Crystal Graph Convolutional Neural Network

Crystal Graph Convolutional Neural Networks (CGCNNs) are machine learning models designed to predict material properties from their atomic structures, represented as graphs. Current research focuses on improving the accuracy and robustness of CGCNNs, particularly their ability to generalize to unseen materials (out-of-distribution prediction), often employing ensemble methods and adversarial training techniques to enhance performance. These models are significantly impacting materials science by accelerating the discovery and design of new materials with desired properties, ranging from improved energy storage to novel alloys, and are even finding applications beyond materials science, such as in vehicle trajectory prediction.