Running the model in LAMMPS or ASE

Dependencies

Physics based KIM models can run from LAMMPS or ASE out of the box. You just need to install them as

kim-api-collections-management install system <model name>

However, for running the ML models you need install few dependencies first, namely - libtorch (C++ API for Pytorch) - libtorchscatter and libtorchsparse

Attention

Installing libtorch in same LD_LIBRARY_PATH as Pytorch can crash your pytorch installation (segfaults), which will mean that you will need to remove the libtorch from the environment path everytime you neet to train. Therefore it is recommended that you create a new conda env, specifically for running the models.

Let us create a fresh environment and install required dependencies

conda create -n kliff_run
conda activate kliff_run
conda install lammps kimpy ase=3.22 python=3.9 unzip -c conda-forge

If you want to use the descriptor based models as well, then install libdescriptor, and set its root path

conda install libdescriptor -c ipcamit -c conda-forge
export LIBDESCRIPTOR_ROOT=$CONDA_PREFIX

Tip

If you are using GPU based systems, then it is advisable to use CMake version 3.18, with recommended CUDA version 11.7. Newer versions of CMake fail to detect CUDA environments correctly. You can install CMake as

conda install cmake=3.18 -c conda-forge

Now, installing libtorch and other dependencies is simple, for most systems you can just download the binaries and use them directly. If you are working in a Linux environment, OpenKIM provides an easy-to-use installation script that you can use as:

wget https://openkim.org/files/MD_173118614730_001/install_dependencies.sh
bash install_dependencies.sh

wget https://gist.githubusercontent.com/ipcamit/25b9de5ab823fe00f8eaae89ef8c60d4/raw/1c2cf62ef3e9f2108e5227feb9cb50d203422a35/install_macos_dependencies.sh
bash install_macos_dependencies.sh

Following the installation, just source the produces env.sh file:

source env.sh

source env.sh

# For runtime
cp $TorchSparse_ROOT/lib/libtorchsparse.dylib  $CONDA_PREFIX/lib/
cp $TorchScatter_ROOT/lib/libtorchscatter.dylib  $CONDA_PREFIX/lib/
cp $TORCH_ROOT/lib/*  $CONDA_PREFIX/lib/

and now you are ready to install the required model driver

Note

KIM-API system install location on MacOS is a hit-or-miss, so it is recommended to use user install location for MacOS. Also due to older libtorch versions not being available for MacOS, you need to install the TorchML driver manually. This is mainly because newer versions of libtorch specifically need C++17 support.

kim-api-collections-management install user TorchML__MD_173118614730_000

wget https://openkim.org/download/TorchML__MD_173118614730_000.txz
tar -xvf TorchML__MD_173118614730_000.txz
sed -i '' 's/libdescriptor.so/libdescriptor.dylib/' TorchML__MD_173118614730_000/CMakeLists.txt
sed -i '' 's/PROPERTY CXX_STANDARD 14/PROPERTY CXX_STANDARD 17/' TorchML__MD_173118614730_000/MLModel/CMakeLists.txt
kim-api-collections-management install user TorchML__MD_173118614730_000

Testing the installation

Let us install a simple Stillinger-Weber potential for Silicon and test it

kim-api-collections-management install user  SW_StillingerWeber_1985_Si__MO_405512056662_006

from ase.calculators.kim.kim import KIM
from ase.build import bulk

si = bulk("Si")
model = KIM("SW_StillingerWeber_1985_Si__MO_405512056662_006")
si.calc = model
print(si.get_potential_energy())
print(si.get_forces())

Using your models

You can install you models now as any other KIM model,

kim-api-collections-management install user SchNet1__MO_111111111111_000

Now that your model is installed, you can use it in LAMMPS and ASE,

ASE

RUN with SW first.

from ase.calculators.kim.kim import KIM
from ase.build import bulk

si = bulk("Si")
model = KIM("SchNet1__MO_111111111111_000")
si.calc = model
print(si.get_potential_energy())
print(si.get_forces())

LAMMPS

Save the following script as in.lammps,

# Define KIM model and get Si diamond lattice parameter for this potential
kim init         SchNet1__MO_111111111111_000 metal
# Setup diamond crystal
boundary         p p p
lattice          diamond 5.44
region           simbox block 0 1 0 1 0 1 units lattice
create_box       1 simbox
create_atoms     1 box
mass             1 28.0855
# Define atom type to species mapping
kim interactions Si
# Compute energy
run 0

You can now run the LAMMPS script as

lmp_serial -i in.lammps

Common Errors

  1. std::optional error

During the installation of dependencies (from install script) you might get an error looking like

... ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h:201:37: note: 'std::integral_constant<bool, false>::value' evaluates to false
    did you forgot #include <optional>
    Requires CXX17 support..

