Simulation and Post-processing
The unified gpumdkit-skill connects GPUMDkit utilities, GPUMD and NEP input preparation, controlled execution, validation, and post-processing. Its main file selects focused references for each task, including requests such as "run an Arrhenius diffusion study" or "train a NEP model and validate it in GPUMD."
Decision boundary
An agent must ask before filling any unresolved scientific or execution choice. This includes the potential, species/type mapping, charge, temperatures, time step, ensemble, pressure, run length, output interval, fit policy, convergence criteria, scheduler resources, and overwrite behavior.
Preparing inputs does not authorize running GPUMD or NEP. Long calculations and scheduler submissions require an explicit request.
The skill bundles a portable, categorized summary of model.xyz, run.in, nep.in, GPUMD/NEP parameters, and output formats; users do not need a local GPUMD documentation tree. Local GPUMDkit and Python-script help govern post-processing commands. If parser behavior differs from the bundled snapshot, the agent must report the executable version and exact error, then request version-specific information instead of guessing.
One skill, focused references
Agents activate only gpumdkit-skill, then load the smallest relevant set from references/:
| Task | References |
|---|---|
| GPUMD simulation | gpumd.md, then the required gpumd-inputs.md, gpumd-setup.md, gpumd-ensembles.md, gpumd-computes.md, and gpumd-outputs.md catalogs |
| NEP training or prediction | nep.md, then nep-data.md, nep-parameters.md, and nep-outputs.md as needed |
| Arrhenius workflow | gpumd.md, arrhenius.md, calculators.md, and visualization.md |
| GPUMDkit utility | The matching conversion, sampling, workflow, calculator, analyzer, or visualization reference |
| GPUMDkit development | contributing.md plus the affected module reference |
This progressive-loading structure keeps discovery simple while avoiding the context cost of loading every module for every request.
End-to-end workflow
- Define the observable, deliverables, and acceptance criteria.
- Confirm the structure, potential/model, software version, and scientific parameters.
- Prepare inputs and check species, cell, PBC, groups, units, and command syntax.
- Present the resolved protocol and unresolved questions.
- Run only when explicitly requested, beginning with a short approved smoke test when appropriate.
- Check completion, thermodynamic stability, trajectory integrity, and observable-specific convergence.
- Run GPUMDkit post-processing and report units, fit choices, uncertainty, exclusions, and limitations.
Arrhenius example
Before setting up an Arrhenius study, confirm whether the target is diffusivity, ionic conductivity, or both; the mobile species and charge; temperature points; ensemble sequence; time step; production length; replicas; MSD sampling; and fit/uncertainty policy.
Use one integer-temperature directory per run:
project/
500K/{model.xyz,nep.txt,run.in}
600K/{model.xyz,nep.txt,run.in}
700K/{model.xyz,nep.txt,run.in}
MSD can come directly from GPUMD compute_msd or from an extxyz trajectory:
cd 500K
gpumdkit.sh -calc msd dump.xyz Li <frame_dt_fs> [max_corr_steps]
gpumdkit.sh -plt msd save
gpumdkit.sh -plt sdc save
frame_dt_fs is the time between stored trajectory frames:
After validating every temperature point, run from the parent directory:
The current Arrhenius scripts fit 40%-80% of each MSD series. This is an established GPUMDkit choice and should not be changed without maintainer approval. The fit must still be checked against the observed diffusive regime.
For one temperature, ionic conductivity can be calculated with a user-confirmed species and charge:
The automatic conductivity Arrhenius scripts have narrower assumptions:
They count only Li/Na, assume charge magnitude one, use the first temperature's composition/replication, and extrapolate the fit to 300 K. Do not use them as a general multivalent or mixed-ion workflow. Confirm the Nernst-Einstein approximation and any extrapolation with the user.
Skill discovery
The command prints the canonical skill path and installation hints. The single portable skill is located at ${GPUMDkit_path}/skills/gpumdkit-skill/; simulation details live in its references/gpumd.md, nep.md, and arrhenius.md files.