v0.15.0: nequip.net

[cw-tan]

nequip.net is now online

It hosts our new NequIP and Allegro foundation potentials that rank among the most accurate models on the Matbench Discovery leaderboard. These include compliant models trained solely on MPTrj and non-compliant ones trained on OMat24.

Downloading and Using Pretrained Potentials

The packaged model files can be downloaded from nequip.net and used in typical NequIP workflows. Additionally, we provide convenience commands to auto-download and compile the models for inference in our ASE and LAMMPS integrations. Examples of these commands can be found on nequip.net when clicking into each model's webpage. The following is one such example for AOT Inductor compilation for ASE inference.

nequip-compile \
  nequip.net:mir-group/NequIP-OAM-L:0.1 \
  mir-group__NequIP-OAM-L__0.1.nequip.pt2 \
  --mode aotinductor \
  --device cuda \
  --target ase

---

[0.15.0] CHANGELOG

Added

• Possible to now use ChainedScheduler and SequentialLR for LR scheduling
• Auto-download capabilities for pretrained potentials from nequip.net

Changed

• Loosen LAMMPS ML-IAP constraints of using absolute paths and ensuring that the original checkpoint or package file is present in the same exact location during ML-IAP runtime

Fixed

• OpenEquivariance accelerated compiled training with PyTorch 2.8
• NequIPCalculator usage restrictions are loosened, e.g. when mixing ASE calculators

---

This discussion was created from the release v0.15.0: nequip.net.