GPUMD periodic Si/Ge interface NEMD tutorial

1. Check package: sha256sum -c MANIFEST.sha256
2. Reproduce reference figures: bash view.sh
3. Prepare new inputs: python3 prepare.py
4. Enter a case, inspect config.json and job.slurm, then sbatch job.slurm.
   A compatible GPU partition and account are required.
5. After jobs: python3 analyze.py --cases cases --output analysis-new

Units: Angstrom, ps, K, eV. Model: coherent strained diamond cells,
a=5.5 A, Ny=Nz=3, Nx=16 or24, fully periodic. Si/Ge occupies equal
halves. Pure-Si cells are geometric fitting controls, not a chemical interface.
Hot group6, cold19: targets330/270K. 26 groups; all are nonempty.
Production400ps, exclude first100ps in analysis. Two seeds per platform.
Two periodic heat paths: flux = mean(abs(bath slopes))/(2*cross-section).
Heat-bath energies are energy transferred FROM system TO reservoirs.
Keep signed temperature jumps and resistances; two-seed ranges are not95%CI.
Short strained nanoscale examples are not converged bulk/interface properties.

Synchronous energies use: compute 0 1000 1 temperature potential.
Columns1-26: current group temperature(K).
Columns27-52: current group total potential energy(eV), NOT per atom.
Columns53-54: current cumulative reservoir energies(eV).
For each row, KE=sum(1.5*Ngroup*8.617343e-5*Tgroup); PE=sum(columns27-52).
The heat-BDP build used here does not refresh dump_thermo energy;
thermo.out is retained unchanged but its stale KE/PE are not used.
Do not combine averaged group energies with endpoint reservoir energies.

final.xyz preserves the exact last native frame; full native trajectories
are kept privately. No licensed VASP files or executables are included.
Official potentials are reproduced unchanged with their original attribution.
Sources: https://www.gpumd.org/gpumd/input_parameters/ensemble_heat.html
https://www.gpumd.org/gpumd/input_parameters/compute.html

The final.xyz global energy/stress metadata uses the same stale thermo buffer.
Use its coordinates, forces and velocities; use compute.out for current PE.
Do not use this final.xyz global energy as a training label.

For a fresh short trajectory with bath heat imbalance, increase the production
length rather than changing seed to select a favorable run. Set config.json
production_steps=1200000 (1.2ns); initial warmup and numerical gates unchanged.
Reference16 cases are400ps. The fresh final-ZIP V100 replay uses1.2ns.
