SevenNet Module¶

The SevenNet module is a component of the aMACEing toolkit designed to facilitate the creation of input files for SevenNet simulations.

SevenNet Repository

Capabilities¶

The SevenNet module supports the creation of input files (ASE/LAMMPS) for the following calculation types:

Geometry Optimization (GEO_OPT)
Cell Optimization (CELL_OPT)
Molecular Dynamics (MD)
Multi-Configuration Molecular Dynamics (MULTI_MD)
Reference Trajectory Recalculation (RECALC): Recomputes energies and forces along an existing trajectory
Fine-tuning of Foundation Models (FINETUNE)

Usage¶

Command-line Usage¶

Interactive Q&A session:

amaceing_sevennet

This guides you through:

Selecting a coordinate file
Defining the simulation box
Choosing the calculation type
Selecting the foundation model
Setting calculation-specific parameters

Direct Command Line Usage:

amaceing_sevennet -rt="RUN_TYPE" -c="{'parameter1': 'value1', 'parameter2': 'value2', ...}"

Where RUN_TYPE is one of: GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE, TRAIN

For MD:

amaceing_sevennet -rt="MD" -c="{'project_name': 'NAME', 'coord_file': 'FILE', 'pbc_list': '[10 0 0 0 10 0 0 0 10]', 'foundation_model': '7net-0', 'temperature': '300', 'thermostat': 'Langevin', 'nsteps': '10000', 'timestep': '0.5', 'write_interval': '10', 'log_interval': '10', 'dispersion_via_simenv': 'n', 'print_ext_traj': 'y', 'simulation_environment': 'ase'}"

For FINETUNE:

amaceing_sevennet -rt="FINETUNE" -c="{'project_name': 'NAME', 'foundation_model': '7net-0', 'train_file': 'FILE', 'batch_size': 'INT', 'epochs': 'INT', 'seed': '1', 'lr': '0.01', 'force_loss_ratio': 1.0, 'device': 'cuda'}"

For TRAIN:

amaceing_sevennet -rt="TRAIN" -c="{'project_name': 'NAME', 'train_file': 'FILE', 'batch_size': 'INT', 'epochs': 'INT', 'seed': '1', 'lr': '0.01', 'force_loss_ratio': 1.0, 'device': 'cuda'}"

For RECALC:

amaceing_sevennet -rt="RECALC" -c="{'project_name': 'NAME', 'coord_file': 'FILE', 'pbc_list': '[10 0 0 0 10 0 0 0 10]', 'dispersion_via_simenv': 'n', 'foundation_model': '7net-mf-ompa', 'modal': 'mpa', 'simulation_environment': 'ase'}"

Note

Do NOT use double quotes inside the dictionary. Also do NOT use commas inside of lists in the dictionary.

Python API¶

from amaceing_toolkit.workflow import sevennet_api

config = {
    'project_name': 'koh_h2o_geoopt',
    'coord_file': 'system.xyz',
    'pbc_list': [14.2067, 0, 0, 0, 14.2067, 0, 0, 0, 14.2067],
    'max_iter': 100,
    'foundation_model': '7net-mf-ompa',
    'modal': 'mpa',
    'dispersion_via_simenv': 'y',
    'simulation_environment': 'ase'
}

sevennet_api(run_type='GEO_OPT', config=config)

Output Files¶

The module generates:

Python script for the calculation (<runtype>_sevennet.py)
HPC runscripts for execution (runscript.sh and gpu_script.job)
For fine-tuning: YAML configuration file (config_finetune.yml)
Log file with configuration parameters (sevennet_input.log)
For recalculation: Files with recalculated energies and forces
For multi-configuration MD: Directory structure with files for each configuration

Foundation Models¶

The module supports various foundation models:

7net-mf-ompa: (recommended) multi-fidelity model trained on Materials Project data, Alexandria data (modal: mpa) and Meta Open Materials 2024 (modal: omat24) data
7net-omat: model trained on Meta Open Materials 2024 data
7net-l3i5: model trained on Materials Project data (increased maximum spherical harmonics degree to 3)
7net-0: model trained on Materials Project data (default model, only model available for fine-tuning)
custom: User-provided model path or model from the model logger

Technical Details¶

Thermostats: Langevin, NoseHooverChainNVT, Bussi and NPT
Environment management: Runs in a separate conda environment to avoid package conflicts
Dispersion corrections: Optional inclusion of dispersion via ASE or LAMMPS (But: ASE dispersion corrections are only available on GPU)
Model Logger: Automatic tracking of fine-tuned models