mir-group/NequIP-OAM-L
NequIP GNN foundational potential for materials.
This is a 'large' NequIP model, optimised for speed and accuracy with a greater priority placed on accuracy. This model is pre-trained on the OMat24 dataset (~101M frames), and fine-tuned on the sAlex (~10.5M frames) and MPTrj (~1.5M frames) datasets.
This is the recommended NequIP model for most applications in inorganic solids, having been trained on the largest available open-access datasets.
We find this model to currently lie on the upper-right quadrant of the Pareto front when compared to other leading foundation models (preprint incoming), showing an optimal balance of speed and accuracy.
Key model hyperparameters:
- Radial cutoff: 6 Å
- Maximum spherical harmonic rotation order (
l_max): 3 - Tensor features: 128 (l=0), 64 (l=1), 32 (l=2,3)
- Number of layers: 6
- ZBL: True
- Parity: False
See the nequip docs for details on fine-tuning NequIP/Allegro models.
Supported Elements
Supported Model Modifiers
Enable OpenEquivariance tensor product kernel for accelerated NequIP training and inference.
Modify per-type scales and shifts of a model.
The new scales and shifts should be provided as dicts. The keys must correspond to the type_names registered in the model being modified, and may not include all the possible type_names of the original model. For example, if one uses a pretrained model with 50 atom types, and seeks to only modify 3 per-atom shifts to be consistent with a fine-tuning dataset's DFT settings, one could use
shifts:
C: 1.23
H: 0.12
O: 2.13In this case, the per-type atomic energy shifts of the original model will be used for every other atom type, except for atom types with the new shifts specified.
Args:
-
scales: the new per-type atomic energy scales
-
shifts: the new per-type atomic energy shifts (e.g. isolated atom energies of a dataset used for fine-tuning)
-
scales_trainable (bool): whether the new scales are trainable
-
shifts_trainable (bool): whether the new shifts are trainable
Papers Using This Model
Know of a paper that uses this model? Click "Add a paper that uses this model" to let us know.
Model Information
Published Date
August 28, 2025
Tags
License
CC-BY-4.0
Architecture
NequIP GNN
Model Size
9.6M