MatBind aligns crystal structure, diffraction, DOS, and text into shared embedding space

Researchers present MatBind, a contrastive learning framework that integrates heterogeneous materials data—crystal structures, simulated powder X-ray diffraction (pXRD), density of states (DOS), and natural language—into a unified embedding space. Using crystal structure as the central physical anchor, the model aligns these modalities to enable emergent zero-shot cross-modal retrieval.

• The framework induces alignment between modalities never explicitly paired during training.
• Retrieval performance improves systematically when multiple modalities are combined at query time.
• The learned embedding space organizes materials according to physically meaningful properties without explicit supervision.

Treating heterogeneous materials data as complementary projections of a single physical reality allows for more effective cross-modal querying and analysis.