SIAT Researchers Unveil MatBrain, First Lightweight Dual-Model Agent for Materials Science
On September 10, a research team led by Professor Yu Xuefeng at the Institute of Materials Artificial Intelligence, Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy of Sciences, published a new study in Nature Machine Intelligence and released MatBrain, described as the first lightweight dual-model collaborative agent for materials research.
MatBrain consists of two lightweight models: Mat-R1, which handles material knowledge understanding, scientific reasoning and result evaluation, and Mat-T1, which manages material databases, structure generation models and computational software. The two models work in an iterative loop of execution, analysis, feedback and re-execution to complete research tasks. According to the team, MatBrain achieves 66% higher comprehensive prediction accuracy than general large models such as GPT-5 and DeepSeek-R1 in materials science tasks, while cutting deployment costs by 95%. Shi Tongyu, the paper's first author and a postdoctoral researcher at SIAT, said it is difficult for a single lightweight model to excel at both scientific reasoning and tool invocation, so a specialized division of labor with model collaboration offers a better solution.
The team is now integrating MatBrain as the "decision brain" of the embodied intelligent robot scientist Dr. Mind, aiming to improve its performance in dexterous hand control, centrifuge operation, precise pipetting, micro powder addition and sample characterization. A trial version of MatBrain has been released for public download.
The team has also obtained a core patent for the dual-model collaborative lightweight materials science method and registered software copyright for the Materials Science Model Collaboration Ecosystem Platform V1.0. The achievement was selected among the Chinese Academy of Sciences' top ten progresses in network and informatization work in 2025. The related technology has been applied to an advanced functional materials pilot platform built by Zhongke Advanced Materials Technology Co., Ltd., which was approved as a national-level manufacturing pilot platform in 2025. In addition, the team has cooperated with Xingfa Group, a leading phosphorus chemical company, to develop the "Xingfa Cloud Brain" intelligent R&D platform for phosphorus chemical and new materials fields.
