Vertical Model for Polyimide Design Assisted by Knowledge-fused Large Language Models

Large Language Models (LLMs) show underexplored potential in specialized scientific domains. We develop domain-knowledge-fused LLMs (D-LLMs) integrated with Vertical Models (VMs) for polymer graph-structure representations. A polyimide-focused text mining assistant automatically extracted 543 experimental parameter sets from 84 articles into a structured database, achieving >91% F1 scores across 8,000 multi-parameter evaluations. Leveraging research process prompting and pre-coding toolkit, a “3T” VM was built to balance three performance parameters--Tensile Strength, Glass Transition Temperature, and Transmittance. This model was employed on an online prediction platform, as well as for the prediction of 20,424 combinations. The structural mapping aligns with high-throughput predictions, providing insights into rational experimental design. This D-LLM/VM framework demonstrates efficient design of high-performance polyimides while establishing an extensible methodology for functional materials and device innovation.