Installation

Requirements

  • Python >= 3.9

  • numpy >= 1.22

  • ODAT-SE >= 3.0

  • AENET (predict.x, generate.x, train.x)

The following is also required for the training tutorial:

Installing ODAT-SE

To install from PyPI:

pip3 install odat-se[all]

To install from source:

git clone https://github.com/issp-center-dev/ODAT-SE.git
cd ODAT-SE
pip3 install .[all]

Installing AENET

AENET (Atomic Energy Network) is used to build machine learning potentials. Building requires a Fortran compiler (gfortran) and LAPACK/BLAS.

Prerequisites for macOS

If gfortran is not installed, install it via Homebrew:

brew install gcc

This makes the gfortran command available. On macOS, the Accelerate framework is available by default as a LAPACK/BLAS implementation.

Obtaining the source code

git clone https://github.com/atomisticnet/aenet.git
cd aenet

Building the L-BFGS-B library

Before building AENET itself, you need to build the bundled L-BFGS-B library first:

cd lib
make static
cd ..

Verify that lib/liblbfgsb.a has been generated.

Building AENET

In the src/ directory, build by specifying the Makefile appropriate for your environment:

cd src
make -f makefiles/Makefile.gfortran_serial_MacOS   # macOS (Apple Silicon / Intel)
cd ..

Note

For Linux environments, choose one of the following depending on whether MPI is available:

  • Makefile.gfortran_serial — Serial version

  • Makefile.gfortran_mpi — MPI parallel version

  • Makefile.gfortran_openblas_serial — OpenBLAS version

See src/makefiles/ for a full list of available Makefiles.

Once the build is complete, the following executables are generated under bin/:

  • generate.x — Generation of training data (structural descriptors)

  • train.x — Training of neural network potentials

  • predict.x — Energy prediction using trained potentials

Note

The executable filenames include version and compiler suffixes (e.g., predict.x-2.0.4-gfortran_serial). Create symbolic links as needed:

cd bin
ln -s predict.x-2.0.4-gfortran_serial predict.x
ln -s generate.x-2.0.4-gfortran_serial generate.x
ln -s train.x-2.0.4-gfortran_serial train.x

Setting up the PATH:

export PATH=/path/to/aenet/bin:$PATH

Verification:

generate.x
# If "generate.x - training set generation" is displayed, the installation was successful

Note

For more details, see the AENET official website.

Installing Quantum ESPRESSO

Quantum ESPRESSO is a first-principles calculation engine used for structure relaxation in the training tutorial.

Prerequisites (macOS)

cmake is required:

brew install cmake

Building from source

git clone --depth 1 --branch qe-7.4 https://github.com/QEF/q-e.git qe-7.4
cd qe-7.4

Warning

Versions v7.3.1 and earlier fail to build the mbd library with GCC 15 (gfortran 15). Using v7.4 or later is recommended.

Configure the build with cmake. For the case without MPI and using the internal FFTW:

cmake -B build \
  -DCMAKE_Fortran_COMPILER=gfortran \
  -DCMAKE_C_COMPILER=gcc-15 \
  -DQE_ENABLE_MPI=OFF \
  -DQE_FFTW_VENDOR=Internal

Note

  • Adjust gcc-15 according to your environment (e.g., on Linux where gcc is GNU, simply use gcc).

  • If you need the MPI parallel version, remove -DQE_ENABLE_MPI=OFF and specify -DCMAKE_Fortran_COMPILER=mpif90.

  • If an external FFTW3 is already installed, -DQE_FFTW_VENDOR=Internal is not needed.

To build only pw.x:

cmake --build build --target pw -j4

To build everything:

cmake --build build -j4

After building, add pw.x to your PATH:

export PATH=/path/to/qe-7.4/build/bin:$PATH

Verification:

pw.x --version
# If "Program PWSCF v.7.4" is displayed, the installation was successful

Note

Installing odatse-aenet

git clone https://github.com/k-yoshimi/odatse-aenet.git
cd odatse-aenet
pip3 install .

After installation, the odatse-aenet command becomes available.

Verification

odatse-aenet --version

If the version number is displayed, the installation was successful.