Graph neural network analysis of layered material phases

We apply graph neural network (GNN)-based analysis to automatically classify different crystalline phases inside computationally-synthesized molybdenum disulfide monolayer by reactive molecular dynamics (RMD) simulations on parallel computers. We have found that addition of edge-based features like distance increases the model accuracy up to 0.9391. Network analysis by visualizing the feature space of our GNN model clearly separates 2H and 1T crystalline phases inside the network. This work demonstrates the power of the GNN model to identify structures inside multimillion-atom RMD simulation data of complex materials.