Unsupervised Motif Discovery in Zintl Phases via Sparse Hierarchical Graph Learning

<jats:p>Graph neural networks (GNNs) are a highly effective way to learn structureproperty relationships for materials, but most architectures overlook physical intuition that properties are governed by substructures and motifs rather than individual atom interactions. Here, we show that motifs can be learned automatically by training a sparse hierarchical GNN end-to-end using only property labels. Unlike attention-based methods, ours is interpretable by design rather than through post-hoc analysis, and unlike existing motif-based approaches, it uses no hand-crafted chemical heuristics. We demonstrate this on Zintl phases, intermetallics of strong interest for thermoelectrics whose properties are driven by anionic and cationic substructure interactions. Our model matches imperfect chemical heuristic pooling by up to 9× better than random assignment, with discrepancies that are justifiable rather than arbitrary, enabling interpretable substructure-property relationships and informative priors downstream for generative materials design.</jats:p>