unlockNN

A Python package for interpreting and extracting uncertainties in neural network models of chemical systems based upon Gaussian processes.

Statement of need

Neural networks (NNs) are powerful tools for materials property prediciton (MPP) based on structural information. After training, they offer a cheaper alternative to density function theory (DFT) and are therefore promising for high throughput screening of materials. However, most current implementations of NNs for MPP lack uncertainty quantifiers. Knowledge of the certainty in an estimate is particularly important for machine learning models, as the reliability of a prediction depends on the existence of functionally similar structures in the training dataset, which cannot be readily determined.

UnlockNN contains utilities for adding uncertainty quantification to Keras-based models. This is achieved by replacing the last layer of the model with a variational Gaussian process (VGP), a modification of a Gaussian process that improves scalability to larger data sets. The caveat is that the modified model must undergo further training in order to calibrate the uncertainty quantifier; however, this typically only requires a small number of training iterations.

UnlockNN also contains a specific configuration for adding uncertainty quantification to MEGNet: a powerful graph NN model for predicting properties of molecules and crystals.

Installation

The package can be installed by cloning this repository and building it using either anaconda or pip, or it can be downloaded directly from PyPi.

To install from PyPi, run pip install unlockNN. To install from source:

git clone https://github.com/a-ws-m/unlockNN.git
cd unlockNN
conda env create -f environment.yml  # Optional: create a virtual environment with conda
pip install .

The dev_environment.yml contains additional dependencies for development, testing and building documentation. It can be installed using conda env create -f dev_environment.yml.

Documentation

Full documentation is available for the project here.

Benchmarks

Benchmark details and results can be found in the benchmarking subdirectory.

License and attribution

Code licensed under the MIT License.

Development notes

Reporting issues

Please use the Issue tracker to report bugs in the software, suggest feature improvements, or seek support.

Contributing to unlockNN

Contributions are very welcome as we look to make unlockNN more flexible and efficient. Please use the Fork and Pull workflow to make contributions and follow the contribution guidelines:

• Use the environment defined in dev_environment.yml. This installs black, the formatter used for this project, as well as utilities for building documentation (sphinx and the insegel theme), enabling the testing suite (pytest and pytest-cov) and publishing to PyPi (build, but this will be handled by the package maintainer).
• Use black to format all Python files that you edit: black {edited_file.py} or python -m black {edited_file.py}.
• Write tests for new features in the appropriate directory. Run tests using pytest tests/, or optionally with pytest --cov=unlocknn tests/ to generate coverage on the fly.
• After testing that pytest works for your current environment, run tox in the root directory of the project to check that all versions of Python are compatible.
• Use Google-style Docstrings. Check docstrings with pydocstyle: pydocstyle {edited_file.py}.
• Check docstrings are valid Sphinx RST and that the documentation compiles without errors: in the docs directory, run make html.
• Feel free to clean up others' code as you go along.

List of developers

Contributors to unlockNN:

• Alexander Moriarty

Huge thanks to Keith Butler, Aron Walsh and Kazuki Morita for supervising the project at its inception and for their immense support.