MACE4IR: A foundation model for molecular infrared spectroscopy

MACE4IR is a domain-specific machine-learned interatomic potential trained for accurate prediction of infrared (IR) spectra of molecules (consists of ~80 elements from the periodic table). It is built upon the MACE framework and trained using a selected subset of the QCML dataset QCML dataset, consisting of molecular systems relevant in catalysis, chemical synthesis, and energy applications.

This repository provides pretrained models for direct inference or further fine-tuning. The models can be used with the MACECalculator class for integration into ASE-based molecular simulations and IR spectrum generation workflows.

Model Highlights

• Developed for IR spectral prediction using both Harmonic approximation and MD trajectories.
• Trained on high-quality quantum data and validated against reference DFT and experimental spectra.
• Supports fast and transferable inference across molecular temperatures and structural variants.

Installation, Setup and Usage

The step-by-step tutorials on using the pretrained model with ASE and MACE, including installation instructions, are available here: Tutorial_I and Tutorial_II.

Training Information and Data

The MACE4IR model was trained on 10M datapoints from the QCML dataset. These are available on zenodo (doi: 10.5281/zenodo.16761021 and doi: 10.5281/zenodo.16920067) and in the future, we may upload to HuggingFace Datasets.

Fine-tuning

To fine-tune the model on new molecules or configurations, please follow the steps mentioned here: Fine-tuning Guide.

Updates

New models incorporating MBD correction and D4 dispersion correction will be available soon. Stay tuned!

License

The pretrained models are released under the Academic Software License. They are free to use for academic purposes, but commercial use is not permitted. For full details, please refer to the LICENSE.md file.

Contact

For questions, bug reports, or collaboration inquiries, please open an issue/discussion in this repository.

If you use our work, please cite:

@misc{bhatia2025mace4irfoundationmodelmolecular,
  title        = {MACE4IR: A foundation model for molecular infrared spectroscopy},
  author       = {Nitik Bhatia and Ondrej Krejci and Silvana Botti and Patrick Rinke and Miguel A. L. Marques},
  year         = {2025},
  eprint       = {2508.19118},
  archivePrefix= {arXiv},
  primaryClass = {physics.chem-ph},
  url          = {https://arxiv.org/abs/2508.19118},
}