01
Format Conversion
Convert VASP, CP2K, ABACUS, LAMMPS, CIF, etc. to extended XYZ, and back again when you need it.
-out2xyz-cp2k2xyz-cif2exyz-lmp2exyz-xyz2dp
Read the docs
PaperPublished in MGE Advances 2026
GPUMDkit
A user-friendly toolkit for GPUMD and NEP
Format conversion, structure sampling, workflow automation, calculators, analyzers and publication-ready plots, all behind one command. Use the interactive menu or script everything from the command line.
$
conda create -n gpumdkit -c gpumdkit -c conda-forge gpumdkit
move to disturb · click to scatter
0+
interactive tools
0
command-line options
0
plot types
0
Python scripts
0
contributors
0
publications using it
Features
Everything around GPUMD & NEP
Six modules cover the entire machine-learned potential workflow: from raw DFT output to training data, from MD trajectories to figures for your paper.
02
Structure Sampling
Uniform, random and farthest-point sampling, perturbations and active-learning picks.
Read the docs03
Workflows
Batch SCF set-ups for VASP and CP2K, and MD sampling with GPUMD or LAMMPS.
Read the docs04
Calculators
Ionic conductivity, NEB, descriptors, MSD, XRD and phonon bands with a NEP model.
Read the docs05
Analyzers
Composition, outliers, charge balance, minimum distances and energy, force and virial ranges.
Read the docs06
Visualization
42 plot types for NEP training, MD, diffusion, phonons and heat transport. One flag each: gpumdkit.sh -plt train
Live demo
Try it right here
A browser replica of gpumdkit.sh with real output. Type a command, press Tab to complete flags, or pick one on the left.
Active learning
Close the loop, one command at a time
Every stage of a NEP active-learning cycle has a GPUMDkit tool: label, convert, train, explore, select, and go around again.
NEP · GPUMDactive-learning cycle
Web console
Your runs, in any browser
Start gpumdkit.sh -server in a work directory and follow remote jobs from your laptop.
- Live loss curves and NEP training progress, refreshed automatically.
- MD monitor for neighbor.out, run.in step progress, speed and ETA.
- One-click plot actions that match the files in each folder.
- Sandboxed file browser with create, upload and download.
gpumdkit.sh -server
loss.out · 100,000 / 100,000
Agent-ready
Built-in skills for AI coding agents
GPUMDkit ships English and Chinese Agent Skills, so your coding agent can drive GPUMD, NEP and GPUMDkit the same way you do, with guard rails.
Global install targets
Agent Skills
~/.agents/skillsOpenCode
~/.config/opencode/skillsClaude-compatible
~/.claude/skillsRun gpumdkit.sh -skill to print these paths and ready-to-paste ln -s commands. Project-local folders such as .agents/skills work too.
01
Knows every command
Routed references for conversion, sampling, workflows, GPUMD, NEP and plotting.
02
Asks before guessing
Unresolved scientific choices go back to you; parameters are never invented.
03
Never runs jobs alone
No simulations, training or scheduler jobs without an explicit request.
04
English & 中文
Two skill sets, installed globally or per project with one symlink.
Get started
Up and running in a minute
Install with Conda (recommended) or from source, then type gpumdkit.sh.
OPTION A · RECOMMENDED
Conda
conda create -n gpumdkit -c gpumdkit -c conda-forge gpumdkit conda activate gpumdkit # optional extras for some features pip install neptrain calorine
Update later with conda update -c gpumdkit -c conda-forge gpumdkit.
OPTION B · FROM SOURCE
Git
git clone https://github.com/zhyan0603/GPUMDkit.git cd GPUMDkit source ./install.sh # ready to go gpumdkit.sh
Pull the latest version of your branch with gpumdkit.sh -update.
Gallery
Publication-ready, out of the box
Every figure below came from a single gpumdkit.sh -plt command. Click one to see how.
Research
Trusted in published research
From superionic conductors to thermal transport, GPUMDkit is part of the toolchain behind recent papers.
0
publications and counting
Community
Built together, cited together
Contribute on GitHub
Report issues, share scripts and send pull requests. Every tool started as someone's everyday script.
—Stars
—Forks
21contributors
Join the QQ group
Ask questions and get answers straight from the developers.
825696376scan to join
Cite GPUMDkit
If GPUMDkit helps your research, please cite:
GPUMDkit: A User-Friendly Toolkit for GPUMD and NEP
MGE Advances, 2026, 4, e70074
Zihan Yan, Denan Li, Xin Wu, Zhoulin Liu, Chen Hua, Boyi Situ, Hao Yang, Shengjie Tang, Benrui Tang, Ziyang Wang, Shangzhao Yi, Huan Wang, Dian Huang, Ke Li, Qilin Guo, Zherui Chen, Ke Xu, Yanzhou Wang, Ziliang Wang, Gang Tang, Shi Liu, Zheyong Fan, Yizhou Zhu
Spend less time on scripts
More time on science
Open source under GPL-3.0, available on Conda, and growing with every release.