【Frontiers of HaiChuan Chemical Technology】Chinese Academy of Sciences develops AI agents in the field of materials science

Thread Content

According to a report from the Dongguan Institute of Materials Science and Technology, Chinese Academy of Sciences (referred to as “Dongguan Materials Institute”) on the 23rd, the institute recently released MatChat 2.0, an AI agent focused on the field of materials science. This agent integrates the capabilities of the Qianwen large model to a high degree, achieving reliable performance in scientific research scenarios with a hallucination rate of 1.3%; its vector database with hundreds of millions of entries enables second-level response times for literature searches, marking a new stage in materials science research. It is reported that MatChat2.0 is built upon over 800,000 carefully selected academic papers, covering all branches of materials science such as new energy materials, electronic information materials, structural materials, and biomedical materials. The vast amount of content in this knowledge base was vectorized using the Qianwen large model, resulting in the creation of a vector database with over 300 million vectors. This constitutes the largest semantic retrieval \"infrastructure\" in the field of materials science in China, enabling precise understanding of technical terms, complex formulas, and implicit knowledge. It can provide professional, intelligent, and reliable support services in various research scenarios such as material synthesis and preparation, performance characterization and testing, academic paper writing, and project application. According to the Dongguan Materials Institute, in research environments where efficiency is of paramount importance, MatChat2.0 reduces the average response time to seconds by introducing dynamic routing algorithms and knowledge retrieval optimization techniques, thereby enabling a transition from \"waiting for retrieval\" to \"instantaneous dialogue\". At its core lies a unique four-layer collaborative architecture: the intelligent context persistence layer, the intent classification and routing layer, the agent reasoning and execution layer, and the academic knowledge database engine. Furthermore, MatChat2.0 reduces the hallucination rate significantly to 1.3% through its unique three-tier verification mechanism. The Dongguan Institute of Materials states that, as an important hub for materials science innovation in the Guangdong-Hong Kong-Macao Greater Bay Area, it focuses its research on areas such as information materials, energy materials, functional ceramics, and advanced metals. It is accelerating its efforts to seize a leading position in the field of new materials, while also laying the groundwork for an innovative ecosystem based on the integration of AI and materials. To date, the institute has developed core application platforms such as the Atomly Materials Science Database, the GPTFF artificial intelligence potential field, and the MatChat AI agent, as well as a range of digital tools including robot scientists, a scientific data open platform, and think tank agents, thereby establishing an integrated closed-loop mechanism for intelligent research in materials science.