Elsevier and LG AI Research have deployed a chemistry-specific AI vision model that extracts substance information from images, drawings, and reaction schemes in patents and scientific literature, feeding the data into the Reaxys discovery platform. The technology captures visual chemistry at a scale and speed previously impossible, directly addressing a long-standing bottleneck where critical compound data remained trapped in non-searchable figures.
Researchers in fields like inorganic and organometallic chemistry, where complex structures resist standard text extraction, stand to benefit first. The model combines molecule detection, reaction-diagram parsing, and optical chemical structure recognition (OCSR) in a single system. LG AI Research's published benchmarking shows it outperforms alternatives when extracting chemistry from a full document page.
Why chemical drawings resist standard AI
Chemical structures encode meaning through bonds, atoms, stereochemistry, and spatial relationships. A model that misreads a bond identifies the wrong compound. One that misses a structure gives chemists an incomplete picture. This matters most during novelty searching, competitive intelligence, and synthesis planning-tasks where a single overlooked substance can send a research team down a dead end.
Much of the substance and reaction information chemists rely on is communicated through figures rather than searchable text. Without AI trained specifically on these visual conventions, researchers often check documents by hand to confirm whether a compound or reaction has already been described. The new pipeline runs each extraction against existing Reaxys benchmarks before going live, so accuracy is not traded for scale.
How the extraction pipeline works
The technology from LG AI Research captures substance information from images in both patent and journal content. It feeds into Elsevier's content extraction and curation processes for Reaxys, which already indexes substances, reactions, bioactivities, biological targets, and substance properties. The full pipeline underwent testing across Elsevier's data and workflow tools before wider deployment.
Mirit Eldor, Managing Director, Life Sciences at Elsevier, said: "Every hour a chemist spends deciphering figures or images to see what has already been made is an hour that could instead be spent on chemistry discovery. Our partnership with LG AI Research gives that time back, lifting more chemistry out of the image and into Reaxys - curated, searchable and ready to act on."
What comes next
Reaction extraction is the next stage of the collaboration. This will extend image-based extraction beyond individual substances to broaden the reaction evidence available through Reaxys. The two organizations are also exploring further challenges to tackle together, combining LG AI Research's specialist AI capabilities with Elsevier's chemistry content and curation expertise.
Hwayoung Edward Lee, lead of the AI Biz Transformation Unit at LG AI Research, said: "Understanding scientific images requires AI engineered specifically for chemistry - where every bond and spatial layout holds critical meaning. We designed our AI vision model to decode these complex visual representations with human-expert precision."
The work follows Elsevier's Responsible AI Principles and Privacy Principles. For research scientists looking to build practical skills in applying AI to scientific workflows, the AI Learning Path for Research Scientists offers structured guidance on tools and techniques relevant to data extraction and analysis tasks like those described here.
Why this matters for science and research professionals
A structure buried in a patent figure is no longer a dead end. Chemists who once spent hours manually checking drawings can now search that visual evidence directly in Reaxys. The immediate payoff is faster, more confident decisions during prior-art searches and synthesis planning. For research teams evaluating AI for Science & Research, this deployment shows how domain-specific vision models can turn unstructured scientific images into structured, queryable data without sacrificing the precision that chemistry demands.
Your membership also unlocks:
