DeePMD-kit plugin for various graph neural network models

deepmd-gnn is a DeePMD-kit plugin for various graph neural network (GNN) models, which connects DeePMD-kit and atomistic GNN packages by enabling GNN models in DeePMD-kit.

Supported packages and models include:

• MACE (PyTorch version)
• NequIP (PyTorch version)

After installing the plugin, you can train the GNN models using DeePMD-kit, run active learning cycles for the GNN models using DP-GEN, and perform simulations with MACE and NequIP models using molecular dynamic packages supported by DeePMD-kit, such as LAMMPS and AMBER. You can follow DeePMD-kit documentation to train the GNN models using its PyTorch backend, after using the specific model parameters.

Highlights

The official MACE and NequIP packages remain the reference implementations of their model architectures. DeePMD-GNN focuses on what is useful beyond the standalone official workflows: bringing those equivariant GNN models into the DeePMD-kit ecosystem for training, active learning, and production molecular simulations.

• Unified DeePMD-kit workflow: train and freeze MACE or NequIP models with dp --pt train and dp --pt freeze, using DeePMD-kit data pipelines, inputs, losses, and model serialization instead of maintaining a separate workflow for each upstream package.
• Broader MD engine access: run frozen GNN models through molecular dynamics interfaces supported by DeePMD-kit, including LAMMPS and AMBER, rather than relying only on the engines and plugins maintained by the official MACE or NequIP projects.
• Parallel periodic simulations: use DeePMD-kit's message-passing communication path for LAMMPS/MPI runs. MACE and NequIP models advertise their message passing to DeePMD-kit, so ghost-atom features can be exchanged between message-passing layers through border_op.
• Active learning and QM/MM integration: plug MACE or NequIP models into DP-GEN active learning loops and DeePMD-kit DPRc/AmberTools workflows through the same interface used by other DeePMD-kit models.
• Deployment-oriented MACE extras: export MACE models with the DeePMD-kit/LAMMPS PyTorch exportable backend (pt_expt), train with optional cuEquivariance acceleration, use DeePMD-kit's torch.compile path, and conservatively convert selected official MACE-OFF checkpoints for downstream validation in DeePMD-kit-supported engines.

Credits

If you use this software, please cite the following paper:

• Jinzhe Zeng, Timothy J. Giese, Duo Zhang, Han Wang, Darrin M. York, DeePMD-GNN: A DeePMD-kit Plugin for External Graph Neural Network Potentials, J. Chem. Inf. Model., 2025, 65, 7, 3154-3160, DOI: 10.1021/acs.jcim.4c02441.

Installation

Install via conda

If you are in a conda environment where DeePMD-kit is already installed from the conda-forge channel, you can use conda to install the DeePMD-GNN plugin:

conda install deepmd-gnn -c conda-forge

Build from source

First, clone this repository:

git clone https://gitlab.com/RutgersLBSR/deepmd-gnn
cd deepmd-gnn

Python interface plugin

Python 3.10 or above is required. A C++ compiler that supports C++ 17 is required. NVCC is required to build CUDA OPs.

pip install .

Only PyTorch 2.10 or above is supported.

C++ interface plugin

DeePMD-kit version should be v3.0.0b4 or later.

Follow DeePMD-kit documentation to install DeePMD-kit C++ interface with PyTorch backend support and other related MD packages. After that, you can build the plugin

# Assume libtorch has been contained in CMAKE_PREFIX_PATH
mkdir -p build
cd build
cmake .. -D CMAKE_INSTALL_PREFIX=/prefix/to/install
cmake --build . -j8
cmake --install .

libdeepmd_gnn.so will be installed into the directory you assign. When using any DeePMD-kit C++ interface, set the following environment variable in advance:

export DP_PLUGIN_PATH=/prefix/to/install/lib/libdeepmd_gnn.so

Usage

Follow Parameters section to prepare a DeePMD-kit input file.

dp --pt train input.json
dp --pt freeze

A frozen model file named frozen_model.pth will be generated. You can use it in the MD packages or other interfaces. For details, follow DeePMD-kit documentation.

