DeePMD-kit

Start from a pretrained Deep Potential model, fine-tune it for your system, and deploy it at simulation scale.

Pretrained models · Fine-tuning · Documentation · Quick start · Model guide · Tutorials · Examples · Releases

[!IMPORTANT] A pretrained model can be your starting point, not just your end result. Download a built-in pretrained DPA4 checkpoint, fine-tune the full model for your system, then test, export, and deploy it through the same DeePMD-kit workflow.

DeePMD-kit turns quantum-mechanical reference data into fast, scalable interatomic potentials. Use it across molecular and materials science—from finite molecules and covalent systems to periodic solids and metals—and scale from laptop fine-tuning to distributed training and MPI-parallel molecular dynamics.

The DPA4 model family traces a Pareto frontier across Matbench Discovery CPS and saturated inference throughput.

⚡ Why DeePMD-kit

[!TIP] On supported descriptors and workloads, model compression can deliver more than 10× inference speedup and reduce memory usage by as much as 20×. Actual gains depend on the model, system, and hardware.

Backend and interface support varies by model and feature. The web documentation marks compatibility and limitations on each feature page.

🧭 Two starting points, one path to dynamics

flowchart LR
    A["Pretrained DPA4 model"] --> C["Fine-tune on target data"]
    B["Model configuration"] --> D["Train from scratch"]
    E["Target reference data"] --> C
    E --> D
    C --> F["Test, compress, export"]
    D --> F
    F --> G["Python and native APIs"]
    F --> H["Molecular dynamics"]

1. Choose a starting point: download a pretrained DPA4 checkpoint for adaptation, or configure DPA4 or DPA4C to train from scratch.
2. Prepare target data in DeePMD's NumPy format or convert structures and trajectories with dpdata.
3. Fine-tune or train: adapt the full pretrained DPA4 model, or optimize a new DPA4 or DPA4C model with single-task, multi-task, and distributed training workflows.
4. Validate and export with dp test, dp freeze, backend conversion, embedding extraction, and supported compression paths.
5. Run simulation through Python or native APIs, or load the model into a supported molecular-dynamics engine.

🚀 Start in minutes

DeePMD-kit requires Python 3.10 or later. The fastest installation path is:

curl -fsSL https://dp1s.deepmodeling.com | bash
dp --version
dp -h

The installation guide covers pip, conda-forge, containers, offline packages, GPU builds, LAMMPS, i-PI, and source installation.

Fine-tune a pretrained DPA4 model

Download a built-in checkpoint, start from its matching released training configuration, and fine-tune it on your target data. This example uses DPA4-Neo, one of the recommended general-purpose sizes:

dp pretrained download DPA4-Neo-OMat24-v20260805
curl -fsSL \
    https://huggingface.co/deepmodelingcommunity/DPA4-OMat24/resolve/main/DPA4-Neo-OMat24-v20260805.json \
    -o input_finetune.json

The DPA4 OMat24 release provides Nano, Mini, Neo, Air, and Plus checkpoints together with their matching training configurations. The downloaded input_finetune.json matches the Neo checkpoint above; for another size or version, use the correspondingly named JSON file. Keep its complete model section unchanged, including the full-periodic-table type_map; replace the training and validation data, and use a smaller learning rate for fine-tuning. Then run:

dp --pt train input_finetune.json \
    --finetune ~/.cache/deepmd/pretrained/models/DPA4-Neo-OMat24-v20260805.pt

These are PyTorch single-task checkpoints, so no model branch selection is needed. They target inorganic materials in the OMat24 chemical space; validate accuracy before using them outside that domain.

The fine-tuning guide covers full-model adaptation. DPA-ADAPT reuses supported pretrained DPA representations for downstream property-prediction tasks.

Pretrained model names can also be resolved and cached automatically by Python:

from deepmd.infer import DeepPot

potential = DeepPot("DPA4-Neo-OMat24-v20260805")

Train a model from scratch

Training from scratch remains a first-class workflow for new architectures, fully custom systems, and physical targets without a suitable pretrained checkpoint. Clone the examples and start with the compact water system:

git clone https://github.com/deepmodeling/deepmd-kit.git
cd deepmd-kit/examples/water/dpa4

# Accuracy-first DPA4 model
dp --pt train input.json

# Or the throughput-first DPA4C model
cd ../dpa4c
dp --pt-expt train input.json

Ready-to-run inputs include:

• DPA4 water training
• DPA4C high-throughput water training
• DPA4 multi-task training
• DPA-ADAPT property prediction

For a guided end-to-end example, open the web quick-start notebook.

🧠 Choose a model family

For conservative energy/force interatomic potentials, start with the DPA4 family. The choice between its two primary models follows the constraint that matters most for your workload:

DPA4 uses the PyTorch backend (dp --pt). DPA4C currently uses the PyTorch Exportable backend (dp --pt-expt); its compressed CUDA path requires float32.

For other physical targets, use the model guide to select a compatible model and backend. The guide also compares data formats, precision, compression, and deployment constraints.

For energy/force potentials, DPA4 and DPA4C span accuracy–throughput trade-offs for different deployment budgets.

🔬 Go beyond conventional force fields

Explore the complete set of models and physical targets in the web documentation.

🔌 Deploy into the scientific ecosystem

Inference interfaces

• Python
• C and C++
• Node.js
• Model embedding export

Simulation and workflow integrations

• LAMMPS, i-PI, ASE, JAX MD, and nvalchemi
• Ecosystem integrations for OpenMM, Amber, CP2K, GROMACS, ABACUS, DP-GEN, and MLatom
• External MACE and NequIP models through the DeePMD-GNN plugin

See the integration hub for maintained interfaces, third-party projects, supported scope, and installation guidance.

The native C and C++ interfaces load machine-learning backends as runtime plugins. Applications can therefore open the backend required by a model without directly linking every framework.

[!NOTE] Working with an AI coding or scientific agent? Start with Install with an AI agent, or browse the official Agent Skills for model selection, training, fine-tuning, Python inference, and LAMMPS workflows.

npx -y skills add https://github.com/deepmodeling/deepmd-kit/tree/master/skills \
    --skill deepmd-install -y

If direct GitHub access fails, clone the official Gitee mirror and install from the local checkout:

git clone --depth 1 \
    https://gitee.com/deepmodeling/deepmd-kit.git \
    deepmd-kit-skill-source
npx -y skills add ./deepmd-kit-skill-source/skills \
    --skill deepmd-install -y

📚 Documentation and community

• Read the full web documentation.
• Follow hands-on material in the DeepModeling tutorials.
• Browse examples for training, inference, and integrations.
• Ask questions or report problems in GitHub Issues.
• Join development through the contributing guide.

Citation

If DeePMD-kit contributes to published work, cite the general software paper that matches the version used and the method-specific papers listed in CITATIONS.bib:

• Wang et al., “DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics,” Computer Physics Communications 228 (2018), 178–184 (describes the initial version).
• Zeng et al., “DeePMD-kit v2: A software package for Deep Potential models,” The Journal of Chemical Physics 159 (2023), 054801 (covers features until v2.2.3).
• Zeng et al., “DeePMD-kit v3: A Multiple-Backend Framework for Machine Learning Potentials,” Journal of Chemical Theory and Computation 21 (2025), 4375–4385 (covers features until v3.0).

License

DeePMD-kit is licensed under the GNU Lesser General Public License v3.0 or later.