Materials design based on a material-motif network and heterogeneous graphs

TL;DR

The paper tackles the limitation of traditional structure–property representations by introducing a motif-centric bipartite network that links materials to recurring local coordination motifs. It combines automated motif identification via ChemEnv/CSM with a BiNE-based embedding to capture both explicit material–motif connections and implicit transitive relations, yielding improved formation energy and bandgap predictions and enabling motif-guided screening for various functional classes. Centrality analysis reveals motif hubs that structure the material space into clusters, and the approach demonstrates interpretable, transferable descriptors that can augment high-throughput materials discovery. Overall, motif connectivity provides a compact, interpretive framework that complements existing descriptors and can be extended with additional atomic-scale features to further enhance predictive performance.

Abstract

Machine learning models for functional materials design require precise and informative representations of material systems. Common representations encode atomic composition and bonding but often do not include local coordination environments across chemically diverse crystals. Recurring structural motifs provide a motif level description of crystalline solids and can serve as interpretable descriptors for structure property learning. To analyze the motif connectivity in materials, we construct a bipartite material motif network from 131,548 Materials Project entries, with materials and motifs as the two node sets. Edges connect materials to their constituent motifs and are weighted by motif distortion, which quantifies the strength of each material motif association. Network connectivity is analyzed to identify motif-defined material clusters that capture recurring local geometries relevant to structure property trends. Most shared motifs act as hubs that connect otherwise disconnected regions of the network, enabling motif guided screening by expanding from known motifs to nearby materials in the same neighborhoods. A network embedding step converts this weighted connectivity into vector representations of materials. Using these motif informed embeddings, property prediction yields a formation energy mean absolute error (MAE) of 0.157 eV per atom and a bandgap MAE of 0.601 eV. These results indicate that motif connectivity provides a compact, interpretable representation that complements existing descriptors for scalable screening and structure property modeling.

Figures (6)

• Figure 1: A sample network for a small set of complex metal oxides (red circular nodes) and their structure motifs with different motif types. Only most common motif types are shown and remaining ones are included as other types.
• Figure 2:
• Figure 3:
• Figure 4:
• Figure 6: Sub-networks showing material clusters based on shared motifs for (a) solar energy conversion, (b) transparent conducting oxides, (c) superconductivity, and (d) battery electrodes. Motif nodes are shown in blue and material nodes in red. Each subgraph highlights representative materials previously studied for the indicated functionality and their associated motifs.