ComProScanner with Gemini-3-Flash achieves 97% accuracy in extracting materials data from scientific figures

Existing automated extraction pipelines for materials science literature have been limited to text and tables, leaving the vast amount of quantitative data locked in scientific figures unexploited. To address this, Aritra Roy and colleagues extended ComProScanner—a fully end-to-end multi-agent framework for composition-property database construction—with native vision-language model (VLM) capabilities. The new extension introduces a FigureExtractor utility that performs caption-keyword-based figure filtering across publishers, and a GraphExtractorTool agent that passes extracted figures to a configurable VLM to recover composition-property pairs from charts and plots. Four VLMs were evaluated based on the LMArena Diagram leaderboard and a cost criterion of under $1.50 per million input tokens.

Benchmarking on 50 piezoelectric ceramic articles from the established d33 test corpus, Gemini-3-Flash-Preview topped performance with a composition accuracy of 0.97 and a normalized F1 score of 0.97, while also being the most cost-effective model among those tested. The team also introduced a range-based value error threshold parameter, providing a more physically meaningful assessment of numeric property values extracted from figures than exact value matching. This work establishes VLM-integrated ComProScanner as the first materials-specific, fully automated, multimodal literature mining platform capable of extracting structured composition-property data from text, tables, and figures within a single unified pipeline.