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.

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.

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"
}

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