A Multi-Agent Framework for Multimodal Polymer Structure Recognition

Machine-readable polymer structures are essential for data-driven materials research, but polymer 2D-structures reported in the literature are often embedded in figures and accompanied by textual descriptions. Existing structure-recognition methods for simpler small molecules typically overlook polymer attachment sites and are not readily applicable to copolymers. Here, we present a multi-agent workflow that integrates multimodal agents with specialized tools, including detection and structural parsing models, to extract SMILES strings for both homopolymers and copolymers from polymer figures and link them to the textual composition. A small real-world dataset was curated for anion exchange membrane and gas separation polymers, on which the proposed workflow achieves an F1 score of 0.72, with an average cost of approximately \$0.02 per figure, demonstrating the potential of the workflow for practical polymer structure extraction.