Large language models unleash AI’s potential for autonomous and explainable materials discovery

Tokyo, Japan – Discovering new inorganic materials is central to advancing technologies in catalysis, energy storage, semiconductors, and more. But finding a material with just the right properties is extremely difficult. What if an AI system could think like a human expert, explore this enormous space automatically, and suggest promising new materials on its own?

In a study published this month in Cell Reports Physical Science, researchers from the Institute of Industrial Science, The University of Tokyo announced the development of MatAgent, an AI framework built around a large language model (LLM). This LLM serves as the system’s “brain,” guiding the search for new inorganic materials using natural language reasoning.

While many other AI methods can generate promising materials automatically, such models often act as “black boxes,” offering little insight into why a suggestion was made. In contrast, MatAgent’s LLM is able to reason step by step, explaining its decisions in plain language and adjusting its ideas based on feedback—much like a human researcher would.

“Large language models are very good at reasoning with words,” says lead researcher Izumi Takahara. “We realized that this ability could be applied to materials design. By letting an LLM guide the search and explain its logic, we can make the design process more efficient and more transparent.”

To support this reasoning ability, the team equipped MatAgent with tools inspired by the reasoning process of human experts. The LLM can review what it has tried before, learn from successful attempts, check trends on the periodic table, and consult a database filled with expert explanations. Using this information, the LLM proposes a new material and describes why it might work better than earlier ideas.

“The design of materials is naturally iterative,” adds senior author Teruyasu Mizoguchi. “Giving the LLM a way to remember past steps and refer to expert knowledge allows it to refine its suggestions in a meaningful way.”

MatAgent repeats a simple but powerful loop: it applies reasoning to suggest a new combination of elements, predicts what the atomic structure might look like, and then checks whether the material has the desired properties. This information is fed back into the LLM, which uses it to improve its next idea.

Because the system is based on an LLM, users can give instructions in everyday language—such as “avoid toxic elements” or “only use common elements”—and MatAgent will automatically incorporate these constraints into its search. Importantly, the LLM explains every step it takes, giving users a clear view of the thinking process behind the proposed materials.

The researchers believe that MatAgent will accelerate the discovery of new materials in areas where finding viable candidates is extremely challenging but essential for technological progress. They are now exploring how to extend the system so that it can balance multiple target properties at once as well as consider how easy a proposed material might be to make.

###

The article, “Accelerated Inorganic Materials Design with Generative AI Agents,” was published in Cell Reports Physical Science at 10.1016/j.xcrp.2025.103019.

 

About Institute of Industrial Science, The University of Tokyo The Institute of Industrial Science, The University of Tokyo (UTokyo-IIS) is one of the largest university-attached research institutes in Japan. UTokyo-IIS is comprised of over 120 research laboratories—each headed by a faculty member—and has over 1,200 members (approximately 400 staff and 800 students) actively engaged in education and research. Its activities cover almost all areas of engineering. Since its foundation in 1949, UTokyo-IIS has worked to bridge the huge gaps that exist between academic disciplines and real-world applications.
 

END