Exporting MACE models with the PyTorch exportable backend

MACE models can also be trained and frozen with DeePMD-kit's PyTorch exportable backend (pt_expt):

dp --pt-expt train input.json
dp --pt-expt freeze -o frozen_model.pt2

Use this path when you need a .pt2 model for the DeePMD-kit/LAMMPS PyTorch exportable runtime. The pt_expt path currently supports MACE models; NequIP models should still use the regular --pt workflow.

The pt_expt workflow requires DeePMD-kit 3.2.0 or later. Multi-layer MACE models include the extra communication artifact needed by LAMMPS/MPI inside the exported .pt2 package, so the resulting file can be passed to LAMMPS in the same way as other DeePMD-kit frozen models.

Training MACE with cuEquivariance

MACE training can use cuEquivariance kernels when the installed MACE package provides CuEquivarianceConfig and the cuEquivariance runtime packages are available. Install the matching extra for the CUDA runtime used by PyTorch:

pip install "deepmd-gnn[cueq]"

The cueq extra installs the CUDA 12 cuEquivariance op package. For CUDA 11 environments, use deepmd-gnn[cueq-cu11] instead. When installing from a source checkout, use pip install ".[cueq]" or pip install ".[cueq-cu11]".

Then enable it in the model section:

"model": {
  "type": "mace",
  "enable_cueq": true
}

Then train with either PyTorch backend:

dp --pt train input.json
dp --pt-expt train input.json

On a single RTX 5090, using the water example data with sel = "auto", hidden_irreps = "128x0e + 128x1o", batch_size = 1, and disp_freq = 100, 2000-step runs were used to sample steady-state training speed:

backend	cuEquivariance	steady avg after step 300	speedup
dp --pt	no	0.1955 s/step	1.00x
dp --pt	yes	0.1061 s/step	1.84x
dp --pt-expt	no	0.2025 s/step	1.00x
dp --pt-expt	yes	0.1007 s/step	2.01x

The cuEquivariance effect is similar in both backends, but not bit-for-bit identical: in this run, the steady cuEquivariance speed differed by about 5%. The first cuEquivariance step includes a one-time kernel initialization cost (about 56 s in this run), which is amortized over long training jobs.

cuEquivariance accelerates training only. When freezing a checkpoint whose input had "enable_cueq": true, deepmd-gnn disables cuEquivariance for the frozen model and converts the MACE weights back to the e3nn format. This lets both dp --pt freeze and dp --pt-expt freeze export normally, but the frozen model does not use cuEquivariance kernels.

Training MACE with torch.compile

MACE training through the pt_expt backend can use DeePMD-kit's native torch.compile path. Enable it in the training section:

"training": {
  "enable_compile": true
}

Then run the regular exportable-backend training command:

dp --pt-expt train input.json

On a single RTX 5090, using the water example data with sel = "auto", hidden_irreps = "128x0e + 128x1o", batch_size = 1, and disp_freq = 100, a 2000-step run was used to sample the steady-state training speed:

mode	steady avg after step 200	speedup
eager pt_expt	0.2030 s/step	1.00x
torch.compile	0.1495 s/step	1.36x

The first compiled step includes a one-time Inductor trace/compile cost (97.17 s in this run), which is amortized over normal long training jobs.

Do not enable enable_cueq and enable_compile together for now. With cuEquivariance enabled, DeePMD-kit's symbolic make_fx compile trace fails in the cuEquivariance uniform_1d fake-tensor path with a data-dependent symbolic shape guard.

Running LAMMPS + GNN models with period boundary conditions

GNN models use message passing neural networks, so the neighbor list built with traditional cutoff radius will not work, since the ghost atoms also need to build neighbor list. The MACE and NequIP models work like regular DeePMD-kit message-passing models. No extra environment variable is needed when freezing the model. They advertise message passing to DeePMD-kit, so LAMMPS can use the regular neighbor list with a cutoff radius of $r_c$ and DeePMD-kit communicates the ghost-atom features through MPI between message-passing layers. This requires a DeePMD-kit build whose LAMMPS/PyTorch interface provides the message-passing communication op, border_op.

