DeepH-pack is the official implementation of the DeepH (Deep Hamiltonian) method described in the paper Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation and in the Research Briefing.
DeepH-pack supports DFT results made by ABACUS, OpenMX, FHI-aims or SIESTA and will support HONPAS soon.
For more information, see the documentation and the talk (in Chinese).
@article{deeph,
author = {Li, He and Wang, Zun and Zou, Nianlong and Ye, Meng and Xu, Runzhang and Gong, Xiaoxun and Duan, Wenhui and Xu, Yong},
title = {Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation},
journal = {Nature Computational Science},
volume = {2},
number = {6},
pages = {367-377},
ISSN = {2662-8457},
DOI = {10.1038/s43588-022-00265-6},
url = {https://doi.org/10.1038/s43588-022-00265-6},
year = {2022},
type = {Journal Article}
}
@article{deephe3,
author = {Gong, Xiaoxun and Li, He and Zou, Nianlong and Xu, Runzhang and Duan, Wenhui and Xu, Yong},
title = {General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian},
journal = {Nature Communications},
volume = {14},
number = {1},
pages = {2848},
ISSN = {2041-1723},
DOI = {10.1038/s41467-023-38468-8},
url = {https://doi.org/10.1038/s41467-023-38468-8},
year = {2023},
type = {Journal Article}
}
@article{xdeeph,
author = {Li, He and Tang, Zechen and Gong, Xiaoxun and Zou, Nianlong and Duan, Wenhui and Xu, Yong},
title = {Deep-learning electronic-structure calculation of magnetic superstructures},
journal = {Nature Computational Science},
volume = {3},
number = {4},
pages = {321-327},
ISSN = {2662-8457},
DOI = {10.1038/s43588-023-00424-3},
url = {https://doi.org/10.1038/s43588-023-00424-3},
year = {2023},
type = {Journal Article}
}
To use DeepH-pack, following environments and packages are required:
Prepare the Python 3.9 interpreter. Install the following Python packages required:
- NumPy
- SciPy
- PyTorch = 1.9.1
- PyTorch Geometric = 1.7.2
- e3nn = 0.3.5
- pymatgen
- h5py
- TensorBoard
- pathos
- psutil
In Linux, you can quickly achieve the requirements by running
# install miniconda with python 3.9
wget https://repo.anaconda.com/miniconda/Miniconda3-py39_4.10.3-Linux-x86_64.sh
bash Miniconda3-py39_4.10.3-Linux-x86_64.sh
# install packages by conda
conda install numpy
conda install scipy
conda install pytorch==1.9.1 ${pytorch_config}
conda install pytorch-geometric=1.7.2 -c rusty1s -c conda-forge
conda install pymatgen -c conda-forge
# install packages by pip
pip install e3nn==0.3.5
pip install h5py
pip install tensorboard
pip install pathos
pip install psutil
with ${pytorch_config}
replaced by your own configuration.
You can find how to set it in the official website of PyTorch.
Prepare the Julia 1.6.6 interpreter. Install the following Julia packages required with Julia's builtin package manager:
- Arpack.jl
- HDF5.jl
- ArgParse.jl
- JLD.jl
- JSON.jl
- IterativeSolvers.jl
- DelimitedFiles.jl
- StaticArrays.jl
- LinearMaps.jl
- Pardiso.jl
In Linux, you can quickly achieve the requirements by first running
# install julia 1.6.6
wget https://julialang-s3.julialang.org/bin/linux/x64/1.6/julia-1.6.6-linux-x86_64.tar.gz
tar xzvf julia-1.6.6-linux-x86_64.tar.gz
# open the julia REPL
julia
Then enter the pkg REPL by pressing ]
from the Julia REPL. In the pkg REPL run
(@v1.6) pkg> add Arpack
(@v1.6) pkg> add HDF5
(@v1.6) pkg> add ArgParse
(@v1.6) pkg> add JLD
(@v1.6) pkg> add JSON
(@v1.6) pkg> add IterativeSolvers
(@v1.6) pkg> add DelimitedFiles
(@v1.6) pkg> add StaticArrays
(@v1.6) pkg> add LinearMaps
Follow these instructions to install Pardiso.jl.
One of the supported DFT packages is required to obtain the dataset and calculate the overlap matrix for large-scale material systems. DeepH-pack supports DFT results made by ABACUS, OpenMX, FHI-aims or SIESTA and will support HONPAS soon.
- ABACUS: Install ABACUS package for density functional theory Hamiltonian matrix calculation to construct datasets. DeepH-pack requires ABACUS version >= 2.3.2.
- OpenMX:
- Install OpenMX package version 3.9 for density functional theory Hamiltonian matrix calculation to construct datasets.
