Vision LLM pipeline turns published COF diffraction figures into reusable data
Researchers describe a pipeline pairing a vision large language model with classical computer vision to extract powder X-ray diffraction curves from published figures, releasing 11,445 calibrated curves in the open COF-PEAK database.
In plain words
Powder X-ray diffraction patterns, the main structural evidence for covalent organic frameworks, are usually published only as images in figures. The authors built a system in which a vision large language model locates each plot and sketches its curves, while classical computer vision refines the traces to pixel level. This converts decades of figure images into numerical data that can be analysed again.
What was announced
The challenge is format variety: PXRD figures range from a single curve to crowded stacks with labels, insets and neighbouring plots that defeat traditional computer vision. The pipeline divides the work, with the vision LLM interpreting figure layout, finding the diffraction plot and its curves, and drawing a coarse trace of each, before classical CV refines every trace to pixel-level accuracy.
Applied to 3,954 COF articles, the extraction produced 11,445 calibrated diffraction curves, released in the open database COF-PEAK. Against hand digitization of 60 panels, extracted peak positions differed by a median of 0.07 degrees. The authors note that papers usually report only a few peak angles, whereas the full curves let researchers track how peak prominence, which they describe as a possible marker of crystallinity, changes with synthesis conditions.
What it does not show
- The source is an abstract only; the full paper's methods, validation details and any additional results are not covered here.
- Accuracy is reported against a hand digitization of 60 panels, so agreement across the full dataset is not established in this excerpt.
- The claimed link between peak prominence and crystallinity is described as possible, not demonstrated, in the abstract.
For security and IT teams
- Access
- The extracted curves are released in the open database COF-PEAK.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.