SevenNet

SevenNet (Scalable EquiVariance Enabled Neural Network) is a graph neural network interatomic potential package that supports parallel molecular dynamics simulations with LAMMPS. Its underlying GNN model is based on nequip.

The project provides parallel molecular dynamics simulations using graph neural network interatomic potentials, which enable large-scale MD simulations or faster MD simulations.

The installation and usage of SevenNet are split into two parts: training + command-line interface + ASE calculator (handled by Python) and molecular dynamics (handled by LAMMPS).

PLEASE NOTE: SevenNet+LAMMPS parallel after this commit: 14851ef (v0.9.3 ~ 0.9.5) has a serious bug: it gives wrong forces when the number of mpi processes is greater than two. The corresponding pip version is yanked for this reason. The bug has been fixed in the main branch, and I'll release the corrected version ASAP.

Features

Pre-trained GNN interatomic potential SevenNet-0, with fine-tuning interface
ASE calculator support
Multi-GPU accelerated molecular dynamics with LAMMPS
Accelerated D3 (van der Waals) dispersion, written in CUDA.

Supporting MD frameworks and its features. While all modes support both CPU and GPU, GPU is much faster.

Features	ASE calculator	LAMMPS serial	LAMMPS parallel
Working?	✅	✅	✅
Multi-GPU	❌	❌	✅
Stress	✅	✅	⏳
D3 correction	⏳	✅	⏳

✅: Support, ⏳: Planned, ❌: Not planned.

SevenNet

Installation

Python >= 3.8
PyTorch >= 1.12.0

Please install PyTorch from PyTorch official before installing the SevenNet. Note that for SevenNet, torchvision and torchaudio are redundant. You can safely exclude these packages from the installation commands.

Here are the recommended versions we've been using internally without any issues.

PyTorch/2.2.2 + CUDA/12.1.0
PyTorch/1.13.1 + CUDA/12.1.0
PyTorch/1.12.0 + CUDA/11.6.2

Using the newer versions of CUDA with PyTorch is usually not a problem. For example, you can compile and use PyTorch/1.13.1+cu117 with CUDA/12.1.0.

PLEASE NOTE: You must install PyTorch before installing SevenNet. They are not marked as dependencies since it is coupled with the CUDA version.

PLEASE NOTE: Currently backward compatibility is ensured for only checkpoint files.

After the PyTorch installation, run

pip install sevenn

To download the latest version of SevenNet(not stable!), run

pip install https://github.com/MDIL-SNU/SevenNet.git

Usage

SevenNet-0

SevenNet-0 is a general-purpose interatomic potential trained on the MPF dataset of M3GNet or MPtrj dataset of CHGNet.

While SevenNet-0 can be applied to downstream tasks as it is, it is recommended to fine-tune SevenNet-0 before addressing real downstream tasks.

SevenNet-0 (11July2024)

This model was trained on MPtrj. We suggest starting with this model as we found that it performs better than the previous SevenNet-0 (22May2024). Check Matbench Discovery leaderborad for this model's performance on materials discovery. For more information, click here.

Whenever the checkpoint path is the input, this model can be loaded via 7net-0 | SevenNet-0 | 7net-0_11July2024 | SevenNet-0_11July2024 keywords.

Acknowledgments: This work was supported by the Neural Processing Research Center program of Samsung Advanced Institute of Technology, Samsung Electronics Co., Ltd. The computations for training models were carried out using the Samsung SSC-21 cluster.

SevenNet-0 (22May2024)

This model was trained on MPF.2021.2.8. This is the model used in our paper. For more information, click here.

Whenever the checkpoint path is the input, this model can be loaded via 7net-0_22May2024 | SevenNet-0_22May2024 keywords.

SevenNet Calculator for ASE

ASE (Atomic Simulation Environment) is a set of tools and Python modules for atomistic simulations. SevenNet-0 and SevenNet-trained potentials can be used with ASE for its use in python.

For pre-trained models,

from sevenn.sevennet_calculator import SevenNetCalculator
sevennet_0_cal = SevenNetCalculator("7net-0", device='cpu')  # 7net-0, SevenNet-0, 7net-0_22May2024, 7net-0_11July2024 ...

For user trained models,

from sevenn.sevennet_calculator import SevenNetCalculator
checkpoint_path = ### PATH TO CHECKPOINT ###
sevennet_cal = SevenNetCalculator(checkpoint_path, device='cpu')

Training

sevenn_preset fine_tune > input.yaml
sevenn input.yaml -s

Other valid preset options are: base, fine_tune, and sevennet-0. Check comments in the preset yaml files for explanations. For fine-tuning, note that most model hyperparameters cannot be modified unless explicitly indicated.

To reuse a preprocessed training set, you can specify ${dataset_name}.sevenn_data to the load_dataset_path: in the input.yaml.

Multi-GPU training

We support multi-GPU training features using PyTorch DDP (distributed data parallel). We use one process (or a CPU core) per GPU.

torchrun --standalone --nnodes {number of nodes} --nproc_per_node {number of GPUs} --no_python sevenn input.yaml -d

Please note that batch_size in input.yaml indicates batch_size per GPU.

sevenn_graph_build

sevenn_graph_build -f ase my_train_data.extxyz 5.0

You can preprocess the dataset with sevenn_graph_build to obtain *.sevenn_data files. The cutoff length should be provided. See sevenn_graph_build --help for more information.

sevenn_inference

sevenn_inference checkpoint_best.pth path_to_my_structures/*

This will create dir sevenn_infer_result. It includes .csv files that enumerate prediction/reference results of energy and force. See sevenn_inference --help for more information.

sevenn_get_model

This command is for deploying lammps potentials from checkpoints. The argument is either the path to checkpoint or the name of pre-trained potential.

sevenn_get_model 7net-0

This will create deployed_serial.pt, which can be used as lammps potential under e3gnn pair_style.