If you are using Intel MKL and Intel MPI environments, you can use the following variable definitions for makefile
Or edit the makefile yourself according to your environment to install OpenMX version 3.9.
CC = mpiicc -O3 -xHOST -ip -no-prec-div -qopenmp -I${MKLROOT}/include/fftw -I${MKLROOT}/include FC = mpiifort -O3 -xHOST -ip -no-prec-div -qopenmp -I${MKLROOT}/include LIB = ${CMPLR_ROOT}/linux/compiler/lib/intel64_lin/libiomp5.a ${MKLROOT}/lib/intel64/libmkl_blas95_lp64.a ${MKLROOT}/lib/intel64/libmkl_lapack95_lp64.a ${MKLROOT}/lib/intel64/libmkl_scalapack_lp64.a -Wl,--start-group ${MKLROOT}/lib/intel64/libmkl_intel_lp64.a ${MKLROOT}/lib/intel64/libmkl_intel_thread.a ${MKLROOT}/lib/intel64/libmkl_core.a ${MKLROOT}/lib/intel64/libmkl_blacs_intelmpi_lp64.a -Wl,--end-group ${CMPLR_ROOT}/linux/compiler/lib/intel64_lin/libifcoremt.a -lpthread -lm -ldl
- A modified OpenMX package is also used to compute overlap matrices only for large-scale materials structure. Install 'overlap only' OpenMX according to the readme documentation in this repository.
- Install OpenMX package version 3.9 for density functional theory Hamiltonian matrix calculation to construct datasets.
If you are using Intel MKL and Intel MPI environments, you can use the following variable definitions for makefile
- SIESTA: Install SIESTA package for density functional theory Hamiltonian matrix calculation to construct datasets. DeepH-pack requires SIESTA version >= 4.1.5.
Run the following command in the path of DeepH-pack:
git clone https://github.com/mzjb/DeepH-pack.git
cd DeepH-pack
pip install .
To perform efficient ab initio electronic structure calculation by DeepH method for a class of large-scale material systems, one needs to design an appropriate dataset of small structures that have close chemical bonding environment with the target large-scale material systems. Therefore, the first step of a DeepH study is to perform the DFT calculation on the above dataset to get the DFT Hamiltonian matrices with the localized basis. DeepH-pack supports DFT results made by ABACUS, OpenMX, FHI-aims or SIESTA and will support HONPAS soon.
For more information, see the documentation.
Preprocess
is a part of DeepH-pack. Through Preprocess
, DeepH-pack will
convert the unit of physical quantity, store the data files in the format
of text and HDF5 for each structure in a separate folder, generate local
coordinates, and perform basis transformation for DFT Hamiltonian matrices.
We use the following convention of units:
Quantity | Unit |
---|---|
Length | Å |
Energy | eV |
You need to edit a configuration in the format of ini, setting up the
file referring to the default file DeepH-pack/deeph/preprocess/preprocess_default.ini
.
The meaning of the keywords can be found in the
documentation.
For a quick start, you must set up raw_dir, processed_dir and interface.
With the configuration file prepared, run
deeph-preprocess --config ${config_path}
with ${config_path}
replaced by the path of your configuration file.
Train
is a part of DeepH-pack, which is used to train a deep learning model using the processed dataset.
Prepare a configuration in the format of ini, setting up the file referring to the default DeepH-pack/deeph/default.ini
. The meaning of the keywords can be found in the documentation. For a quick start, you must set up graph_dir, save_dir, raw_dir and orbital, other keywords can stay default and be adjusted later.
With the configuration file prepared, run
deeph-train --config ${config_path}
with ${config_path}
replaced by the path of your configuration file.
Tips:
-
Name your dataset. Use dataset_name to name your dataset, the same names may overwrite each other.
-
Hyperparameters of the neural network. The neural network here contains some hyperparameters. For a specific problem your should try adjusting the hyperparameters to obtain better results.
-
The keyword orbital. The keyword orbital states which orbitals or matrix elements are predicted. It is a little complicated to understand its data structure. To figure out it, you can refer to the documentation or the method make_mask in class
DeepHKernel
defined inDeepH-pack/deeph/kernel.py
.Alternatively, a Python script at
DeepH-pack/tools/get_all_orbital_str.py
can be used to generate a default configuration to predict all orbitals with one model. -
Use TensorBoard for visualizations. You can track and visualize the training process through TensorBoard by running
tensorboard --logdir=./tensorboard
in the output directory (save_dir):
Inference
is a part of DeepH-pack, which is used to predict the
DFT Hamiltonian for large-scale material structures and perform
sparse calculation of physical properties.
