Structural Motif-Based Material Network for Material Discovery and Property Prediction

The effectiveness of machine learning (ML) algorithms in material science depends on the precise and accurate representation of material systems. Structure motifs are considered structure descriptors of solid-state materials and are strong predictors of material properties. This work introduces a novel approach for constructing a network of 145,249 solid-state materials within the Materials Project database connected by common structure motifs. Network analysis shows that the most shared motifs act as hubs, effectively linking several materials within the network. We utilize a bipartite network embedding technique to obtain high-dimensional vector representation of both material and motif nodes, capturing both direct and transitive links in the network. The t-SNE-transformed embeddings exhibited distinct clustering patterns for motifs of different types and for materials sharing most common motifs in the network. This clustering behavior highlights the repetitive nature of structural motifs and their critical role as indicators of specific material properties. The learned embeddings, when used in a neural network model, can effectively predict material properties such as formation energies and band gaps and classify metals and non-metals. The combination of t-SNE visualization, property prediction, and classification shows the crucial role of structural motifs in understanding material behavior, predicting properties, and classifying materials. Our approach provides a robust framework that integrates ML techniques with structural motif information to explore and categorize vast material spaces, accelerating discovery and design of novel functional materials.