A repository to demonstrate fitting a minimal message-passing network. Intended as a teaching example more than a high-quality implemetnation.
The code in these libraries is close to what is used in early versions of nfp. If you are looking to do more state-of-the art deep learning on molecules, I would recommend learning a package like nfp, megnet, schnetpack, or deepchem.
The module requirements are described in environment.yml and can be installed with:
conda env create --file environment.yml --forceThe first notebook must be run to save the data in a protobuf format accessible by later notebooks.
Then you can either run the "small neural network" to get a minimal (117-parameter) MPNN for comparison, or the "large model" to compare models on different material properties.