Finding the ideal material for an application might theoretically be as easy as ordering dinner thanks to artificial intelligence. Scientists would simply tell a big language model what they wanted, and it would provide them with the ideal material.
Unfortunately, the majority of materials are made up of 3D arrangements of crystals rather than the text streams that are used to train today’s huge language models, which is why that approach isn’t working well yet. AI outputs are skewed toward a few prominent kinds of crystals due to the experimental literature’s strong bias.
Researchers have now developed a multimodal AI model that addresses these issues. The model uses first-principles computational databases and experimental research publications to generate rich linguistic descriptions about 3D atomic structures. The unified description helps AI understand what makes materials effective by outlining not just chemical composition and structure, but also their properties, processing challenges, and potential applications.
By enabling the model to “see” the atomic arrangements and bonding patterns, the researchers demonstrated in a recent publication published in Nature Communications that their approach could identify better material designs than models trained on text streams alone. For instance, the model’s top results included compounds that had been independently determined to be viable candidates when the researchers requested it to find materials suitable for solid-state lithium-ion battery electrolytes. Additionally, the researchers demonstrated how their model was able to learn geometric representations of material structures in a manner that text-only models were unable to.
According to Ju Li, the Carl Richard Soderberg Professor in Power Engineering at the departments of Materials Science and Engineering and Nuclear Science and Engineering, “structure is really at the heart of the materials science field.” You may deduce a lot about the characteristics and functionality of materials if you comprehend their structure. For the model, we aimed to establish the relationship between a material’s atomic structure and its characteristics and processing.
First author and former postdoc Yang Jeong Park, recent University of Illinois Urbana-Champaign computer engineering graduate Mayank Kumaran, Chia-Wei Hsu SM ’26, and Elsa Olivetti, Jerry McAfee (1940) Professor of Engineering at DMSE, collaborate with Li on a paper about the model.
Obtaining 3D Models To View
In order to create the products of the future, materials scientists are constantly searching for novel crystals. AI has changed that process over the past ten years by suggesting and creating new crystalline formations, which comprise over 90% of Earth’s solid materials and have highly organized atomic structures. Solid crystalline materials are treated as 3D graphs in the majority of those models.
The emergence of massive language models has provided researchers with an additional means of creativity. However, while many possible crystal structures are kept in computational databases, current big language models rely largely on text from published articles and books, which limits their understanding of the physical world. Additionally, text-based models do not comprehend visual geometry.
The researchers made the decision to develop a vision-language model for crystalline materials in order to get around such difficulties.
Using a technique called contrastive learning, which compares similar and dissimilar samples of structures and textual descriptions to obtain a cohesive understanding, they trained a language model to provide textual descriptions of crystal structures. This is comparable to aligning the brain’s language and vision regions.
After training, the researchers compared the predictions of two iterations of the model: one that was trained on a dataset that only included text from publications, and the other that was taught on a dataset that included descriptions provided by AI. They discovered that whereas the other model failed to learn how to connect geometric structures with text descriptions, the model trained on AI-generated descriptions did.
According to Li, “we wanted to create an AI model that works broadly across materials.” For this reason, instead of depending just on previously published articles, we chose to synthesize narratives. We discovered that the models are more accurate at predicting the function when we use that narrative dataset.
Materials Science Acceleration
The researchers next demonstrated how their model outperformed a baseline in identifying crystalline materials with a certain attribute when given a natural language query expressing that property.
According to Olivetti, “there are occasionally these step changes within materials science around algorithmic integration toward useful tools, which the community uses to further advance the field.” This method enables us to connect structures to properties and processing and gain a deeper understanding of three-dimensional structures. I’m hoping that this will be one of those adjustments.
Li contrasts the model with AlphaFold, an AI model that has advanced medicine discovery and scientific research by connecting protein structures and functions.
The researchers hope other materials scientists will utilize their innovative method to expedite their work because they think it can work with a variety of crystalline materials.
According to Li, “seeing a three-dimensional crystal and how its atoms are connected is super foundational.” “I believe that anyone who works in materials science will find this useful.”
The National Research Foundation of Korea, the U.S. Defense Advanced Research Projects Agency, and the U.S. Defense Threat Reduction Agency via the University Research Alliance provided partial funding for the project.