MACE0.3.16 optional native model/converter examples.
Python3.11, PyTorch2.8.0+cu126, e3nn0.4.4, ASE3.29. Optional graph_longrange0.4.0 MIT; cuequivariance/cuequivariance-torch0.6.1 Apache2.0. CPU naive fallback is used for CuEq; no accelerated-CUDA performance claim.
Run bash run.sh to draw actual reference data. Run bash setup.sh once to prepare only this project's .deps, then bash run.sh verify to evaluate fixed owned models and all fields. The default run does not train or run DFT/MD.
Native converters: bash convert.sh. New external-field training: sbatch train.sh, after reviewing train-input/train.yml and the accompanying tutorial. Existing native.log is protected from overwrite. Reusing the supplied training input is an example, not a guarantee of a useful model.
IMPORTANT actual scientific scope: PolarMACE's original held-test energy/force/dipole errors meet the tutorial targets, but its pinned real-space branch fails rotation covariance and does not recover the DFT polarizability. Full failed scientific checks are retained in reference-polar/VALIDATION.json. This owned model is supplied solely for reproducible method/diagnostic demonstrations, not production molecular dynamics. Do not equate low pointwise fitting errors with invariant or response-accurate potentials.
The H2O label source is completed CP2K2026.2 PBE/DZVP-MOLOPT-SR-GTH,600/50Ry,12A isolated box. Neutral geometry coordinates are shifted so oxygen is at origin, with labels unchanged. external_field is the potential gradient G in V/A; physical electric field is -G. total_spin is multiplicity=1, total_charge=0. Train16/valid3/test9 conditions; six held fields are amplitude extrapolations at a training geometry, not unseen-geometry validation.
Models and DFT numerical labels are owned teaching artifacts. No vendor foundation weights, VASP proprietary files, thirdparty source/binaries, credentials or internal execution receipts included. MACELES requires a separate LES dependency with CC-BY-NC-4.0 upstream; this package does not install or run it.
