Data-driven artificial intelligence (AI) has attracted growing attention as a technology capable of efficiently exploring new materials. However, important experimental data in materials research often exists embedded as images in the figures of scientific papers, making it difficult to utilize effectively.
A research team led by Professor Hao Li and Director Shin-ichi Orimo from the Advanced Institute for Materials Research (AIMR), Tohoku University, together with Research Assistant Ryuhei Sato from the Graduate School of Engineering, the University of Tokyo, has developed DIVE (Descriptive Interpretation of Visual Expression), a multi-agent AI workflow. DIVE is capable of systematically reading experimental data from figures and tables in scientific papers, scientifically interpreting the data, and organizing it in structured form. The team successfully demonstrated that the system can propose new hydrogen storage material candidates from literature in a short amount of time. The findings were published in Chemical Science.
The research team developed the DIVE multi-agent AI workflow with the goal of not just reading figures and tables in scientific papers but also interpreting the extracted data on the basis of scientific reasoning. In DIVE, multiple AI agents with distinct roles—understanding figure content, interpreting captions, and verifying numerical consistency—work in concert to carry out data extraction and validation in a step-by-step manner. This approach achieved substantial improvements not only in accuracy but also in the range of applicability when compared with conventional methods that rely on a single multimodal model extracting everything at once.
In benchmarks for the hydrogen storage materials field, DIVE achieved extraction accuracy 10-15% higher than general commercial multimodal models, and over 30% better than open-source models. The team further organized more than 30,000 data entries from over 4,000 publications to build DigHyd (Digital Hydrogen Platform), an AI agent-based infrastructure capable of systematic analysis. Using this database as a foundation, they constructed an inverse-design workflow for proposing candidate materials, demonstrating the ability to suggest new hydrogen storage material candidates in as little as approximately two minutes.
DIVE can be applied beyond hydrogen storage materials to streamline database construction and materials exploration across diverse fields, including batteries, catalysts, and thermoelectric materials. Future work will focus on expanding compatibility with a wider range of figure/table formats, as well as enhancing autonomous materials design workflows that make use of the extracted data. The team aims to build a new research infrastructure in which AI reads scientific literature—figures and tables included—to accelerate materials discovery.
Journal Information
Publication: Chemical Science
Title: "DIVE" into hydrogen storage materials discovery with AI agents
DOI: 10.1039/D5SC09921H
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.