Filip Ekström Kelvinius, Rickard Armiento and I got a paper accepted for Physical Review Materials.
The paper is called Graph-based machine learning beyond stable materials and relaxed crystal structures. We investigate the usefulness of graph neural networks, specifically the Crystal Graph Convolutional Neural Network (CGCNN), for predicting formation energies for hypotetical materials that are: (i) unrelaxed or partially relaxed, i.e. their structures are not completely known, and (ii) unstable, i.e. far from the convex hull of the corresponding phase diagram. We also explore options for improving the performance in these scenarios by transfer learning, either from models trained on a large database of mostly stable systems, or a different but related phase diagram. Models pre-trained on stable materials do not significantly improve performance, but models trained on similar data transfer very well. We demonstrate how our findings can be utilized to generate phase diagrams with a major reduction in computational effort.