12	113.44	99.89

To facilitate the development of models to predict the Hamiltonian matrix, here we provide the results of QHNet on various total training steps for later on comparison. Note that the training procedure is based on float32, and the pretrained model QHNet version is QHNet_w_bias.

Dataset	Training strategies	MAE	$\epsilon$	$\psi$
water	RLROP	10.36	36.21	99.99
Ethanol	LSW (10,000, 1,000,000)	12.78	62.97	99.99
Ethanol	RLROP	13.12	51.80	99.99
Malondialdehyde	LSW (10,000, 1,000,000)	11.97	55.57	99.94
Malondialdehyde	RLROP	13.18	51.54	99.95
Uracil	LSW (10,000, 1,000,000	9.96	66.75	99.95

Citation

Please cite our paper if you find our paper useful.

@inproceedings{yu2023efficient,
  title={Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian},
  author={Yu, Haiyang and Xu, Zhao and Qian, Xiaofeng and Qian, Xiaoning and Ji, Shuiwang},
  booktitle={International Conference on Machine Learning},
  year={2023},
  organization={PMLR}
}

Acknowledgments

This work was supported in part by National Science Foundation grants CCF-1553281, IIS-1812641, OAC-1835690, DMR-2119103, DMR-2103842, IIS-1908220, IIS-2006861, and IIS-2212419, and National Institutes of Health grant U01AG070112. Acknowledgment is made to the donors of the American Chemical Society Petroleum Research Fund for partial support of this research.