Predict crystal properties using MatErials Graph Network (MEGNet) models.
Background:MEGNet, or MatErials Graph Network is an implementation of graph networks for materials science by the Materials Virtual Lab. These models are trained on data from the Materials Project. Please refer to the open-source repo and the following work for implementation details and benchmarks: Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; Ong, S. P. Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals. Chemistry of Materials 2019, acs.chemmater.9b01294. doi:10.1021/acs.chemmater.9b01294.
Because the MEGNet models are trained on Materials Project (MP)-relaxed structures, predicting using MP ids will likely be more accurate. If using uploaded crystals, you should ideally use the relaxed structures from DFT-PBE calculations. Nevertheless, our preliminary tests find that the MEGNet prediction error using experimental structures, while substantially larger, are still reasonably well-controlled.