Insider Brief
- Researchers developed ChemELLM, a large language model specialized in chemical engineering, which outperformed mainstream LLMs on a new benchmark called ChemEBench.
- The model was adapted from Spark-70B using 19 billion tokens of domain-specific data for pretraining and 1 billion tokens for fine-tuning on chemical engineering tasks.
- ChemEBench evaluates LLMs across basic, advanced, and professional competencies, with ChemELLM demonstrating superior knowledge and problem-solving ability in the field.
PRESS RELEASE — The development of chemical technologies is a multi-stage process that typically begins with laboratory research, progresses through scale-up and basic engineering, and culminates in industrial deployment. This complex process requires synergistic collaboration among experts from diverse disciplines such as chemistry, physics, mathematics, electrical engineering, process design, and architecture to address technical bottlenecks while balancing economic viability.
However, interdisciplinary collaboration is often hindered by disciplinary boundaries, posing significant challenges to maintaining consistency in design intentions during chemical process development.
Emerging strategies such as data-driven artificial intelligence (AI) technologies have gained recognition for their potential to streamline development pipelines and enhance process efficiency. Particularly, the advent of large language models (LLMs), trained on extensive corpora encapsulating complex, cross-disciplinary information, offers unprecedented opportunities to revolutionize scientific workflows.
Recently, a research team led by Prof. Mao Ye (Dalian Institute of Chemical Physics, Chinese Academy of Sciences) & Prof. Xin Li (iFLYTEK Co., Ltd.) has developed ChemELLM, a domain-specialized LLM designed for chemical engineering applications.
Built upon the Spark 70B foundation model, ChemELLM underwent domain-adaptive pretraining and instruction fine-tuning using ChemEData, a carefully curated corpus of high-quality chemical engineering data. Additionally, to assess the knowledge and problem-solving capabilities of LLMs in this filed, the team introduced ChemEBench, a comprehensive benchmark designed for chemical engineering. The results were published in the Chinese Journal of Catalysis (DOI:10.1016/S1872-2067(25)64725-5).
ChemEData, a specialized dataset containing 19 billion tokens for pre-training and 1 billion tokens for fine-tuning, was constructed. Domain pre-training was conducted on the Spark-70B foundation model using a 19-billion-token chemical engineering corpus. This approach enables ChemELLM to acquire domain-specific knowledge while retaining Spark-70B’s foundational capabilities. During the supervised fine-tuning phase, 2.75 million high-quality data (1 billion tokens) were utilized to align the model with the specific language patterns and terminology of chemical engineering.
The ChemEBench benchmark integrates three progressive evaluation stages-basic knowledge, advanced knowledge, and professional skills-to comprehensively assess LLMs in this specialized domain. Evaluation results highlight ChemELLM’s superior performance over mainstream LLMs (including O1-Preview, GPT-4o, and DeepSeek-R1) on ChemEBench, demonstrating its excellence in chemical engineering tasks.
