China's chemical engineering large language model upgraded with task execution ability

SHENYANG -- The latest version of China's first large language model dedicated to the chemical industry was released on Monday, marking a significant leap from knowledge acquisition to intelligent execution of chemical tasks, and providing new support for managing complex chemical operations, said northeast China's Dalian Institute of Chemical Physics (DICP) under the Chinese Academy of Sciences.

The Chemical Engineering Large Language Model (ChemELLM) 3.0 Pro was jointly developed by DICP, iFlytek, Alibaba Cloud Computing, and other institutions and enterprises.

The ChemELLM is a key component of the DICP's new system for intelligent chemical research and development, aiming to bridge the gap between laboratory research and industrial production. To date, more than 300 chemical enterprises, universities and research institutes have registered to use the model, with cumulative API calls exceeding 14 million.

The latest version of ChemELLM has evolved from an assistant for knowledge acquisition into an intelligent engineering partner that can collaborate with professionals on complex tasks, offering a new foundation for the smart transformation of chemical research and development, engineering design and production operations.

The chemical industry involves complex materials, large spans in scale, diverse data types and stringent engineering constraints, making it difficult for general-purpose large models to meet professional needs.

The new version breaks away from the conventional application model centered on knowledge Q&A and content generation, establishing a four-tier architecture comprising a large model, intelligent agents, professional skills and tools, and application scenarios.

In this architecture, the large model functions as the cognitive core, responsible for professional knowledge comprehension, multimodal information parsing, solution generation and task planning. Intelligent agents act as the execution entities, handling process decomposition, tool invocation, result verification and dynamic adjustments around task objectives, forming a closed loop that integrates cognition, reasoning, planning, execution and validation.

According to assessments by a chemical-domain evaluation system, the ChemELLM 3.0 Pro achieved text-based Q&A accuracy and multimodal Q&A accuracy of 81.96 percent and 80.75 percent, respectively, with overall scores showing respective improvements of 20.2 percent and 31.4 percent compared with the 3.0 version.

The DICP said it will develop the ChemELLM 4.0 with enhanced reasoning and multi-tool collaborative planning capabilities. By using typical mature chemical processes such as methanol-to-olefins as validation scenarios, the institute will integrate key stages including laboratory research, engineering design and plant operations to create a novel paradigm enabling a seamless "lab-to-plant" transition in a single step.