AI Helps Index Hidden Chemistry in Scientific Images

Chemists have long struggled to find data that exists only as drawings and reaction schemes in patents and journals. According to a report by GN technics/ai (en-US), a new AI vision technology developed by LG AI Research is now integrated into Elsevier’s Reaxys platform. This tool identifies substances and reactions hidden in visual formats, turning them into searchable text. The system aims to stop researchers from spending hours manually checking images to verify if a compound has already been described.

The core problem is that much of the critical chemical knowledge is not written in text but encoded in spatial relationships, bonds, and stereochemistry. Standard search engines fail here because they cannot interpret the meaning of a drawn hexagon or a specific bond angle. By extracting this visual data, the new system makes it possible to search for molecules that were previously invisible to digital queries, a significant shift for fields like synthetic planning and competitive intelligence.

Specialized AI for Complex Structures

The technology behind this advancement is not a general-purpose image reader. LG AI Research built a model that combines molecule detection, reaction diagram parsing, and optical chemical structure recognition into a single system. This approach is necessary because chemical drawings are semantically dense; a slight misinterpretation of a bond can lead to identifying the wrong compound entirely. The model is trained to understand the specific logic of chemical notation, ensuring that the extracted data reflects the actual science rather than just visual shapes.

Accuracy is the primary trade-off in this system. To ensure that the AI does not generate false positives, the extraction pipeline is rigorously validated against existing benchmarks before being deployed. This means the system prioritizes precision over raw speed, which is crucial for scientific research where a single error can derail a project. The validation process ensures that the data added to Reaxis is reliable, even as the scale of extraction grows.

Benefits for Research Workflows

For researchers, the immediate benefit is time. Managing Director Mirit Eldor of Elsevier noted that deciphering figures manually consumes valuable hours that could be spent on discovery. By lifting this chemistry out of the image and into a searchable database, the partnership allows scientists to focus on innovation rather than verification. This is particularly useful in inorganic and organometallic chemistry, where structures are complex and harder to index manually.

However, the current implementation has limits. While it excels at identifying individual substances and properties, it does not yet fully capture the broader context of reactions. This is a known gap that the partners acknowledge. The system is a step forward in making hidden data visible, but it is not a complete solution for all types of chemical inquiry. Researchers still need to interpret the extracted data within their specific experimental context.

Future Focus on Reaction Extraction

The next phase of this collaboration will focus on extracting reaction data from images. Currently, the system is strong at identifying what substances are present, but it is less adept at understanding the dynamic process of how one compound transforms into another. Expanding the capability to parse reaction schemes will broaden the evidence available in Reaxys, providing a more complete picture of chemical processes. This extension aims to address the most complex aspects of visual chemistry.

Elsevier and LG AI Research are also exploring other customer challenges to tackle with this technology. The goal is to continue refining the AI’s ability to handle the nuances of scientific imagery. As the technology matures, it may address more specialized areas of chemistry that have historically been difficult to digitize. The ultimate aim is to create a comprehensive digital library where no chemical insight is lost simply because it was drawn on a page.