A generator of polynomial machine learning potentials

Polynomial machine learning potentials

• Polynomial machine learning potential repository
• A. Seko, J. Appl. Phys. 133, 011101 (2023)

Required libraries and python modules

• numpy
• scipy
• Eigen3 (if building pypolymlp)
• pybind11 (if building pypolymlp)
• phonopy (if using phonon datasets and/or computing force constants)
• phono3py (if using phonon datasets and/or computing force constants)
• symfc (if computing force constants)
• spglib (optional)
• joblib (optional)

Installing pypolymlp using pip package

• Intel Linux
• Compatible with python 3.9, 3.10, and 3.11

> pip install $(pypolymlp)/dist/pypolymlp-0.1-py3-none-any.whl

Manually installing pypolymlp

1. Conda package management system

> conda create -n pypolymlp
> conda activate pypolymlp

> conda install numpy scipy pybind11 eigen cmake
(optional)
> conda install spglib
> conda install phono3py
> conda install joblib

2. Building a shared library (libmlpcpp)

The process of building libmlpcpp may take approximately five minutes to one hour.

> cd $(pypolymlp)/src/pypolymlp/cxx
> cmake -S . -B build
> cmake --build build
> cmake --install build

or

> cd $(pypolymlp)/src/pypolymlp/cxx
> make

If necessary, the stack size may need to be set to unlimited.

ulimit -s unlimited

3. Install pypolymlp using pip

> cd $(pypolymlp)
> pip install .

How to use pypolymlp

• Polynomial MLP development
• Property calculators
  • Energy, forces on atoms, and stress tensor
  • Force constants
  • Elastic constants
  • Equation of states
  • Structural features (Polynomial invariants)
  • Local geometry optimization
  • Phonon properties, Quasi-harmonic approximation
  • Self-consistent phonon calculations
• Utilities
  • Random structure generation
  • Estimation of computational costs
  • Enumeration of optimal MLPs
  • Compression of vasprun.xml files
  • Automatic division of DFT dataset
  • Atomic energies
• Python API (MLP development)
• Python API (Property calculations)