Prediction of Material Properties using Crystal Graph Convolutional Neural Networks

Machine Learning principles have found application to scientific domains, such as materials science, owing to their significant efficiency compared with performing ab initio calculations. In recent years, Graph Convolutional Networks (GCNs) have gained popularity in the same domain, for the task of prediction and classification of crystal properties, as they have been proven to be effective for node-representation learning while working with graph-like structures. Fundamentally, similar to conventional Convolution Neural Networks (CNNs), GCNs are also composed of convolution and pooling layers. While there is ample literature focusing on the convolution operations performed in similar models, not enough focus has been given to the pooling mechanisms employed in the same models. In this work, the proposed model employs GCNs in the domain of materials science while incorporating a relatively more sophisticated pooling mechanism that takes into account the structure of the crystals being featurized, with which we've observed better results for all the attributes tested as compared to traditional first-principles methods used for computing the same.