Lightweight Graph Neural Network for Crystal Structure Prediction

In this work, we propose a framework for reliable crystal structure prediction using simpler GNN models leading to good accuracy and faster speed as compared to computationally expensive GNN models. This has the potential to accelerate crystal structure prediction which is often a primary task in search for novel functional materials. We have made a comparison between CGCNN, MEGNet and ALIGNN model in this work and showed that CGCNN model, though simpler and faster, can attain similar accuracy as more expensive GNN models like ALIGNN by following the strategies outlined in our manuscript. The documentation to perform proposed model selection, data augmentation and geometry optimization with each of these models is provided in README.md file within respective directories.