MIT Develops Multimodal Model That Teaches Crystal Structures to AI

Researchers at the Massachusetts Institute of Technology (MIT) have developed a multimodal artificial intelligence model that allows AI to better understand and characterize materials based on their crystal structures. Published in the journal Nature Communications, the study demonstrated that by allowing the model to "see" 3D atomic arrangements and bonding patterns, it can identify better material designs compared to text-based models.

Although artificial intelligence holds the potential to facilitate the discovery of the right material, this approach does not yet operate with full efficiency because the data streams on which current large language models are trained are text-oriented, while most materials consist of 3D crystal arrangements. Furthermore, the bias of the experimental literature toward a small number of popular crystal classes steers AI outputs toward these materials. The MIT team addressed this issue with a new model that combines 3D atomic structures with rich linguistic descriptions from experimental research papers and computational databases.

The developed system provides a unified description that includes not only the chemical composition and structure of a material, but also its properties, processing considerations, and potential applications. When asked for suitable materials for solid-state lithium-ion battery electrodes, the top results from the system independently included materials identified as promising candidates. The researchers noted that the model learns to align structures with text descriptions through a contrastive learning approach.

Ju Li, one of the authors of the study, stated that the AI model could assume a role similar to AlphaFold, a model that correlates protein structures with function and advances biological research and drug development processes. The researchers hope that the approach can work with many crystal materials and help accelerate the work of other materials scientists.