MatSciFig: 391K Materials Science Figure-Panel Pairs Released via Open Pipeline—Decades of Lab Data Unlocked for AI

Materials science figures encode decades of experimental knowledge that is structurally inaccessible to AI—one caption typically spans many sub-panels, making direct image-text pairing impossible at scale. MatSciFig is the first large-scale panel-level multimodal dataset for the domain, produced by an open-source compound-figure decomposition pipeline that achieves 0.9227 mAP₅₀ for panel localization.