Our paper on MatErials Graph Networks (MEGNet) for machine learning in crystals and molecules have been published in Chemistry of Materials. The article is available here. Key advances include the incorporation of state variables such as temperature, pressure and entropy, transfer learning from models with large data (e.g., formation energies) to models with smaller data (e.g., elastic constants) and extraction of chemical trends from learned elemental embeddings. These advances address key limitations in ML in materials science, such as data size limitations and physical interpretability.
We have also released all our codes and data in our open Github repo at https://github.com/materialsvirtuallab/megnet to enable others to reproduce and improve on our models.