What's new

03 September 2026: 0.6.1 Release

• Foundation models re-exported. The distributed models were rebuilt to fix a NaN in virial_sigma (the uncertainty contribution to the virial). Cached models now record where they were downloaded from, and the new grace_models update re-downloads only the copies that are out of date.
• Tunable uncertainty threshold. grace_uq build --threshold-percentile Q puts gamma = 1 at the Q-th percentile of the training distribution instead of the default robust outlier fence — for training sets where a large fraction of atoms reads gamma > 1. The artifact records which calibration was used, and grace_uq info reports it.
• Bulk prediction. A new predict_structures helper evaluates a whole list of structures with a single calculator instance; grace_predict and grace_uq predict now both use it.

07 July 2026: 0.6.0 Release

• Uncertainty estimates. Models can now report a per-atom uncertainty score (gamma) that flags when a prediction is an extrapolation — useful for spotting unreliable regions and for active learning. Comes with a new grace_uq command line tool.
• LAMMPS Kokkos support. You can now export GRACE models for the fast Kokkos pair styles in LAMMPS. The uncertainty score can be carried along, so it's available during LAMMPS runs too.
• Foundation models on HuggingFace (AMS-ICAMS-RUB/grace-foundation-models). The released models now include uncertainty support out of the box,
• GRACE-3L-OMAT/OAM. There are new, larger 3-layer models (GRACE-3L-OMAT-large and GRACE-3L-OMAT-large-ft-AM).
• Faster and lighter by default. Foundation models now use fp32 precision by default, which roughly halves memory use and about doubles speed. (fp64 versions are still available under the -fp64 name.)
• Better tooling. A new grace_dashboard shows training curves and metrics in the browser; the grace_models command was reworked to list, inspect, and download models (including Kokkos weights); and training can now save and resume in the middle of an epoch.

Important Note

If a model was fitted with gracemaker version < 0.5.1, it will not be compatible with newer versions due to a format change.
You can convert it to the new format using the following command:

grace_utils -p seed/1/model.yaml -c seed/1/checkpoints/checkpoint.best_test_loss.index update_model

This will generate new files with the "-converted" suffix, which you can replace the old files (model.yaml and checkpoints) with.

GRACE - GRaph Atomic Cluster Expansion

Project GRACEmaker is a heavily modified and in large parts rewritten version of the PACEmaker software geared towards support for multi-component materials and graph architectures.

Documentation

Please see documentation for installation instructions and examples.

Tutorial

You can find tutorial materials here

Also in a video format

Support

Also, you may join ACE support Zulip channel for additional resources: https://acesupport.zulipchat.com/join/xtwxu2grjbtg64m3vnhypi6p/

Reference

Please see

• Y.Lysogorskiy, A. Bochkarev, R.Drautz, Graph atomic cluster expansion for foundational machine learning interatomic potentials. npj Computational Materials 12(1), 114 (2026)

@article{lysogorskiy2026graph,
  title={Graph atomic cluster expansion for foundational machine learning interatomic potentials},
  author={Lysogorskiy, Yury and Bochkarev, Anton and Drautz, Ralf},
  journal={npj Computational Materials},
  volume={12},
  number={1},
  pages={114},
  year={2026},
  doi={10.1038/s41524-026-01979-1},
  url={https://www.nature.com/articles/s41524-026-01979-1}
}

• Anton Bochkarev, Yury Lysogorskiy, and Ralf Drautz Graph Atomic Cluster Expansion for Semilocal Interactions beyond Equivariant Message Passing. Phys. Rev. X 14, 021036 (2024)

@article{PhysRevX.14.021036,
  title = {Graph Atomic Cluster Expansion for Semilocal Interactions beyond Equivariant Message Passing},
  author = {Bochkarev, Anton and Lysogorskiy, Yury and Drautz, Ralf},
  journal = {Phys. Rev. X},
  volume = {14},
  issue = {2},
  pages = {021036},
  numpages = {28},
  year = {2024},
  month = {Jun},
  publisher = {American Physical Society},
  doi = {10.1103/PhysRevX.14.021036},
  url = {https://link.aps.org/doi/10.1103/PhysRevX.14.021036}
}