NH3 HF/6-31G(d) quantum labels and TensorFlow DeePMD3.2.0 tensor training example.
16 actual original Gaussian geometries,48 rigid augmented frames; 8/4/4 original train/validation/test split. Provided labels avoid requiring a Gaussian license to retrain. No CHK/full FCHK/orbitals/program included.
Ordinary login user: bash analyze.sh; bash run.sh(default P100),or bash run.sh v100(V10016G). One GPU job trains three 3000-step models, freezes and tests each, then does CPU/GPU predictions inside the allocation. No energy-force potential,no MD,no IR/Raman spectra or material transferability claim.
N-only dipole assigned whole neutral molecule as a label convention,not a physical atomic partition. All-center polar uses global labels only. Original N-only polar models remain as failed-error controls.
