Skip to content

GRACE

GRACE models are exported TensorFlow SavedModels, loaded through tensorpotential. The model path is the SavedModel directory, not a single file.

What the model is

  • Architecture. GRACE (graph atomic cluster expansion) extends the atomic cluster expansion from local stars to tree graphs: 1L models are local and fast, 2L models add one level of equivariant message passing for semilocal interactions (Bochkarev, Lysogorskiy & Drautz, PRX 2024).
  • Training data of the example checkpoint. The -OMAT-base models are trained on OMat24 (PBE / PBE+U, bulk inorganic, see the MACE page) with energies, forces and stresses weighted equally; -OMAT variants are then fine-tuned with a larger energy weight, and -OAM variants are further fine-tuned on MPtrj and a subset of Alexandria (Lysogorskiy et al., 2025).
  • Forces are the gradient of the energy (conservative); the models are evaluated in the precision of the exported build listed below.
  • What this means for you. Same reference level and the same blind spots as the other OMat24 models (no dispersion, bulk-only training). 1L is the choice when speed matters most, 2L when accuracy does. Check the model distribution page for its license.

Install

Environment Python Runtime stack
.venv-grace 3.10, 3.11, 3.12, 3.13 tensorpotential==0.6.0 (tensorflow 2.20.0)
./scripts/install_mliport.sh --engines grace --device auto

Pinned profiles

Profile Identity dtype task head model SHA-256
grace_omat GRACE-2L-OMAT-medium-base (upstream default precision build) upstream/model-defined bulk None d497af4c7d9a...
grace_omat_fp64 GRACE-2L-OMAT-medium-base fp64 build float64 bulk None 53b945efce37...

Run it

.venv-grace/bin/mliport sp structure.cif \
  --model /path/to/saved_model --model-type grace

Official exported SavedModels come from the upstream AMS-ICAMS-RUB release channels; point --model (MODEL_PATH) at the directory you unpack.

Things to know

  • No heads. GRACE models do not expose an mliport-level head; --task is the explicit periodic-boundary semantic (bulk or molecule).
  • GPU isolation. TensorFlow's GPU visibility has to be fixed per process. The CLI restarts a command with CUDA_VISIBLE_DEVICES when it is unset; the Python API and the queue worker need the same isolation.
  • GPU memory and neighbor lists are tunable. TensorFlow normally reserves nearly all visible memory, so memory growth is on by default (--gpu-memory-growth, --gpu-memory-limit-mb). A Verlet-style neighbor cache (--neighbor-cache, --neighbor-skin, default skin 1.5 Å) is on by default and rebuilds the table when an atom moves more than half the skin.
  • Precision follows the model: the pinned profiles include the upstream default build and an fp64 build, as listed above.
  • Stress depends on the checkpoint.

mliport is MIT-licensed. The upstream software and model must be cited separately; see Citations.