MIT model links crystal structures to text to improve materials search

MIT researchers developed an AI model that combines three-dimensional atomic structures with language descriptions to identify crystalline materials matching requested properties. In a Nature Communications paper, they reported that the system outperformed a baseline in searches using natural language. The model learns connections between crystal geometry and descriptions covering composition, properties, processing and potential applications. Training with generated descriptions enabled it to align structures with text where a version trained only on research-paper text could not. For solid-state lithium-ion battery electrolytes, its top results included materials independently identified as promising candidates; the demonstration identified candidates rather than establishing their performance in batteries.