Graph-FINDER converts scientific plots into machine-readable materials data

The multimodal pipeline combines computer vision, natural-language processing and LLMs to digitise multi-line plots without manual calibration. Authors report extracting more than 40,000 datapoints from 874 vanadium-dioxide studies, then testing transfer on 350 Ge–Sb–Te studies.

Why it matters: Much experimental evidence remains trapped in figures. Reliable extraction can turn literature into structured datasets for screening and knowledge graphs.

Impact: Automated extraction still requires provenance, uncertainty flags and human review for ambiguous axes, legends and publication artefacts. npj Computational Materials

Space navigation · Europe/United Kingdom