Researchers Introduce Dual AI Agent for Materials Research

Shenzhen (ANTARA) - Chinese researchers have launched a lightweight dual-model collaborative AI agent for materials research, enabling AI to not only “think” but also to “act.”

Named MatBrain, the agent comprehensively improves prediction accuracy in materials science by 66 per cent compared to large-scale models, including GPT-5 and DeepSeek-R1, while simultaneously slashing hardware deployment costs by 95 per cent, according to Xinhua, citing the journal Nature Machine Intelligence.

Materials science is the field that studies the structure, properties, and behaviour of various materials, such as metals, ceramics, polymers, and composites, focusing on understanding and developing materials for various beneficial applications.

This lightweight collaborative approach is far superior to leading multipurpose large models in tasks such as predicting material properties and designing crystal structures.

Developed by the Materials Artificial Intelligence Centre of the Shenzhen Institutes of Advanced Technology (SIAT) under the Chinese Academy of Sciences, MatBrain consists of 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 acts as a “research executor” that accesses material databases, structure generation models, and computational software.

The two models operate through an iterative “execution-analysis-feedback-re-execution” cycle to complete research tasks.

Conventional materials research typically requires long development cycles, from concept validation to application, involving continuous theoretical analysis, computation-based screening, and trial-and-error processes through experimentation.

In practice, scientific reasoning and tool usage are very different tasks. A large model must choose between various software tools, arrange the correct parameters, read intermediate results, and flexibly determine the next steps. No single effective pathway can be applied to all types of tasks.

“It is difficult to excel in scientific reasoning and the use of research tools simultaneously within a single lightweight model,” said Shi Tongyu, one of the paper’s authors and a postdoctoral researcher at SIAT.

“Switching to multipurpose models with hundreds of billions of parameters can indeed increase research capabilities, but it incurs enormous computational and deployment costs.”

To address this issue, the team developed MatBrain. Mat-R1, with 30 billion parameters, handles the “thinking” process, which involves understanding material structures and properties, assessing whether computational results are plausible, and determining the next research direction.

Meanwhile, Mat-T1, with 14 billion parameters, handles the “acting” process, which involves accessing material databases, crystal structure generation models, and material property calculation programs.

“Both models have clear roles and a division of labour. If the results are nonsensical or the information is insufficient, Mat-R1 will propose new research requirements so that Mat-T1 can continue its execution,” said Shi.

Research results show that MatBrain is capable of completing materials research tasks, including structure generation, property prediction, stability analysis, and synthesis route planning, and also plays a role in open-ended material discovery.

In the application of electrocatalysts, the model completed the entire research chain, from forming scientific hypotheses and generating 30,000 candidate structures to multi-stage computational screening and experimental validation.