MACE-Osaka24 models

This repository provides the model and training scripts for a multi-domain universal machine learning interatomic potentials (MLIPs), the MACE-Osaka24 models, capable of accurately describing both crystalline and molecular domains.

The MACE-Osaka24 model is a universal MLIP trained on datasets of both crystals and molecules, which were generated using a dataset integration technique called "Total Energy Alignment" that combines first-principles calculations under various conditions.

Its architecture is based on the first-generation MACE model. To use the models please install the MACE code.

Models

The first generation of models are available in the MACE-Osaka24.

If you use the models, in addition to citing the original MACE papers, please cite:

@misc{shiota2024taming,
    title={Taming Multi-Domain, -Fidelity Data: Towards Foundation Models for Atomistic Scale Simulations},
    author={Tomoya Shiota and Kenji Ishihara and Tuan Minh Do and Toshio Mori and Wataru Mizukami},
    year={2024},
    eprint={2412.13088},
    archivePrefix={arXiv},
    primaryClass={physics.chem-ph}
}

Training scripts

We provide training scripts for the models in this repository. The latest training command line is found in mace-osaka24/mace-osaka24-large.sh.

Training data

The integrated inorganic–organic domain dataset used to train the models—composed of the inorganic MPtrj dataset and the organic SPICE, QMug, water clusters, and Tripeptides (OFF23) datasets—is available at figshare. If you use any of these datasets, please cite the following paper.

@article{deng2023chgnet,
      title={CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling},
      author={Bowen Deng and Peichen Zhong and KyuJung Jun and Janosh Riebesell and Kevin Han and Christopher J. Bartel and Gerbrand Ceder},
      year={2023},
      eprint={2302.14231},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci}
}

@misc{kovacs2023maceoff23,
      title={MACE-OFF23: Transferable Machine Learning Force Fields for Organic Molecules}, 
      author={Dávid Péter Kovács and J. Harry Moore and Nicholas J. Browning and Ilyes Batatia and Joshua T. Horton and Venkat Kapil and William C. Witt and Ioan-Bogdan Magdău and Daniel J. Cole and Gábor Csányi},
      year={2023},
      eprint={2312.15211},
      archivePrefix={arXiv},
}

@article{eastman2023spice,
  title={Spice, a dataset of drug-like molecules and peptides for training machine learning potentials},
  author={Eastman, Peter and Behara, Pavan Kumar and Dotson, David L and Galvelis, Raimondas and Herr, John E and Horton, Josh T and Mao, Yuezhi and Chodera, John D and Pritchard, Benjamin P and Wang, Yuanqing and others},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={11},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

@article{donchev2021quantum,
  title={Quantum chemical benchmark databases of gold-standard dimer interaction energies},
  author={Donchev, Alexander G and Taube, Andrew G and Decolvenaere, Elizabeth and Hargus, Cory and McGibbon, Robert T and Law, Ka-Hei and Gregersen, Brent A and Li, Je-Luen and Palmo, Kim and Siva, Karthik and others},
  journal={Scientific data},
  volume={8},
  number={1},
  pages={55},
  year={2021},
  publisher={Nature Publishing Group UK London}
}

@article{isert2022qmugs,
  title={QMugs, quantum mechanical properties of drug-like molecules},
  author={Isert, Clemens and Atz, Kenneth and Jim{\'e}nez-Luna, Jos{\'e} and Schneider, Gisbert},
  journal={Scientific Data},
  volume={9},
  number={1},
  pages={273},
  year={2022},
  publisher={Nature Publishing Group UK London}
}

@misc{shiota2024taming,
    title={Taming Multi-Domain, -Fidelity Data: Towards Foundation Models for Atomistic Scale Simulations},
    author={Tomoya Shiota and Kenji Ishihara and Tuan Minh Do and Toshio Mori and Wataru Mizukami},
    year={2024},
    eprint={2412.13088},
    archivePrefix={arXiv},
    primaryClass={physics.chem-ph}
}

Example

In this example, the energy of a silicon crystal and acetic acid is calculated using universal multi-domain MLIP MACE-Osaka24 and Atomic Simulation Environment (ASE).

from ase.build import bulk
from ase.build import molecule
from mace.calculators import MACECalculator

si = bulk('Si', 'diamond', a=5.43)
calculator = MACECalculator(model_path='/path-to-mace-osaka24/mace-osaka24-large.model', device='cpu')
si.calc = calculator 

energy_si = si.get_potential_energy()
print("Single-point energy of diamond Si:", energy_si)

acid = molecule('CH3COOH')
calculator = MACECalculator(model_path='/path-to-mace-osaka24/mace-osaka24-large.model', device='cpu')
acid.calc = calculator 

energy_acid = acid.get_potential_energy()
print("Single-point energy of acetic acid:", energy_acid)

Contributors

This project was developed by:

• Tomoya Shiota (@TShiotaSS)
• Kenji Ishihara (@kenji-ishihara-os)
• Toshio Mori (@forest1040)
• Wataru Mizukami (@wmizukami)