Standardize on CHGNet-PES-MatPES-*-1M-2026.9 and refresh MatGL recommendations

[bowen-bd]

Summary

Standardizes AtomisticSkills on the current CHGNet checkpoints, CHGNet-PES-MatPES-PBE-1M-2026.9 and CHGNet-PES-MatPES-r2SCAN-1M-2026.9, which are already the matgl 4.1 and matgl_server defaults. CHGNet-MPtrj-2023.12.1-2.7M-PES no longer loads in matgl 4, and its data and mentions are removed. Examples, reference data and docs that used the earlier MatPES checkpoint (2025.2.10) now use 1M-2026.9; the code still accepts the older names.

Model lists and recommendations now point to the current MatGL checkpoints (CHGNet 1M-2026.9 and TensorNet). This covers the MCP load_model description, the environment rules, the skill standards, the developer guide, the site and the skill docs.

Changes

• Defaults: MatGLWrapper defaults to CHGNet-PES-MatPES-PBE-1M-2026.9; the CHGNet tests use it.
• mat-elemental-energies: new 89-element libraries for both 1M-2026.9 checkpoints, using the library protocol. Neither collapses the O₂ molecular crystal. Superseded library files are removed. The example now compares CHGNet r2SCAN with MP r2SCAN: MAE 38.1 meV/atom.
• mat-pourbaix-diagram:
  • H₂O references for the new CHGNet.
  • Na–Ni example re-run with CHGNet r2SCAN: NiO formation energy −1.18 eV/atom, against MP r2SCAN −1.10 and GGA+U −1.22.
• ml-mlip-benchmark: CHGNet example re-run on the same 100 MP r2SCAN structures. Energy MAE (meV/atom): MACE-MATPES-R2SCAN-0 73.7, CHGNet r2SCAN 1M 77.5, TensorNet r2SCAN 88.3.
• ml-mlip-nvalchemi: batched vs sequential static inference re-measured for both new checkpoints. ΔF = 0, ΔE ≤ 9.5e-7 eV; speedup 3.4× at N = 5 and 10.7–10.9× at N = 20.
• ml-mlip-speed: stored benchmark updated, and a measured CHGNet 1M entry added (0.107 ms/atom at 2000 atoms); plots regenerated.
• ml-property-predictor:
  • The MatGL path uses a CHGNet backbone, and a dataset script is added.
  • The bulk-modulus example was re-run: validation MAE 20.8 GPa against a 56.9 GPa mean-value baseline; 19.0 GPa (R² = 0.88) on the held-out test split.
  • Model checkpoints are no longer committed in the example folder.
• ml-matgl-finetune: WBM example re-run with the new CHGNet (10 epochs, frozen backbone). Energy RMSE goes from 382.2 to 328.8 meV/atom, force MAE from 154.6 to 146.3 meV/Å and stress MAE from 4.97 to 3.66 meV/ų; energy MAE rises from 241.3 to 250.6 meV/atom. A fresh-process reload of the saved checkpoint reproduces the fine-tuned prediction.
• mat-sample-pes-by-md: LiMnO₂ example re-run with the new CHGNet.
• mat-lammps-md: the input file and run-script default now use the new CHGNet. Not re-run: it needs a MatGL ML-IAP LAMMPS build.
• mat-dft-mixing-functionals: the "no correction needed" case uses the current CHGNet.
• Warnings and links: warnings about MPtrj-trained models now cite MACE-MP. Fixes a broken link from ml-foundation-potentials to mat-dft-mixing-functionals.

Fixes found while re-running

• mat-pourbaix-diagram:
  • calculate_pourbaix.py looked for elemental references under the pre-rename skills/elemental-energies/ path.
  • get_pourbaix_structures.py failed on the EntryID objects returned by newer mp_api.
• ml-matgl-finetune: train_matgl.py is ported to matgl 4.1. This covers the MGLDataset arguments, the device of element_refs, memory pinning, and inference_mode=False for force validation.
• ml-property-predictor:
  • run_matgl.py's reload check read a result key that does not exist.
  • Per-atom normalization broadcast predictions of shape [B] against atom counts of shape [B, 1] into a [B, B] matrix, so the training loss and validation metrics were wrong for batches with unequal atom counts. Predictions are now 1-D, and validation errors are weighted by structure (regression test added).
  • A failed pretrained-model load now raises an error instead of falling back to a freshly initialized model.
• ml-matgl-finetune metrics and checkpoint:
  • The history callback divided matgl's per-atom energy MAE/RMSE by the mean atom count a second time.
  • The saved checkpoint now uses model_state_dict, the key MatGLWrapper.load_checkpoint() reads.

Testing

• venv/run cpu python -m pytest tests/test_*.py -q: 1082 passed, 5 skipped.
• venv/run mlip python -m pytest tests/matgl -q (CPU): 47 passed, including the new batching test.
• Pre-commit hooks and tools/check_skill_commands.py pass.
• Re-run on an NVIDIA GB10: the elemental-energy libraries, the Pourbaix, benchmark, nvalchemi, speed, property-predictor, fine-tuning and PES-sampling examples.

🤖 Generated with Claude Code