Picture a world where identifying the ideal material for a specific application can be as easy as ordering pizza. Theoretically, artificial intelligence could drastically simplify this task, as researchers could use a large language model to describe their requirements and then receive the requisite crystal structure. Despite this, reality tarnishes this idealised concept – especially when it comes to data representation.
Most of the materials we use daily are composed of crystals – elaborate 3D atom configurations – and storing these structures tends to require specialised formats. This is considerably different from the regular text-based streams large language models currently train on. Additionally, the majority of experimental literature seems to favour a handful of crystal classes, inadvertently introducing bias into AI outputs.
Help is, however, on the horizon. A group of MIT researchers have pioneered an innovative multimodal AI model that seamlessly merges 3D atomic structures with natural language descriptions. This results in thorough linguistic narratives extracted from experimental research papers and computational databases. These extensive descriptions encompass everything from a material’s chemical composition and structure to its properties, processing requirements, and potential uses – massively enhancing AI’s understanding of influential materials.
In a publication in Nature Communications, these researchers show their system’s capability in creating superior materials compared to models that only learn from text. They’ve opened the model’s eyes to atomic arrangements and bonding patterns, which has led to the successful identification of material fitting for solid-state lithium-ion battery electrolytes – matching up perfectly with independently recognized candidates that hold promise. Furthermore, their model mastered the representation of geometric structures that could not be captured by text-only models.
Learning about this accomplishment, Ju Li, a professor in Power Engineering at MIT, shed light on how pivotal understanding the structure of materials is in their science. “Bridging the gap between a material’s atomic structure and its properties and processing was our objective for the model,” stated Li.
For eons, materials scientists have been on the hunt for fresh crystals to create tomorrow’s products. Over the past decade, AI has been instrumental in propelling this search, by helping design new crystalline structures comprising of more than 90 percent of the Earth’s solid materials. Unfortunately, the existing language models are largely confined to textual data drawn from published works, thus limiting their understanding of the physical realm of potential crystal structures held in computational databases.
Undeterred by this challenge, the MIT researchers fashioned a vision-language model specially intended for crystalline materials. Utilizing contrastive learning, they trained a language model to create textual descriptions of crystal structures. They even tested two versions of the model – one that utilized AI-generated descriptions and another that used texts from research papers. The AI-generated version demonstrated successful correlations between text and geometric structures.
The researchers also established their model’s abilities to process natural language inquiries and trace crystalline materials with particular properties outpacing other base models. Key project member Elsa Olivetti saw immense potential in the model remarking, “Our method enables us to explore three-dimensional structures in greater depth and create links to their properties and processing. This breakthrough can significantly propel the progress of materials science.”
Li drew comparisons of their model with AlphaFold, another AI model popular for mapping protein structures and functions attracting considerable attention in the fields of biological research and drug discovery. The team is optimistic about their technology being a boon for a large pool of crystalline materials and hope this will expedite advancements in the field.
Acknowledging the importance of this effort and its potential benefits, Ju Li expressed, “Being able to visualize a three-dimensional crystal and its atomic connections is fundamentally crucial. Anyone in material science will find this tool invaluable.”
This revolutionary work was made possible with the support of the U.S. Defense Threat Reduction Agency, the U.S. Defense Advanced Research Projects Agency, and the National Research Foundation of Korea. For more details, visit the original news article.
If you’re keen to tap the potential of AI automation for your business, delve into more with implementi.ai.
This website uses cookies.