A machine learning-based crystal graph network and its application in development of functional materials

Abstract

An active area of MGI (Materials Genome Initiative)/MGE (Materials Genome Engineering) is to accelerate the development of new materials by means of active learning and “digital trial-error” using a prediction model of material property. Machine learning methods have widely been employed for predicting crystalline materials properties with crystal graph neural networks (CGNN). The prediction accuracy of the state-of-the-art (SOTA) CGNN models based on big models and big data is generally higher. However, for the development of some classes of materials, the datasets obtained by experiments are usually lacking due to costly experiments and measurement costs. The lack of datasets will impact the accuracy of CGNN models and may result in overfitting during training models. This paper proposes a simplified crystal graph convolutional neural network (S-CGCNN) which possesses higher prediction accuracy while reducing the vast amount of train datasets and computation costs. The S-CGCNN model has successfully predicted properties of crystalline materials, such as piezoelectric materials and dielectric materials, and increased the prediction accuracy up to 12%–20% than existing SOTA CGNN models. Furthermore, the distribution map between properties and compositions of materials has been built to screen the latent space of candidate materials efficiently by principal component analysis.