The parallel model can be obtained in a similar way

sevenn_get_model 7net-0 -p

This will create a directory with multiple deployed_parallel_*.pt files. The directory path itself is an argument for the lammps script. Please do not modify or remove files under the directory. These models can be used as lammps potential to run parallel MD simulations with GNN potential using multiple GPU cards.

Installation for LAMMPS

PyTorch (same version as used for training)
LAMMPS version of 'stable_2Aug2023_update3' LAMMPS
(Optional) CUDA-aware OpenMPI for parallel MD
MKL-include

PLEASE NOTE: CUDA-aware OpenMPI does not support NVIDIA Gaming GPUs. Given that the software is closely tied to hardware specifications, please consult with your server administrator if unavailable.

PLEASE NOTE: Virial stress (pressure) outputs of SevenNet parallel are not correct! Work in progress...

If your cluster supports the Intel MKL module (often included with Intel OneAPI, Intel Compiler, and other Intel-related modules), load the module. If it is unavailable, read the 'Note for MKL' section before running cmake.

CUDA-aware OpenMPI is optional but recommended for parallel MD. If it is not available, in parallel mode, GPUs will communicate via CPU. It is still faster than using only one GPU, but its efficiency is low.

Ensure the LAMMPS version (stable_2Aug2023_update3). You can easily switch the version using git. After switching the version, run sevenn_patch_lammps with the lammps directory path as an argument.

git clone https://github.com/lammps/lammps.git lammps_sevenn --branch stable_2Aug2023_update3 --depth=1
sevenn_patch_lammps ./lammps_sevenn {--d3}

Add --d3 option to install GPU accelerated Grimme's D3 method pair style (currently available in main branch only, not pip). For its usage and details, click here.

You can refer to sevenn/pair_e3gnn/patch_lammps.sh for the detailed patch process.

Build LAMMPS with cmake (example):

cd ./lammps_sevenn
mkdir build
cd build
cmake ../cmake -DCMAKE_PREFIX_PATH=`python -c 'import torch;print(torch.utils.cmake_prefix_path)'`
make -j4

If the compilation is successful, you will find the executable at {path_to_lammps_dir}/build/lmp. To use this binary easily, for example, create a soft link in your bin directory (which should be included in your $PATH).

ln -s {absolute_path_to_lammps_dir}/build/lmp $HOME/.local/bin/lmp

This will allow you to run the binary using lmp -in my_lammps_script.lmp.

Note for MKL

You may encounter MKL_INCLUDE_DIR NOT-FOUND during cmake. This usually means the environment variable is not set correctly, or mkl-include is not present on your system.

Install mkl-include with:

conda install -c intel mkl-include

If you encounter an error, remove -c intel. This is a known bug in the recent Conda version.

Append the following to your cmake command:

-DMKL_INCLUDE_DIR=$CONDA_PREFIX/include

If you see hundreds of undefined reference to XXX errors with libtorch_cpu.so at the end of compilation, check your $LD_LIBRARY_PATH. PyTorch depends on MKL libraries (this is a default backend for torch+CPU), therefore you already have them. For example, if you installed PyTorch using Conda, you may find libmkl_*.so files under $CONDA_PREFIX/lib. Ensure that $LD_LIBRARY_PATH includes $CONDA_PREFIX/lib.

For other error cases, you might want to check pair-nequip, as the pair-nequip and SevenNet+LAMMPS shares similar requirements: torch + LAMMPS.

Usage for LAMMPS

To check installation

{lammps_binary} -help | grep e3gnn

You will see e3gnn and e3gnn/parallel as pair_style.

For serial model

units         metal
atom_style    atomic
pair_style e3gnn
pair_coeff * * {path to serial model} {space separated chemical species}

For parallel model

units         metal
atom_style    atomic
pair_style e3gnn/parallel
pair_coeff * * {number of message-passing layers} {path to the directory containing parallel model} {space separated chemical species}

For example,

pair_style e3gnn/parallel
pair_coeff * * 4 ./deployed_parallel Hf O

The number of message-passing layers is equal to the number of *.pt files in the ./deployed_parallel directory.

Use sevenn_get_model for deploying lammps models from checkpoint for both serial and parallel.

One GPU per MPI process is expected. The simulation may run inefficiently if the available GPUs are fewer than the MPI processes.

PLEASE NOTE: Currently, the parallel version raises an error when there are no atoms in one of the subdomain cells. This issue can be addressed using the processors command and, more optimally, the fix balance command in LAMMPS. This will be patched in the future.

Future Works

Notebook examples and improved interface for non-command line usage
Development of a tiled communication style (also known as recursive coordinate bisection, RCB) in LAMMPS.

Citation

If you use SevenNet, please cite (1) parallel GNN-IP MD simulation by SevenNet or its pre-trained model SevenNet-0, (2) underlying GNN-IP architecture NequIP

(1) Y. Park, J. Kim, S. Hwang, and S. Han, "Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations". J. Chem. Theory Comput., 20(11), 4857 (2024) (https://pubs.acs.org/doi/10.1021/acs.jctc.4c00190)

(2) S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, "E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials". Nat. Commun., 13, 2453. (2022) (https://www.nature.com/articles/s41467-022-29939-5)

MDIL-SNU/SevenNet