amaceing-toolkit 0.7.0

A Python package for the creation of input files for CP2K, MACE-torch, MatterSim, SevenNet, ORB, and Grace as well as the post-processing and evaluation of these simulations.

aMACEing_toolkit

Table of contents

• Description
• Supported calculation types and simulation engines
• Installation
  • (i) Installation from source
  • (ii) Installation via pip
• Usage
  • Command line interface
  • Python API
• Documentation
• Examples
• Citation
• License

Description

aMACEing_toolkit is a scientific software package that facilitates the creation, execution, and analysis of molecular dynamics simulations using multiple simulation engines. It provides a unified workflow for quantum-chemical calculations with CP2K and machine-learned interatomic potentials (MLIPs) including MACE, MatterSim, SevenNet, ORB and Grace.

The toolkit offers:

1. Input Generation: Easy-to-use interactive interfaces and API functions for creating input files for multiple simulation engines
2. Workflow Management: Consistent workflow across different software packages
3. Trajectory Analysis: Fast tools for analyzing simulation outputs (RDF, MSD, SMSD, vector autocorrelation)
4. Model Evaluation: Utilities to assess and compare MLIP performance against reference data
5. Run Logging: Tracking and documentation of simulation runs and fine-tuned models

Supported calculation types and simulation engines

Engine	Simulation/training types	File formats
CP2K	GEO_OPT, CELL_OPT, MD, REFTRAJ, ENERGY	Input files
MACE	GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE, FINETUNE_MULTIHEAD, TRAIN	ASE, LAMMPS
MatterSim	GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE	ASE
SevenNet	GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE, TRAIN	ASE, LAMMPS
ORB	GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE	ASE
Grace	GEO_OPT, CELL_OPT, MD, MULTI_MD, RECALC, FINETUNE, TRAIN	ASE, LAMMPS

Installation

(i) Installation from source (recommended)

1. Create a virtual environment:

   conda create -n atk python=3.12
   conda activate atk
   

2. Install PyTorch:

   pip install torch torchvision torchaudio
   

   For CPU-only systems:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
   

3. Install aMACEing_toolkit:

   git clone https://github.com/jhaens/amaceing_toolkit.git
   cd amaceing_toolkit
   pip install -r requirements.txt
   pip install .
   

4. (Optional) For accelerated MACE performance on CUDA systems:

   # CUDA 11
   pip install cuequivariance cuequivariance-torch cuequivariance-ops-torch-cu11
   # CUDA 12
   pip install cuequivariance cuequivariance-torch cuequivariance-ops-torch-cu12
   # Older Mace Versions (< 0.3.11):
   pip install cuequivariance==0.1.0 cuequivariance-torch==0.1.0 cuequivariance-ops-torch-cu12==0.1.0
   

5. (Optional) For MatterSim and SevenNet, create a separate environment:

   conda create -n atk_ms7n python=3.9
   conda activate atk_ms7n
   pip install torch torchvision torchaudio
   pip install mattersim==1.1.2 sevenn==0.11.2.post1
   

6. (Optional) For ORB models, create a separate environment:

   conda create -n atk_orb python=3.10
   conda activate atk_orb
   git clone https://github.com/orbital-materials/orb-models.git
   cd orb-models
   pip install .
   

7. (Optional) For Grace models, create a separate environment:

   conda create -n atk_grace python=3.11
   conda activate atk_grace
   pip install tensorpotential
   

(ii) Installation via pip

conda create -n atk python=3.12
conda activate atk
pip install torch torchvision torchaudio
pip install amaceing_toolkit

Note: The pip installation does not support running MatterSim and SevenNet simulations directly through the toolkit (only input file creation).

Usage

Command line interface

Each functionality can be accessed through dedicated command line tools:

# Access through interactive Q&A sessions
amaceing_cp2k           # CP2K input creation
amaceing_mace           # MACE input creation
amaceing_mattersim      # MatterSim input creation
amaceing_sevennet       # SevenNet input creation
amaceing_orb            # ORB model input creation
amaceing_grace          # Grace model input creation
amaceing_ana            # Trajectory analysis
amaceing_utils          # Utility functions

# Or with direct arguments for non-interactive use
amaceing_cp2k -rt="GEO_OPT" -c="{'project_name': 'test', 'coord_file': 'system.xyz', 'pbc_list': [10 0 0 0 10 0 0 0 10]}"

Python API

from amaceing_toolkit.workflow import cp2k_api

# Example: Generate a CP2K geometry optimization input
config = {
    'project_name': 'system_geoopt',
    'coord_file': 'system.xyz',
    'pbc_list': [14.2, 0, 0, 0, 14.2, 0, 0, 0, 14.2],
    'max_iter': 10,
    'xc_functional': 'BLYP',
}
cp2k_api(run_type='GEO_OPT', config=config)

Logger

The package includes a run and model logger to track simulation runs and fine-tuned models. This ensures reproducibility and easy access to past runs.

amaceing_utils --logger=run         # Get an overview of runs
amaceing_utils --logger=runexport   # Export run logger to PDF
amaceing_utils --logger=model       # Get an overview of models

Documentation

Complete documentation is available at ReadTheDocs.

The documentation provides:

• Comprehensive installation instructions
• Detailed API reference
• User guides for all supported simulation engines
• Configuration guides
• Example workflows
• Tutorials for common use cases:
  • NEW: Getting Started with aMACEing_toolkit Introduction to aMACEing_toolkit
  • Tutorial A
  • Tutorial B

Examples

Example systems and scripts are provided in the examples/ directory:

• 4KOH_92H2O_333K - Potassium hydroxide in water
• CsH2PO4 - Cesium dihydrogen phosphate
• analyzer - Trajectory analysis examples
• api_examples - Python API usage examples

Citation

If you use aMACEing_toolkit in your research, please cite:

Hänseroth, J. and Flötotto, A. and Qaisrani, M. N. and Dreßler, C. "Fine-Tuning Unifies Foundational Machine-learned Interatomic Potential Architectures at ab initio Accuracy." arXiv preprint arXiv:2511.05337 (2025).

@misc{hänseroth2025amaceingtoolkit,
      title={Fine-Tuning Unifies Foundational Machine-learned Interatomic Potential Architectures at ab initio Accuracy}, 
      author={Jonas Hänseroth and Aaron Flötotto and Muhammad Nawaz Qaisrani and Christian Dreßler},
      year={2025},
      eprint={2511.05337},
      archivePrefix={arXiv},
      primaryClass={physics.chem-ph},
      url={https://arxiv.org/abs/2511.05337}, 
}

License

This project is licensed under the MIT License (Non-Commercial Use Only).