This project implements a Crystal Graph Convolutional Neural Network (CGCNN,(https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.120.145301) to predict properties of crystalline materials sourced from the Materials Project. The main difference from the published paper is that this model uses pre-trained mat2vec atomic embeddings in the graph architecture. I also use a logarithmic transformation for the band gap values before training to help the model learn more effectively from the skewed data distribution. After training, the final model achieved a Test MAE of 0.36 eV for band gap prediction and a Test MAE of 0.04 eV/atom for formation energy prediction. The best models for band gap prediction and formation energy prediction are also included in the repository.
To reproduce these results, use the environment from environment.yml. The pre-trained mat2vec model files are not included in this repository due to their size; please download the pretrained_embeddings files from the official mat2vec GitHub repository and place them in a local data/ directory. You will also need to add your personal Materials Project API key to the data fetching script.