A Python package for interpreting and extracting uncertainties in neural network models of chemical systems based upon Gaussian processes.
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.
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.
Full documentation is available for the project here.
Code licensed under the MIT License.
Please use the Issue tracker to report bugs in the software, suggest feature improvements, or seek support.
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 installsblack, the formatter used for this project, as well as utilities for building documentation, enabling the testing suite and publishing to PyPi. - Write tests for new features in the appropriate directory.
- Use Google-style Docstrings. Check docstrings with
pydocstyle. - Feel free to clean up others' code as you go along.
Contributors to unlockNN:
Huge thanks to Keith Butler, Aron Walsh and Kazuki Morita for supervising the project at its inception and for their immense support.