GPUMDkit: charge/BEC, diffusion, NEMD, phonon and XRD practical examples

module load gpumdkit/1.5.8
bash run.sh

The default replays native tools from actual reference outputs in a new work directory. It does not start DFT or MD. LAST-RUN.txt gives the output path.
python check.py WORK_DIRECTORY verifies numerical values, figures and preserved inputs.
python figures.py redraws the five publication figures from references.

Optional new simulations: copy one reference/md/CASE to a new directory, retain only model.xyz, potential.txt and run.in, then run gpumd/5.9.1-p100 within allocated resources. For qNEP retain train.xyz, test.xyz and nep.in only, then run nep; seed53 is fixed. Never reuse old loss.out or optimizer state accidentally. CPU-only phonon calculation requires an ASE/Calorine/Phonopy environment; supplied reference data can be plotted directly.

ZnS has 4 training frames and 1 monitored reference frame. Finite short trajectories, high-temperature fixed-density Si, one seed per diffusion temperature and two NEMD seeds are teaching examples. A 5000K piecewise split is chosen only to demonstrate syntax, not evidence of a phase transition. Single-run error bars of zero from the native NEMD plot are not zero physical uncertainty. Native NEMD uses an endpoint estimate; the tutorial also compares a linear-fit estimate and its sampling interval.
