SevenNet

• SevenNet (Scalable EquiVariance-Enabled Neural Network) is a graph neural network (GNN)-based interatomic potential package. It supports efficient atomistic simulations—including parallel molecular dynamics (MD) with LAMMPS—using pretrained models and a fine-tuning interface.
• GitHub Repository: https://github.com/MDIL-SNU/SevenNet

Installation

Requirements

• Python ≥ 3.8
• PyTorch ≥ 1.12.0 (recommended versions used internally include PyTorch 2.2.2 + CUDA 12.1.0, PyTorch 1.13.1 + CUDA 12.1.0, or PyTorch 1.12.0 + CUDA 11.6.2)

Important: Install PyTorch manually according to your hardware before installing SevenNet.

Installing SevenNet

To install via pip:

pip install sevenn

For the latest development version from GitHub:

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

Make sure to consult SevenNet GitHub Repository for new features and updates, as SevenNet is under active development.

Usage

Using the ASE Calculator

SevenNet provides an ASE interface through its calculator. Here are two common usage examples:

Basic ASE Calculator

Load a pretrained SevenNet model using a specified keyword and, if applicable, a modality (e.g., mpa or omat24):

from sevenn.calculator import SevenNetCalculator

# Load a multi-fidelity model (using modal='mpa' or 'omat24' based on training DFT settings)
calc = SevenNetCalculator(model='7net-mf-ompa', modal='mpa')

Using CUDA-accelerated D3 Calculations

For dispersion-corrected calculations with GPU acceleration:

from sevenn.calculator import SevenNetD3Calculator

# Replace '7net-0' with the desired model keyword and specify device
calc = SevenNetD3Calculator(model='7net-0', device='cuda')

Tip: When the device is set to 'auto', SevenNet will use GPU acceleration if available.

Command-line Interface for Training & Inference

SevenNet offers multiple CLI commands for managing preprocessing, training, inference, and model deployment:

1. Input Generation
   Use sevenn_preset to generate an input YAML file:

   sevenn_preset base > input.yaml
   

   Preset keywords include base, fine_tune, multi_modal, sevennet-0, and sevennet-l3i5.

2. Preprocessing (Optional)
   Preprocess your dataset with:

   sevenn_graph_build {dataset_path} {cutoff_radius}
   

   This creates preprocessed files (e.g., sevenn_data/graph.pt) that can speed up training and inference.

3. Training
   With your input.yaml and preprocessed data ready, start training:

   sevenn input.yaml -s
   

   For multi-GPU training using PyTorch DDP:

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

   (Note: batch_size in input.yaml is per GPU.)

4. Inference
   Use the trained checkpoint to predict energies, forces, and stresses:

   sevenn_inference checkpoint_best.pth path_to_structures/*
   

   Results are saved in the sevenn_infer_result directory.

5. Deployment for LAMMPS
   Deploy a model as a LAMMPS potential:

   sevenn_get_model 7net-0
   

   For parallel MD potentials:

   sevenn_get_model 7net-0 -p
   

   The resulting files/directories can be used with LAMMPS’ e3gnn pair_style.

MD Simulation with LAMMPS

SevenNet supports MD simulations using LAMMPS:

• Single-GPU MD:
  Use the serial model with the e3gnn pairstyle.
  _Example Input:

  units       metal
  atom_style  atomic
  pair_style  e3gnn
  pair_coeff  * * {path_to_serial_model} {chemical_species}
  

• Multi-GPU MD:
  Use the parallel model with e3gnn/parallel pairstyle.
  _Example Input:

  units       metal
  atom_style  atomic
  pair_style  e3gnn/parallel
  pair_coeff  * * {num_message_layers} {directory_of_parallel_model} {chemical_species}
  

  (Ensure one GPU per MPI process is available.)

For detailed LAMMPS setup, refer to the provided scripts and documentation. A patch script (sevenn_patch_lammps) is available for compiling LAMMPS with SevenNet-specific modifications.

Notebook Tutorials

Several Jupyter Notebook tutorials are available to help users:

• From Scratch: Learn to train SevenNet from scratch, perform predictions, structure relaxations, and generate EOS curves.
• Fine-Tuning: Fine-tune a pretrained model and compare its performance to the original model.

To access these notebooks, clone the repository:

git clone https://github.com/MDIL-SNU/sevennet_tutorial.git