gpumd

name	gpumd
description	Route GPUMD requests to task-specific subskills. GPUMD is a GPU-accelerated molecular dynamics code that pairs naturally with NEP machine-learning potentials. Use when the user asks for GPUMD MD, `model.xyz`, `run.in`, harmonic phonons via `compute_phonon`, thermal transport via EMD/HNEMD/NEMD, diffusion and ionic conductivity, elastic constants, melting, friction, or deposition workflows.
compatibility	Requires a runnable GPUMD environment. The user must supply the `gpumd` executable path, HPC module, or container entry point. A compatible potential file (`nep.txt`, `nep89_*.txt`, Tersoff, etc.) is required.
license	GPL-3.0-only
metadata	{"author":"Jhin","version":"0.2.0","repository":"https://github.com/brucefan1983/GPUMD","documentation":"https://gpumd.org/"}

GPUMD Task Router

Use this skill as the top-level routing layer for GPUMD work. GPUMD is analogous to LAMMPS but focused on GPU execution and tight integration with the NEP potential family. If the task is about fitting a NEP potential rather than running MD with one, route to machine-learning-potentials/nep-gpumd instead.

Subskill map

Subskill	Use when the target observable is
gpumd/md	equilibrium MD, ensemble setup, thermo.out, general trajectory output
gpumd/phonon	harmonic phonons via compute_phonon, kpoints.in, omega2.out
gpumd/transport	thermal conductivity via EMD / HNEMD / NEMD / SHC
gpumd/diffusion	self-diffusion, ionic conductivity, viscosity via compute_msd/_sdc/_viscosity
gpumd/elastic	elastic constants via the strain-fluctuation method
gpumd/mechanics	friction, deposition, impact, and group-based interface workflows

Agent responsibilities

1. Confirm the execution mode.
   • Ask the user for the exact gpumd command, for example gpumd, gpumd-3.9, an HPC module load, or srun -n 1 --gpus=1 gpumd.
   • Do not invent a binary or module name.
2. Classify the request into one subskill path. If the request spans more than one area, start from the dominant observable and only open the extra subskill files that are needed.
3. Collect the minimum shared context before dispatching:
   • the atomistic structure (preferred: model.xyz in extxyz format)
   • the potential file (for example nep.txt, nep89_20250409.txt, Si_2022_NEP3_4body.txt, or an upstream Tersoff file)
   • the target ensemble / temperature / pressure
   • the target observable
4. Write the input files yourself instead of asking the user to hand-write them. Keep the example readable and explain every block.
5. Validate structures before trusting downstream runs:

   python scripts/validate_extxyz_headers.py model.xyz --mode model
   

6. Report clearly which command was run, which files were used, where outputs were written, and which sanity checks should be performed next.

Shared policy for all GPUMD subskills

• Do not invent missing physical parameters.
• Keep the timestep, ensemble, and target observable physically consistent.
• Treat transport, elastic, and diffusion results as provisional until size, sampling, and method-specific convergence are explicitly discussed.
• For low-dimensional systems (monolayers, ribbons, slabs) require an explicit thickness convention before reporting bulk-like intensive quantities.
• For restarts from another NEP or GPUMD version, re-check keyword syntax against the current docs at https://gpumd.org/.
• Group prerequisites: several subskills (transport/SHC, mechanics, diffusion/ionic conductivity) depend on group labels defined in model.xyz via group:I:M columns in the Properties header. Before dispatching to those subskills, confirm whether the user's model.xyz already carries the required group definitions. Use the add_groups tool from tools/gpumd-tools to add them if needed.

Execution templates

Direct local run

gpumd < run.in | tee gpumd.log

Many builds take input from stdin. Some custom wrappers accept gpumd run.in directly; the user should confirm.

HPC / Slurm

Ask the user for the exact module and wrapper. A typical pattern looks like:

module load gpumd/3.9.5 cuda/12.2
srun -n 1 --gpus=1 gpumd < run.in > gpumd.log

Do not use this exact command without confirming the module names.

Smoke test before production

Before a long transport or phonon run, always do a short stability test:

run 100

This catches:

• missing potential file
• malformed model.xyz
• immediately divergent dynamics (bad structure, wrong species order, wrong potential)
• wrong timestep for the chosen system

Resource map

• Core files and ensembles: references/core-files-and-ensembles.md
• GPUMD keyword cheatsheet: references/gpumd-keyword-cheatsheet.md
• Phonon workflow: references/phonon-workflow.md
• Thermal transport workflow: references/thermal-transport-workflow.md
• Diffusion, viscosity, ionic conductivity: references/diffusion-viscosity-ionic.md
• Elastic constants: references/elastic-constants.md
• Mechanics (friction / deposition / impact): references/mechanics-deposition-friction.md
• Tutorial map: references/tutorial-map.md
• Reusable templates: assets/examples/
• Deterministic helpers: scripts/

Cross-skill pointers

• Need to fit a NEP potential first? → machine-learning-potentials/nep-gpumd
• Need to convert VASP / CP2K / QE output into NEP extxyz? → tools/gpumd-tools
• Need to generate labeled DFT data before training a NEP? → refer to the quantum-chemistry DFT skills (dft-vasp, dft-cp2k, dft-qe, dft-siesta, dft-abinit) for static/relax single-point workflows.
• Need to submit GPUMD jobs to HPC queues (Slurm, PBS, LSF)? → dpdisp-submit can manage job submission, monitoring, and auto-recovery.
• Need to convert CIF / POSCAR / other crystal formats to model.xyz? → pymatgen-structure can read and manipulate crystal structures, then export to extxyz via ASE for use as GPUMD input.