AI can now read molecular diagrams from chemistry papers — but not perfectly yet

Researchers fine-tuned an AI model to convert 2D molecular structures from printed chemistry papers into machine-readable code. The model works reasonably well on synthetic examples but still fails on real patent documents compared to the current best approach, which means it's not ready to replace existing chemistry OCR systems.

Chemistry research is buried in millions of printed papers and patents — images of molecular structures that humans can read but computers can't easily parse into usable data. If this problem gets solved, chemists could automatically extract chemical information from decades of published literature and turn it into searchable databases. Right now it's a slow manual process. This attempt shows the AI approach is getting close but hasn't crossed the threshold where it's actually better than specialized chemical software.