Firstly, one should prepare the structure file of large-scale material
and calculate the overlap matrix. Overlap matrix calculation does not
require SCF
. Even if the material system is large, only a small calculation
time and memory consumption are required. Following are the steps to
calculate the overlap matrix using different supported DFT packages:
- ABACUS: Set the following parameters in the input file of ABACUS
INPUT
:and run ABACUS like a normalcalculation get_S
SCF
calculation. ABACUS version >= 2.3.2 is required. - OpenMX: See this repository.
For overlap matrix calculation, you need to use the same basis set and DFT software when preparing the dataset.
Then, prepare a configuration in the format of ini, setting up the
file referring to the default DeepH-pack/deeph/inference/inference_default.ini
.
The meaning of the keywords can be found in the
INPUT KEYWORDS section.
For a quick start, you must set up OLP_dir, work_dir, interface,
trained_model_dir and sparse_calc_config, as well as a JSON
configuration file located at sparse_calc_config for sparse calculation.
With the configuration files prepared, run
deeph-inference --config ${config_path}
with ${config_path}
replaced by the path of your configuration file.
When the directory structure of the code folder is not modified, the scripts in it can be used to generate a dataset of non-twisted structures, train a DeepH model, make predictions on the DFT Hamiltonian matrix of twisted structure, and perform sparse diagonalization to compute the band structure for the example study of bismuthene.
Firstly, generate example input files according to your environment path by running the following command:
cd DeepH-pack
python gen_example.py ${openmx_path} ${openmx_overlap_path} ${pot_path} ${python_interpreter} ${julia_interpreter}
with ${openmx_path}
, ${openmx_overlap_path}
, ${pot_path}
, ${python_interpreter}
, and ${julia_interpreter}
replaced by the path of original OpenMX executable program, modified 'overlap only' OpenMX executable program, VPS and PAO directories of OpenMX, Python interpreter, and Julia interpreter, respectively. For example,
cd DeepH-pack
python gen_example.py /home/user/openmx/source/openmx /home/user/openmx_overlap/source/openmx /home/user/openmx/DFT_DATA19 python /home/user/julia-1.5.4/bin/julia
Secondly, enter the generated example/
folder and run run.sh
in each folder one-by-one from 1 to 5. Please note that run.sh
should be run in the directory where the run.sh
file is located.
cd example/1_DFT_calculation
bash run.sh
cd ../2_preprocess
bash run.sh
cd ../3_train
bash run.sh
cd ../4_compute_overlap
bash run.sh
cd ../5_inference
bash run.sh
The third step, the neural network training process, is recommended to be carried out on the GPU. In addition, in order to get the energy band faster, it is recommended to calculate the eigenvalues of different k points in parallel in the fifth step by which_k interface.
After completing the calculation, you can find the band structure data in OpenMX Band format of twisted bilayer bismuthene with 244 atoms per supercell computed by the predicted DFT Hamiltonian in the file below:
example/work_dir/inference/5_4/openmx.Band
The plotted band structure will be consistent with the right pannel of figure 6c in our paper.
You can train DeepH models using the existing dataset to reproduce the results of our paper.
Firstly, download the processed dataset for graphene (graphene_dataset.zip), MoS2 (MoS2_dataset.zip), twisted bilayer graphene (TBG_dataset.zip) or twisted bilayer bismuthene (TBB_dataset.zip). Uncompress the ZIP file.
Secondly, edit corresponding config files in the DeepH-pack/ini/
. raw_dir should be set to the path of the downloaded dataset. graph_dir and save_dir should be set to the path to save your graph file and results file during the training. For grahene, twisted bilayer graphene and twisted bilayer bismuthene, a single MPNN model is used for each dataset. For MoS2, four MPNN models are used. Run
deeph-train --config ${config_path}
with ${config_path}
replaced by the path of config file for training.
After completing the training, you can find the trained model in save_dir, which can be used to make prediction on new structures by run
deeph-inference --config ${inference_config_path}
with ${inference_config_path}
replaced by the path of config file for inference.
Please note that the DFT results in this dataset were calculated using OpenMX.
This means that if you want to use a model trained on this dataset to calculate properties, you need to use the overlap calculated using OpenMX.
The orbital information required for overlap calculations can be found in the paper.
Train the DeepH model by random graphene supercells and predict the Hamiltonian of carbon nanotube using the ABACUS interface. See README.md in this file for details.
-
He Li
Tsinghua University
mzjb313 [at] gmail [dot] com
-
Zechen Tang
Tsinghua University
- Xiaoxun Gong
- Honggeng Tao
- Zun Wang
- Nianlong Zou
- Ting Bao
-
Yong Xu
Tsinghua University
yongxu [at] tsinghua [dot] edu [dot] cn
-
Wenhui Duan
Tsinghua University
duanw [at] tsinghua [dot] edu [dot] cn