DP-GEN practical tutorial, Si2/Si8, actual Yuansuan HPC calculations

Quickly recreate figures and CSV from real reference data:
  bash run.sh
This replot does not start DFT, training or MD.

To run the full short native workflow in your own shared cluster directory:
  bash prepare.sh
  bash initialize.sh
  bash workflow.sh
  bash status.sh

prepare.sh generates the authorized Si POTCAR using the installed VASPKIT103
and checks it matches the reference dataset. No POTCAR is distributed here.
The controller runs in your own directory on login01; every train/explore/label
stage is submitted through Slurm using the matching P100 modules. Do not run
the stage executables directly on a login node.

initialize.sh performs native init_bulk stages1/2 with skip_relax=true: structure
scaling and perturbation only. It does not relax structures or run AIMD.
The main workflow uses the exact supplied Si8 starting structure and fixed seed.
32 training /8 validation /8 independent test geometries (2/8 atoms) are actual
PBE DFT labels. Four TF models train1000steps, four50-step NPT conditions explore
300/600K and0/10000bar, up to4 candidates are labeled with VASP, then models
are retrained. Re-running workflow.sh uses native record.dpgen to resume.

Run actual model-deviation CLI + independent four-model checks through Slurm:
  bash compare.sh
Run independent native first-principles backend examples:
  bash backend.sh abacus
  bash backend.sh cp2k
These scripts generate inputs using DP-GEN and submit matching V10016 jobs.
Their labels retain code-specific energy references and are not mixed into
the VASP-only reference training dataset. CPU/GPU builds or arbitrary MPI
libraries must not be mixed in one shell; helpers load the matching modules.

This is a workflow teaching dataset, not a production Si potential. The original
short NPT has substantial volume drift: target temperature/pressure are not
proof of equilibration. The separate virial-loss-weight10 control improves
stress RMSE15.75->0.73GPa but worsens force RMSE and still drifts18.9% in volume.
Every result remains in the reference tables; neither model is validated for
equilibrium volume, thermal expansion or transport properties.

The original44 saved frames have43 candidates,1 above the force-disagreement
upper limit,0 accurate frames using the explicitly chosen teaching thresholds
0.001 and0.5eV/Angstrom. These are not universal material thresholds. Disagreement
is not DFT error. One configured exploration round finishing is not convergence.

Units: energy=eV/configuration (VASP free energy F), forces=eV/Angstrom,
virial=eV=-volume*ASE positive-tension stress, positions/cell=Angstrom.
Six high-resolution plots and seven CSV tables reproduce true native data.
Original six-significant-digit LAMMPS dumps and separate16digit precision controls
are kept distinct. No licensed SCF raw output, binary applications or private logs.
