Native DeePMD3.2 electronic population example from24real prior NH3 RHF/6-31G(d) Mulliken labels.16train/4validation/4test originals,three rigid poses each. Per-atom alpha/beta electron counts;not Bader charges or magnetic-material data.
bash analyze.sh;bash labels.sh;bash plot.sh;bash run.sh(defaultP100)or bash run.sh v100. One ordinary Slurm GPU job runs3000native init-model updates from provided RHF symmetric initial weights,freezes,and compares native DeepPopulation GPU/CPU predictions to actual labels. No new quantum,no MD,no inference-time projection,no shared patch. The starting model has prior training;3000is additional optimization,not random initialization.
Current dp test has no DeepPopulation tester;use the actual native API in the supplied workflow. The native loss/data reader works without a runtime alias. Closed-shell channel symmetry is specific to this RHF example and must not be imposed on open-shell/magnetic systems.
DeePMD3.2,PyTorch2.12,GCC13.3,CUDA12.9;prior Gaussian16C.01AVX quantum labels;ASE3.29 environment supplies NumPy/Matplotlib. Model does not provide an energy/force MD potential.
