MIT model links crystal structures to material properties

MIT researchers have developed a multimodal AI model that combines 3D atomic structures with natural language descriptions to help identify crystalline materials for specific applications. The approach connects atomic arrangements and bonding patterns with material properties, processing considerations, and potential uses. Results published in Nature Communications showed better material identification than models trained on text alone.

Existing language models draw heavily on published research, which favors a limited set of popular crystal classes and leaves many structures in computational databases underrepresented. The researchers addressed those gaps by generating descriptive narratives from experimental research and first-principles computational databases. Using contrastive learning, the system compared similar and dissimilar structures and descriptions to align geometric information with language.

In testing, a version trained on the generated narratives learned to connect structural geometry with textual descriptions, while a version trained only on research-paper text did not. The system also outperformed a baseline when asked to find materials with specified properties. For solid-state lithium-ion battery electrolytes, its leading results included materials independently identified as promising candidates. The researchers hope the approach will help materials scientists search more broadly and accelerate discovery.