Researchers at the Massachusetts Institute of Technology (MIT) have developed an artificial intelligence model that links the three-dimensional structures of crystals with textual descriptions. This approach could help identify materials for batteries and other technologies more quickly, based on their properties and intended applications.

The research findings were published in the journal Nature Communications. The development is intended to address one of the challenges of applying large language models in materials science: they are mainly trained on the texts of scientific publications, whereas the properties of solid materials are largely determined by the three-dimensional arrangement of atoms and the nature of the bonds between them.

As a result, AI may overlook important structural features and may also more often suggest materials from crystal groups that are better represented in the scientific literature.

How AI learns to understand crystal structure

The MIT team created a multimodal model that connects the geometry of atomic structures with natural language. Information obtained from experimental scientific studies and databases generated through computational methods was used for training.

Instead of relying only on already published texts, the researchers compiled detailed textual descriptions of crystals. These reflect not only the chemical composition and structure of a material, but also its properties, processing characteristics, and possible areas of application.

The model was trained using contrastive learning: it matched examples that were close in both structural and semantic terms, learning to establish correspondences between the three-dimensional geometry of crystals and their textual descriptions.

The researchers then compared two versions of the model—one trained on prepared descriptions of materials, and another that used only the texts of scientific publications. The former linked the geometric features of crystals with textual information more effectively and predicted the likely purpose of materials more accurately.

According to MIT professor Ju Li, atomic structure lies at the foundation of materials science: by understanding it, scientists can draw conclusions about a material’s properties and performance characteristics.

Searching for materials for batteries

During testing, the researchers asked the model to find materials suitable as solid electrolytes for lithium-ion batteries. Among the most promising options were materials that had previously been independently identified as potential candidates.

This result shows that the model can use not only information explicitly stated in scientific texts, but also patterns related to the three-dimensional structure of materials.

The scientists also tested how well the system could find crystalline materials based on natural-language queries describing the required properties. In these tests, the model performed better than the baseline approach against which it was compared.

The developers hope the technology will help accelerate the search for new materials and shorten the path from analyzing atomic structure to selecting a material for a specific task. In the long term, such a tool could become for materials science what the AlphaFold model has become for biological research, linking protein structure with function.

At the same time, the development does not yet mean that AI is capable of independently guaranteeing the creation of a suitable material: the proposed options require further scientific validation. The authors expect that specialists working with various crystalline materials will be able to use the new approach.