This indicates that for some reason you are installing a mismatched version of torch scatter or torchsparse libraries. Either try to download the install script again, or try the copy of latest install script as

wget https://gist.githubusercontent.com/ipcamit/646573856b7f5735edd7048687d9655a/raw/d9e2cd51436b3f0a14ec5366f2335575801831fe/install_dependencies.sh

  1. Python not found

CMake Error at CMakeLists.txt:28 (add_library):
Target "torchscatter" links to target "Python3::Python" but the target was
not found.  Perhaps a find_package() call is missing for an IMPORTED
target, or an ALIAS target is missing?

This error usually means that you are using python > 3.9 with CMake == 3.18. CMake 3.18 has hardcoded python version <=3.9 string, making it yield an error. You have two options,

  1. use latest CMAKE and no CUDA/GPU

  2. use CMAKE 3.18 python <= 3.9

Another option for using python >3.9 with CMAKE 3.18 is to patch the the CMAKE python module as

sed -i "25s/set(_\${_PYTHON_PREFIX}_VERSIONS /set(_\${_PYTHON_PREFIX}_VERSIONS $(python -c 'import sys; print(f"{sys.version_info[0]}.{sys.version_info[1]}")') /" $CONDA_PREFIX/share/cmake-3.18/Modules/FindPython/Support.cmake

Tip

This will only work on CMake 3.18, and only do it if you absolutely need the GPU support. It prepends your current python version to the Line 25 of CMake module.

  1. dyld[19429]: symbol not found in flat namespace '_error_top' in lammps

LAMMPS is not properly installed. You can try to reinstall LAMMPS,

conda remove lammps
conda install lammps -c conda-forge

  1. GLIBCXX_3.X.X not found

version `GLIBCXX_3.4.29' not found (required by /opt/mambaforge/mambaforge/envs/kliff_run/lib/libkim-api.so.2)

This error usually means that you are using a different version of GCC than the one used to compile the KIM-API. Just set the following environment variable before running the code,

export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

  1. lmp_serial command not found

Sometimes in some conda environments, the LAMMPS binary is called lmp, so try

lmp -in in.lammps

  1. Missing KIM model during ASE/LAMMPS runs

ase.calculators.kim.exceptions.KIMModelNotFound: Could not find model <> installed
in any of the KIM API model collections on this system.
See https://openkim.org/doc/usage/obtaining-models/ for instructions on installing models.

It usually happens when you have multiple kim-api installations, and the ASE/LAMMPS is not able to find the correct one. Easiest solution can be to restart the terminal and load the correct conda environment. If that does not help, try the following (solution i. is recommended),

  1. using kim-api-activate

source kim-api-activate

  1. set LD_LIBRARY_PATH manually

export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

Making changed permanent

Attention

This can potentially mess up your conda install. So only do this if you understand the risks. Also for MacOS you do not need the instructions below, as anyway you are copying the libraries to the conda environment.

If you want to ensure that everytime you activate your conda environment it loads all the dependencies on itself, without source env.sh, you can create activation hooks in conda environment. Please make sure that you are in same folder as your env.sh file and run,

conda activate kliff_run

[[ -f env.sh ]] || { echo "Error: env.sh not found"; return; }

# Create hooks directories
mkdir -p "$CONDA_PREFIX/etc/conda/activate.d" "$CONDA_PREFIX/etc/conda/deactivate.d"

env_prefix=$CONDA_PREFIX

# Generate activation script
echo "#\!/bin/bash" > "$CONDA_PREFIX/etc/conda/activate.d/env_activate.sh"
echo "export _OLD_LD_LIBRARY_PATH=\$LD_LIBRARY_PATH" >> "$CONDA_PREFIX/etc/conda/activate.d/env_activate.sh"
echo "export _OLD_PATH=\$PATH" >> "$CONDA_PREFIX/etc/conda/activate.d/env_activate.sh"
cat env.sh >> "$CONDA_PREFIX/etc/conda/activate.d/env_activate.sh"

conda deactivate
# Generate deactivation script
echo "#\!/bin/bash" > "$env_prefix/etc/conda/deactivate.d/env_deactivate.sh"
echo "export LD_LIBRARY_PATH=\$_OLD_LD_LIBRARY_PATH" >> "$env_prefix/etc/conda/deactivate.d/env_deactivate.sh"
echo "export PATH=\$_OLD_PATH" >> "$env_prefix/etc/conda/deactivate.d/env_deactivate.sh"
echo "unset _OLD_LD_LIBRARY_PATH" >> "$env_prefix/etc/conda/deactivate.d/env_deactivate.sh"
echo "unset _OLD_PATH" >> "$env_prefix/etc/conda/deactivate.d/env_deactivate.sh"

Now everytime you activate the environment, it will automatically load the dependencies. You can test this by opening a new terminal and running

conda activate kliff_run
lmp_serial -in in.lammps