NepTrainKit

NepTrainKit is a desktop application for preparing, auditing, and visualizing training datasets for neuroevolution potentials (NEPs). It complements, rather than replaces, long-running GPUMD training or DFT calculations. Its focus is the repetitive work around those calculations: generating candidate structures, cleaning problematic samples, selecting representative configurations, and passing a well-prepared dataset back into the DFT and GPUMD workflow.

What you can do

• Make Dataset: generate strained, perturbed, defective, surface, doped, magnetic, and solvated candidate structures with composable cards.
• NEP Dataset Display: inspect structures, errors, and distributions; remove problematic samples; and export a clean subset.
• Representative selection: reduce a candidate pool to a smaller, more representative set with methods such as farthest point sampling (FPS).
• Training-result inspection: load NEP- and DeepMD-related outputs, locate high-error structures, and identify data gaps for the next iteration.
• Project tracking: use Data Management to record models, dataset paths, and notes across multiple iterations.

Installation

We recommend installing NepTrainKit in a dedicated Python environment. Python 3.10 through 3.13 is supported.

conda create -n nepkit python=3.10
conda activate nepkit
pip install NepTrainKit

After installation, launch the application with either command:

nepkit
# or
NepTrainKit

NEP compute backend

NepTrainKit does not compile the NEP compute backend into the application. pip installs the separate nep-adapters dependency:

Platform	Installed backend
macOS / Windows	CPU
Linux x86_64	CPU and CUDA in one wheel

The Linux CUDA path requires a compatible NVIDIA driver, but installing the wheel does not require a local CUDA toolkit or NVCC. Source builds and supported CUDA architectures are documented in the NEPAdapters repository.

After launching NepTrainKit, select Auto, CPU, or CUDA under Settings → NEP Backend. Auto uses CUDA when the installed wheel, driver, and model support it; otherwise NepTrainKit explains why it is continuing on CPU. Explicit CUDA requests fail instead of silently changing backend.

Settings → NEP Settings → NEP runtime updates can install a compatible nep-adapters update without replacing NepTrainKit. NepTrainKit also checks this runtime in the background whenever the app opens; no-update and network-failure cases stay silent, while an available update opens an installation prompt. Pip installations store the managed wheel under the user configuration directory; the Windows standalone build stores it under runtime/nep-adapters/versions beside NepTrainKit.exe. A wheel is activated only after SHA256 verification and a fresh-process health check, and takes effect after restart.

Confirm the installed runtime with:

python -c "import nep_adapters as n; print(n.backend_status('cpu')); print(n.backend_status('cuda'))"

Windows package

If you prefer not to compile NepTrainKit locally, download NepTrainKit.windows-x86_64.zip from GitHub Releases. This package requires 64-bit Windows on x86_64 hardware.

Documentation and support

• User documentation: neptrainkit.readthedocs.io
• Release notes: GitHub Releases
• Bug reports and feature requests: GitHub Issues
• Community: QQ group invitation

If this is your first time using NepTrainKit, begin with Quickstart and Cleaning candidate structures before DFT in the user documentation. If you already know which class of configurations you need, go directly to the Make Dataset card reference.

Citation

If NepTrainKit contributes to your research, please cite:

@article{CHEN2025109859,
title = {NepTrain and NepTrainKit: Automated active learning and visualization toolkit for neuroevolution potentials},
journal = {Computer Physics Communications},
volume = {317},
pages = {109859},
year = {2025},
issn = {0010-4655},
doi = {https://doi.org/10.1016/j.cpc.2025.109859},
url = {https://www.sciencedirect.com/science/article/pii/S0010465525003613},
author = {Chengbing Chen and Yutong Li and Rui Zhao and Zhoulin Liu and Zheyong Fan and Gang Tang and Zhiyong Wang},
}

License and third-party code

NepTrainKit is licensed under the GNU General Public License v3.0 or later. See LICENSE for details.

NEP computation is provided by the separate nep-adapters dependency. NepTrainKit no longer vendors the NEP_CPU or GPUMD backend source trees.

See THIRD_PARTY_NOTICES.md for the remaining attribution in this repository. The nep-adapters distribution carries its own backend source notices and licenses.