LAMMPS: ML-IAP#

Caution

Parallel execution of LAMMPS/ML-IAP using CUDA-aware MPI is supported, but not yet sufficiently tested. For multi-rank LAMMPS runs, use the TorchScript e3gnn/parallel workflow described in LAMMPS: PyTorch.

Requirements#

cython == 3.0.11
cupy-cuda12x or cupy-cuda13x
flashTP (optional, follow here)
cuEquivariance (optional, follow here)
OpenEquivariance (optional, follow here)

Install via:

pip install sevenn[mliap12]  # For CUDA 12.x
pip install sevenn[mliap13]  # For CUDA 13.x

Build#

Get LAMMPS source code:

git clone https://github.com/lammps/lammps lammps-mliap
cd lammps-mliap
git checkout stable_22Jul2025_update4

Note

SevenNet’s ML-IAP tests are validated against the LAMMPS stable_22Jul2025_update4 release. Other LAMMPS versions may work, but should be validated with the ML-IAP test suite before production use.

Configure the LAMMPS build with the necessary options:

cd ./lammps-mliap
cmake \
    -B build \
    -D CMAKE_BUILD_TYPE=Release \
    -D CMAKE_CXX_COMPILER=$(pwd)/lib/kokkos/bin/nvcc_wrapper \
    -D PKG_KOKKOS=ON \
    -D Kokkos_ENABLE_CUDA=ON \
    -D BUILD_MPI=ON \
    -D PKG_ML-IAP=ON \
    -D PKG_ML-SNAP=ON \
    -D MLIAP_ENABLE_PYTHON=ON \
    -D PKG_PYTHON=ON \
    -D BUILD_SHARED_LIBS=ON \
    cmake

Build LAMMPS and install the Python bindings:

cmake --build build -j 8
cd build
make install-python

Note

If the compilation fails, consider specifying the GPU architecture of the node you are building on in KOKKOS. For example, add the following flag to your cmake command: -D Kokkos_ARCH_AMPERE86=ON when you are using GPU with Ampere 86 architecture like RTX A5000.

(Optional) Build with GPU-D3 pair style#

To build LAMMPS with the GPU-accelerated Grimme’s D3 method (see CUDA-accelerated Grimme’s D3), follow the steps below:

Clone the SevenNet source repository.

git clone https://github.com/MDIL-SNU/SevenNet.git

Copy the pair_d3-related scripts into lammps/src.

cp {sevenn_src_path}/sevenn/pair_e3gnn/{pair_d3.cu,pair_d3.h,pair_d3_pars.h} {lammps_src_path}/src/

Modify {lammps_src_path}/cmake/CMakeLists.txt. Lines starting with - indicate the original content, and lines starting with + indicate the modified or added content.

# Declare CUDA at the header. This part depends to LAMMPS version.
 project(lammps
         DESCRIPTION "The LAMMPS Molecular Dynamics Simulator"
         HOMEPAGE_URL "https://www.lammps.org"
-        LANGUAGES CXX C)
+        LANGUAGES CXX C CUDA)

...

# Compile pair_d3.cu
-file(GLOB ALL_SOURCES CONFIGURE_DEPENDS ${LAMMPS_SOURCE_DIR}/[^.]*.cpp)
+file(GLOB ALL_SOURCES CONFIGURE_DEPENDS ${LAMMPS_SOURCE_DIR}/[^.]*.cpp  ${LAMMPS_SOURCE_DIR}/[^.]*.cu)

...

# At the end of the CMakeLists.txt, add cuda-related flags
+find_package(CUDA)
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -fmad=false -O3 --expt-relaxed-constexpr")
+target_link_libraries(lammps PUBLIC ${CUDA_LIBRARIES} cuda)

Add the target CUDA architecture in the cmake configuration. An example configuration is shown below.

cmake \
    -B build \
    -D CMAKE_BUILD_TYPE=Release \
    -D CMAKE_CXX_COMPILER=$(pwd)/lib/kokkos/bin/nvcc_wrapper \
    -D PKG_KOKKOS=ON \
    -D Kokkos_ENABLE_CUDA=ON \
    -D Kokkos_ARCH_AMPERE89=ON \  # for ML-IAP Kokkos implementation
    -D BUILD_MPI=ON \
    -D PKG_ML-IAP=ON \
    -D PKG_ML-SNAP=ON \
    -D MLIAP_ENABLE_PYTHON=ON \
    -D PKG_PYTHON=ON \
    -D BUILD_SHARED_LIBS=ON \
    -D CMAKE_CUDA_ARCHITECTURES="89" \  # for CUDA D3 pair style
    cmake

Build LAMMPS and install the Python bindings as above.

Usage#

Potential deployment#

Please check sevenn graph_build for detail.

An ML-IAP potential checkpoint can be deployed using sevenn get_model command with --use_mliap flag.

By default, output file name will be deployed_serial_mliap.pt. (You can customize the output file name using --output_prefix flag.)
You can accelerate the inference with --enable_cueq, --enable_flash, or --enable_oeq flag:

sevenn get_model \
    {pretrained_name or checkpoint_path} \
    --use_mliap \
    --modal {task_name}  # Required when using multi-fidelity model

Single-GPU MD#

Below is an example snippet of a LAMMPS script for using SevenNet models with ML-IAP:

    # ---------- Initialization ----------
    units           metal
    boundary        p p p
    atom_style      atomic
    atom_modify     map yes
    newton          on

    read_data       {your_system_file_path}

    # ----- ML-IAP potential settings ----
    pair_style      mliap unified {your_potential_file_path.pt} 0
    pair_coeff      * * {space separated chemical species}

Note

If the LAMMPS is built with GPU-D3 pair style, you can combine SevenNet with D3 through the pair/hybrid command as the example below. For detailed instruction about parameters, supporting functionals and damping types, refer to CUDA-accelerated Grimme’s D3.

pair_style hybrid/overlay mliap unified {your_potential_file_path.pt} 0 d3 9000 1600 damp_bj pbe
pair_coeff      * * mliap {space separated chemical species}
pair_coeff      * * d3    {space separated chemical species}

Example command to run a LAMMPS simulation with ML-IAP:

/path/to/lammps-mliap/build/lmp -in lammps.in -k on g 1 -sf kk -pk kokkos newton on neigh half