MODNet: Material Optimal Descriptor Network

Introduction

This repository contains the Python (3.8+) package implementing the Material Optimal Descriptor Network (MODNet). It is a supervised machine learning framework for learning material properties from either the composition or crystal structure. The framework is well suited for limited datasets and can be used for learning multiple properties together by using joint learning.

MODNet appears on the MatBench leaderboard. As of 11/11/2021, MODNet provides the best performance of all submitted models on 7 out of 13 tasks.

This repository also contains two pretrained models that can be used for predicting the refractive index and vibrational thermodynamics from any crystal structure.

See the MODNet papers and repositories below for more details:

• De Breuck et al., "Materials property prediction for limited datasets enabled by feature selection and joint learning with MODNet." npj Comput Mater 7, 83 (2021). 10.1038/s41524-021-00552-2 (preprint: arXiv:2004.14766).
• De Breuck et al., "Robust model benchmarking and bias-imbalance in data-driven materials science: a case study on MODNet." J. Phys.: Condens. Matter 33 404002, (2021), 10.1088/1361-648X/ac1280 (preprint: arXiv:2102.02263).
• MatBench benchmarking data repository: modl-uclouvain/modnet-matbench.

Figure 1. Schematic representation of the MODNet.

How to install

First, create a virtual environment (e.g., named modnet) with Python (3.8+) using your favourite environment manager (the following instructions use conda):

conda create -n modnet python=3.9

Activate the environment:

conda activate modnet

Finally, install MODNet from PyPI with pip:

pip install modnet

Warning We strongly recommend pinning your Python environment when using MODNet across multiple machines, or multiple MODNet versions, as changes to the dependencies and sub-dependencies can lead to different values for particular features.

This can be achieved with conda export or pip freeze.

For development (or if you wish to use pinned versions of direct dependencies that MODNet has been tested with), you can clone this git repository and make an editable install inside your chosen environment with pip:

git clone git@github.com:ppdebreuck/modnet
cd modnet
conda create -n modnet python=3.9
conda activate modnet
pip install -r requirements.txt  # optionally use pinned requirements
pip install -e .

Documentation

The documentation is available at ReadTheDocs.

Changelog

A brief changelog can be found in the release summaries on GitHub.

Author

This software was written by Pierre-Paul De Breuck and Matthew Evans with contributions from David Waroquiers and Gregoire Heymans. For an up-to-date list, see the Contributors on GitHub.

License

MODNet is released under the MIT License.