When border_op communication is available, DeePMD-kit/LAMMPS uses that MPI-capable path. The atom-map path is a serial fallback for runs that do not receive a DeePMD message-passing comm_dict; it maps ghost atoms to their corresponding real atoms on the same process and is not a substitute for MPI communication. Request the mapping when using this fallback in LAMMPS (also requires DeePMD-kit v3.0.0rc0 or above).

atom_modify map array

Parameters

MACE

To use the MACE model, set "type": "mace" in the model section of the training script. Below is default values for the MACE model, most of which follows default values in the MACE package:

"model": {
  "type": "mace",
  "type_map": [
    "O",
    "H"
  ],
  "r_max": 5.0,
  "sel": "auto",
  "num_radial_basis": 8,
  "num_cutoff_basis": 5,
  "max_ell": 3,
  "interaction": "RealAgnosticResidualInteractionBlock",
  "num_interactions": 2,
  "hidden_irreps": "128x0e + 128x1o",
  "pair_repulsion": false,
  "distance_transform": "None",
  "correlation": 3,
  "gate": "silu",
  "MLP_irreps": "16x0e",
  "radial_type": "bessel",
  "radial_MLP": [64, 64, 64],
  "std": 1.0,
  "precision": "float32",
  "enable_cueq": false
}

NequIP

"model": {
  "type": "nequip",
  "type_map": [
    "O",
    "H"
  ],
  "r_max": 5.0,
  "sel": "auto",
  "num_layers": 4,
  "l_max": 2,
  "num_features": 32,
  "nonlinearity_type": "gate",
  "parity": true,
  "num_basis": 8,
  "BesselBasis_trainable": true,
  "PolynomialCutoff_p": 6,
  "invariant_layers": 2,
  "invariant_neurons": 64,
  "use_sc": true,
  "irreps_edge_sh": "0e + 1e",
  "feature_irreps_hidden": "32x0o + 32x0e + 32x1o + 32x1e",
  "chemical_embedding_irreps_out": "32x0e",
  "conv_to_output_hidden_irreps_out": "16x0e",
  "precision": "float32"
}

DPRc support

In deepmd-gnn, the GNN model can be used in a DPRc way. Type maps that starts with m (such as mH) or OW or HW will be recognized as MM types. Two MM atoms will not build edges with each other. Such GNN+DPRc model can be directly used in AmberTools24.

Conservative MACE-OFF checkpoint loading

deepmd_gnn.mace_off provides a conservative bridge from selected official MACE-OFF checkpoints into DeePMD-GNN. The goal is to recover the parts that are really stored in the checkpoint while avoiding guesses about wrapper-only runtime semantics.

In practice, this path intentionally does less inference than the broader DPRc/QM/MM machinery:

• it infers ordinary element type_map entries from checkpoint atomic_numbers
• it restores checkpoint-side model semantics such as avg_num_neighbors from the original ScaleShiftMACE object instead of reusing DeePMD's runtime neighbor cap
• it requires an explicit sel value, because sel is a DeePMD runtime neighbor-list cap and is not stored in the MACE-OFF checkpoint
• it does not infer DPRc/MM labels such as mH, mC, HW, or OW
• load_mace_off_model() keeps the original eager MACE checkpoint model inside the wrapper for closer native/wrapper parity, while convert_mace_off_to_deepmd() scripts the model only at export time
• it uses torch.load(..., weights_only=False) to recover the original ScaleShiftMACE object, so callers should only use trusted checkpoints

Example:

deepmd-gnn mace-off convert mace_off23_small_dp.pt --model off23_small --sel 64

This produces a scripted DeePMD-GNN wrapper. That serialization step is useful, but it is still a good idea to validate the final downstream workflow you care about (for example in LAMMPS or AMBER).

Examples

• examples/water
• examples/dprc