Answers are grounded in this paper and cite their sources.
Flow chemistry literature locks much of its quantitative process data inside figures and tables that show molecular structures as images, including the low-yield and failed reactions that machine learning most needs. Text mining cannot read these formats, and general-purpose vision-language models hallucinate plotted points and do not return the same answer twice. We present FlowFigTabMiner, an automated multimodal pipeline that mines this data into structured records by chaining five custom-trained YOLO models with chemistry-specific tools, a VLM text reader and an LLM adjudication step. On a benchmark of 30 figures and 30 tables it reaches an F1 of 0.874, against 0.848 for Gemini 3.1 Pro and 0.778 for Claude Sonnet 4.6. The advantage is clearest on figures. Applied to organolithium flow chemistry, the mined data enabled a kinetic model whose yield predictions were tested on two substrates outside the corpus. Pipeline and dataset are open-sourced.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: