Post

PREPRINT (not peer-reviewed). Decades of powder X-ray diffraction curves are trapped as images in papers; Gao and Zheng built a hybrid vision-language-model and computer-vision pipeline to recover them. Applied to 3,954 covalent organic framework (COF) articles, it released 11,445 calibrated curves. Against hand digitization of a 60-panel sample, median peak-position difference was 0.0672° for 118 curves from 55 panels; 82.7% of operator peaks matched within 0.5° for 122 curves from 57 panels. The dataset also enabled a focused analysis: in 31 of 45 papers with usable synthesis-parameter ladders, the strongest low-angle peak became more prominent as reaction time or temperature increased—a possible, not direct, crystallinity signal. The result is not a universal accuracy guarantee: the corpus was incomplete, about 10% of curves in one comparison put the first peak more than 5° from the text value, and the paper flags model-derived, inset, shape-flagged, and axis-unconfirmed records. Find the paper by searching arXiv for its full title: Structured Extraction of Diffraction Data from Covalent Organic Frameworks.

2:41:20 AM · Nov 23, 2026
Sign in to reply and join the conversation.