LAMMPS Molecular Dynamics with MLIPs

Goal

Run GPU-accelerated LAMMPS molecular dynamics with MLIP backends using three isolated binaries (MACE, MatGL/CHGNet, FairChem) so Python embedding through ML-IAP/mliappy remains stable and reproducible.

Instructions

1. Select the MLIP backend and model family first using the foundation-potential guide:

   • ml-foundation-potentials
   • This determines which conda env and which LAMMPS binary you must use.

2. Check system prerequisites.

# Env: base-agent
nvidia-smi
nvcc --version
g++ --version
cmake --version
mpicxx --version

3. Identify GPU compute capability and set Kokkos arch flag.

# Env: base-agent
nvidia-smi --query-gpu=name,compute_cap --format=csv,noheader

• Example mapping:
  • 8.0 -> Kokkos_ARCH_AMPERE80
  • 8.6 -> Kokkos_ARCH_AMPERE86
  • 8.9 -> Kokkos_ARCH_ADA89
  • 9.0 -> Kokkos_ARCH_HOPPER90

4. Build the environment-matched LAMMPS binary (choose one of the three paths below).

   Path A: MACE

# Env: base-agent
bash conda-envs/mace-agent/install.sh
KOKKOS_ARCH_FLAG=Kokkos_ARCH_AMPERE86 \
LAMMPS_REF="stable_2Aug2023_update2" \
bash conda-envs/mace-agent/install_lammps.sh

• Binary: ./lammps/mace-agent/lmp
• Runtime env: mace-agent

Path B: MatGL/CHGNet

# Env: base-agent
bash conda-envs/matgl-agent/install.sh
KOKKOS_ARCH_FLAG=Kokkos_ARCH_AMPERE86 \
LAMMPS_REF="stable_2Aug2023_update2" \
bash conda-envs/matgl-agent/install_lammps.sh

• Binary: ./lammps/matgl-agent/lmp
• Runtime env: matgl-agent

Path C: FairChem

# Env: base-agent
bash conda-envs/fairchem-agent/install.sh
KOKKOS_ARCH_FLAG=Kokkos_ARCH_AMPERE86 \
LAMMPS_REF="stable_2Aug2023_update2" \
bash conda-envs/fairchem-agent/install_lammps.sh

• Binary: ./lammps/fairchem-agent/lmp
• Runtime env: fairchem-agent

5. Run the selected binary with its matching conda environment.

# Env: mace-agent (example; switch env/binary pair as needed)
conda activate mace-agent
./lammps/mace-agent/lmp -h

6. Launch MD with the same binary-env pair used during build; do not cross-run binaries between MLIP stacks.

Examples

See scripts/three-backends-build-check/README.md for a minimal build/verification matrix across MACE, MatGL, and FairChem. See the respective README.md files under examples/mace/, examples/matgl/, and examples/fairchem/ for model-specific run scripts.

Constraints

• Strict binary-env pairing: each LAMMPS binary must run only with its own conda env.
• No stack mixing: never run MACE binary in matgl-agent/fairchem-agent, etc.
• GPU arch alignment: choose KOKKOS_ARCH_* from actual compute_cap output.
• Python-coupled mode: this workflow targets ML-IAP/mliappy usage.

References

• Thompson et al., "LAMMPS - A flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales", Computer Physics Communications, 2022. DOI
• LAMMPS Manual, ML-IAP package documentation. Link
• Batatia et al., "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields". arXiv
• Deng et al., "CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling". arXiv
• FairChem documentation and model zoo. Link

---

Author: Jurģis Ruža Contact: GitHub @JurgisR