SHENZHEN, Sept. 12 (Xinhua) -- Chinese researchers have unveiled a lightweight dual-model collaborative AI agent for materials research, enabling AI not only to "think" but also to "act."
The agent, named MatBrain, improves comprehensive prediction accuracy in materials science by 66 percent over large models, including GPT-5 and DeepSeek-R1, while cutting hardware deployment costs by 95 percent. The findings were published in Nature Machine Intelligence on Thursday.
Materials science is the study of the structure, properties and behavior of materials such as metals, ceramics, polymers and composites, with a focus on understanding and developing materials for useful applications.
The lightweight collaborative approach far exceeds frontier general-purpose large models in tasks such as materials property prediction and crystal structure design.
Developed by the Materials Artificial Intelligence Center of the Shenzhen Institutes of Advanced Technology (SIAT) under the Chinese Academy of Sciences, MatBrain comprises two lightweight models: Mat-R1 and Mat-T1.
Mat-R1 functions as a "materials expert," handling knowledge comprehension, scientific reasoning and result evaluation, while Mat-T1 serves as a "research executor," calling materials databases, structure-generation models and computational software. The two models work through repeated cycles of "execute-analyze-feedback-re-execute" to complete research tasks.
Traditional materials research typically requires lengthy development cycles from concept validation to application, involving continuous theoretical analysis, computational screening and experimental trial and error.
In practice, scientific reasoning and tool execution are quite different tasks. A single large model must select among different software tools, construct correct parameters, read intermediate results, and flexibly decide next steps -- with no single effective path that works across tasks.
"It is difficult to excel at both scientific reasoning and research tool invocation within a single lightweight model," said Shi Tongyu, co-author of the paper and a postdoctoral researcher at SIAT. "Switching to general-purpose models with hundreds of billions of parameters can improve research capability, but brings enormous computing and deployment costs."
To address this, the team developed MatBrain. The 30-billion-parameter Mat-R1 handles "thinking" -- understanding material structures and properties, judging whether computational results are reasonable, and deciding the next research direction. The 14-billion-parameter Mat-T1 handles the "doing" -- calling materials databases, crystal structure generation models and property calculation programs.
"The two models have clear roles and division of labor. If results are unreasonable or information is insufficient, Mat-R1 will propose new research requirements for Mat-T1 to continue executing," Shi said.
Results show that MatBrain can complete materials research tasks, including structure generation, property prediction, stability analysis and synthesis route planning, and further participate in open-ended materials discovery.
In electrocatalyst applications, the model completed a full research chain from scientific hypothesis formation and generation of 30,000 candidate structures to multi-stage computational screening and experimental validation. ■
