Materials research often involves a lengthy process from scientific hypothesis and theoretical analysis to computational screening and experimental validation. In practical research settings, a research agent needs to select appropriate tools, construct accurate parameters, interpret intermediate results and determine the next step based on those results.

In a study published in Nature Machine Intelligence, a team led by YU Xuefeng from the Shenzhen Institutes of Advanced Technology (SIAT) of the Chinese Academy of Sciences developed a lightweight dual-model collaborative artificial intelligence (AI) agent, MatBrain, to support materials research by enabling AI to both reason about scientific problems and execute research tasks.

MatBrain, a lightweight collaborative agent system, consists of two lightweight models, Mat-R1 and Mat-T1, which have distinct roles for reasoning and execution.

Mat-R1, with 30 billion parameters, focuses on the "thinking" process, including understanding material structures and properties, assessing the validity of computational results and determining subsequent research directions. Mat-T1, with 14 billion parameters, focuses on the "doing" process, including retrieving information from materials databases, generating crystal structures and invoking property-calculation programs.

Based on intermediate results, Mat-R1 evaluates whether the results are reasonable and identifies what needs to be done next, while Mat-T1 executes the corresponding research tasks. This division of labor enables the system to dynamically adjust its research process rather than relying on a fixed workflow.

Experiments showed that MatBrain can complete a range of materials research tasks, including structure generation, materials property prediction, stability analysis and synthesis route planning. The system demonstrated its potential for open-ended materials discovery, in which AI needs to iteratively formulate research questions, use scientific tools and evaluate the resulting evidence.

In an electrocatalyst research application, MatBrain completed an end-to-end research workflow, from formulating a scientific hypothesis and generating 30,000 candidate structures to multi-stage computational screening and experimental validation. This highlights the potential of collaborative AI agents to integrate scientific reasoning with research tools.

These findings demonstrate the potential of lightweight collaborative intelligence for advancing materials research capabilities. This study provides a new approach to building AI agents for scientific research, particularly in scenarios requiring both domain-specific reasoning and the interaction with specialized scientific tools.

Researchers operate the AI system. (Image by SIAT)