peslearn 1.0.0

Automated Construction of Machine Learning Models of Molecular Potential Energy Surfaces.

PES-Learn

PES-Learn is a Python library designed to fit system-specific Born-Oppenheimer potential energy surfaces using modern machine learning models. PES-Learn assists in generating datasets, and features Gaussian process and neural network model optimization routines. The goal is to provide high-performance models for a given dataset without requiring user expertise in machine learning.

This project is young and under active development. It is recommended to take a look at the Tutorials, the FAQ page, and the list of keyword options before using PES-Learn for research purposes. More documentation will be added periodically. Questions and comments are encouraged; please consider submitting an issue.

Features

• Ease of Use

  • PES-Learn can be run by writing an input file and running the code (much like most electronic structure theory packages)
  • PES-Learn also features a Python API for more advanced workflows
  • Once ML models are finished training, PES-Learn automatically writes a Python file containing a function for evaluating the energies at new geometries.

• Data Generation

  • PES-Learn supports input file generation and output file parsing for arbitrary electronic structure theory packages such as Psi4, Molpro, Gaussian, NWChem, etc.
  • Data is generated with user-defined internal coordinate displacements with support for:
    • Redundant geometry removal
    • Configuration space filtering

• Automated Data Transformation

  • Rotation, translation, and permutation invariant molecular geometry representations

• Automated Machine Learning Model Generation

  • Neural network models are built using PyTorch
  • Gaussian process models are built using GPy

• Hyperparameter Optimization

Installation Instructions

PES-Learn has been tested and developed on Linux, Mac, and Windows 10 through the Windows Subsystem for Linux. To install using pip:
Clone the repository:
git clone https://github.com/adabbott/PES-Learn.git
Change into top-level directory:
cd PES-Learn
Install PES-Learn and all dependencies:
python setup.py install
To avoid having to re-install the package whenever a change is made to the code, run pip install -e .

Install using an Anaconda environment

The above procedure works just fine, however for performance and stability, we recommend installing all PES-Learn dependencies in a clean Anaconda environment. After installing Anaconda for Python3, create and activate an environment:
conda create -n peslearn python=3.6
conda activate peslearn
The required dependencies can be installed in one line:
conda install -c conda-forge -c pytorch -c omnia gpy pytorch scikit-learn pandas hyperopt cclib
Then install the PES-Learn package:
git clone https://github.com/adabbott/PES-Learn.git
python setup.py install
pip install -e .

To update PES-Learn in the future, run git pull while in the top-level directory PES-Learn.

To run the test suite, you need pytest: pip install pytest-cov To run tests, in the top-level directory called PES-Learn, run: py.test -v tests/

Citing PES-Learn

PES-Learn: An Open-Source Software Package for the Automated Generation of Machine Learning Models of Molecular Potential Energy Surfaces

Bibtex:

@article{Abbott2019,
  title={PES-Learn: An open-source software package for the automated generation of machine learning models of molecular potential energy surfaces},
  author={Abbott, Adam S and Turney, Justin M and Zhang, Boyi and Smith, Daniel GA and Altarawy, Doaa and Schaefer III, Henry F},
  journal={Journal of chemical theory and computation},
  volume={15},
  number={8},
  pages={4386--4398},
  year={2019},
  publisher={ACS Publications}
}

Funding

This project is a collaboration with the Molecular Sciences Software Institute. The author gratefully acknowledges MolSSI for funding the development of this software.