H2 real HF/6-31G(d) energy/force/Hessian training example.
10actual fixed-geometry analytic frequency calculations plus24independent displaced-force calculations,all real Gaussian/formchk;6train/2validation/2test originals,30rigid poses. Full6x6 Hessian labels in eV/Angstrom^2;not packed triangular data.
bash analyze.sh,bash labels.sh,bash run.sh(defaultP100),or bash run.sh v100. One ordinary Slurm GPU job trains one1500-step PyTorch model,tests the checkpoint,freezes/tests an energy-force graph,and evaluates CPU/GPU within allocation.
Use model.pt for Hessians;this installed TorchScript frozen_model.pth drops Hessian output but keeps matching energy and force. No atomic-energy labels fabricated,no dp test -a.
Fixed geometries are not optimized equilibrium states;do not interpret all Hessian eigenvalues as equilibrium vibration frequencies. No MD/transferability/production H2 accuracy claim.
