adamlaho/MolCryst

MC-MLIPs

Molecular Crystals Database for Machine Learning Interatomic Potentials

Fine-tuned MACE models for polymorphic molecular crystals, trained using the AMLP framework.

Models

Compound	CSD Code	Energy MAE (meV/atom)	Force MAE (meV/Å)	Model	Dataset
Resorcinol	RESORA	1.568	3.903	↓ model	↓ train ↓ valid
Durene	DURENE	1.647	5.193	↓ model	↓ train ↓ valid
Coumarin	COUMAR	1.670	4.296	↓ model	↓ train ↓ valid
Benzamide	BZAMID	0.713	8.786	↓ model	↓ train ↓ valid
Niacinamide	NICOAM	1.513	7.207	↓ model	↓ train ↓ valid
Nicotinamide	NICOAC	1.201	5.824	↓ model	↓ train ↓ valid
Isonicotinamide	EHOWIH	1.912	10.809	↓ model	↓ train ↓ valid
Pyrazinamide	PYRIZIN	1.634	6.732	↓ model	↓ train ↓ valid
Benzoic acid	BENZAC	1.329	7.897	↓ model	↓ train ↓ valid
Acridine	ACRDIN	3.700	8.300	↓ model	↓ train ↓ valid
Mean		1.689	6.895

Training Protocol

• Foundation Model: MACE-MP-0 (mace-mh-1-omol-1%)
• Reference Data: DFT (PBE-D4) optimizations + AIMD trajectories (25-500K)
• DFT Settings: VASP, 750 eV cutoff, EDIFF = 10⁻⁷ eV

Two-stage training:

1. Initial: LR = 2×10⁻³, energy weight = 100, force weight = 10
2. SWA (epoch 200+): LR = 5×10⁻⁵, force weight = 100

Early stopping with patience = 75 epochs. All models trained in float64.

Validation

All models validated for:

• Energy conservation: NVE drift < 10⁻⁵ over 25 ps
• Thermal stability: NVT stable up to 600K
• Structural integrity: RDFs and P₂ order parameters preserved

Usage

We recommend using the AMLP-Analysis module (amlpa.py) for running simulations with these models:

python3 amlpa.py structure.xyz config.yaml

In your config.yaml, point to the downloaded model:

model_paths:
  - 'path/to/model.model'
device: 'gpu'
gpus: ['cuda:0']

Alternatively, you can use the models directly via the MACE calculator:

from mace.calculators import MACECalculator
calc = MACECalculator(model_paths="path/to/model.model", device="cuda")
atoms.calc = calc

For full configuration options (MD, geometry optimization, RDF analysis, etc.), refer to the AMLP documentation.

Repository Structure

MC-MLIPs/
├── models/     -> Trained MACE model files
└── datasets/   -> Training/validation HDF5 files

Citation

If you use these models, please cite:

@article{lahouari2026mcmlips,
  title={Benchmarking Fine-Tuned MACE Interatomic Potentials for Polymorphic Molecular Crystals},
  author={Lahouari, Adam and Ai, Shen and Han, Jihye and Hoffstadt, Jillian and H{\"o}llmer, Philipp and Infante, Charlotte and Jain, Pulkita and Kadam, Sangram and Martirossyan, Maya M. and McCune, Amara and Newton, Hypatia and Paul, Shlok J. and Pena, Willmor and Raghoonanan, Jonathan and Sahu, Sumon and Tan, Oliver and Vergara, Andrea and Rogal, Jutta and Tuckerman, Mark E.},
  journal={arXiv preprint arXiv:XXXXX},
  year={2026}
}

@article{lahouari2026amlp,
  title={Automated Machine Learning Pipeline: Large Language Models-Assisted Automated Data set Generation for Training Machine-Learned Interatomic Potentials},
  author={Lahouari, Adam and Rogal, Jutta and Tuckerman, Mark E.},
  journal={Journal of Chemical Theory and Computation},
  volume={22},
  number={1},
  pages={305--317},
  year={2026},
  publisher={American Chemical Society},
  doi={10.1021/acs.jctc.5c01610},
  url={https://doi.org/10.1021/acs.jctc.5c01610}
}

License

MIT License

Acknowledgments

• MACE development team
• NYU High Performance Computing