NequIP & Allegro Foundation Potentials

These are 'large' NequIP/Allegro models, optimised for speed and accuracy with a greater priority placed on accuracy. The `MP` models are trained only on the `MPTrj` (~1.5M frames) datasets (i.e. matbench-discovery 'compliant'), and so are only recommended for benchmarking and not production work.

The `OAM` modles are pre-trained on the `OMat24` dataset (~101M frames), and fine-tuned on the `sAlex` (~10.5M frames) and `MPTrj` (~1.5M frames) datasets. These are the recommended `NequIP`/`Allegro` models for most applications in inorganic solids, having been trained on the largest available open-access datasets.

 

We find the NequIP OAM 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.

See `nequip.net` and `matbench-discovery` submission for further details – in particular, for details on including model accelerations, and training config files.

See https://nequip.readthedocs.io/en/latest/guide/training-techniques/fine_tuning.html for details on fine-tuning `NequIP`/`Allegro` models.

 

Latest version update (`0.1a`) was a manual fix to the packaged model modifiers to avoid issues with compiling/packaging models fine-tuned on these packaged foundation models (see https://github.com/mir-group/nequip/issues/572).