📊 There's a hidden treasure buried in scientific papers — and most of it is locked inside images. Experimental data embedded in charts and graphs can be read by human eyes, but processing it at scale with computers has been incredibly difficult. Now, a team from Tohoku University and the University of Tokyo has built an AI system called "DIVE" where multiple AI agents work as a team to read scientific figures, extracting over 30,000 data points from 4,000 papers — and proposing new hydrogen storage materials in just 2 minutes.

The Problem: Valuable Data Trapped Inside Paper Figures

In modern materials science, data-driven AI has become a powerful tool for discovering new materials. But there has been a major bottleneck: much of the experimental data that researchers have painstakingly accumulated over decades exists as images — graphs, charts, and tables embedded in scientific papers.

A human researcher can look at a figure in a paper and read the data without much trouble. But getting computers to do this accurately and at massive scale? That's a different story entirely. Existing AI systems — specifically multimodal large language models — have struggled to extract data from scientific figures with sufficient accuracy, and they can only handle a limited range of figure types.

The question that drove this research was straightforward: can we free the data that's been locked inside these figures?

DIVE: An AI "Reading Team" That Works in Stages

In February 2026, a research team led by Professor Hao Li and Director Shin-ichi Orimo at Tohoku University's Advanced Institute for Materials Research (WPI-AIMR), along with Assistant Professor Ryuhei Sato at the University of Tokyo's Graduate School of Engineering, unveiled their solution: a multi-agent AI workflow called DIVE (Descriptive Interpretation of Visual Expression).

What makes DIVE different from conventional approaches is its team-based architecture. Instead of relying on a single AI to do everything at once, DIVE assigns different roles to multiple AI agents, each handling a specific step in the extraction process.

One agent focuses on visually understanding the content of a figure. Another interprets the caption — the text description accompanying the figure — to grasp the scientific context. Yet another agent verifies that the extracted numerical values are scientifically consistent.

This step-by-step approach dramatically improves both accuracy and the range of figure types the system can handle, compared to the traditional method of extracting everything in a single pass.

10–30% More Accurate Than Existing Methods

When the research team benchmarked DIVE against existing methods using hydrogen storage material data, the results were impressive. DIVE achieved extraction accuracy 10–15% higher than standard multimodal AI models. Compared to open-source models, the improvement exceeded 30%.

And it's not just about accuracy. The team reports that DIVE also broadened the range of figures it can handle, compared with the conventional method of pulling everything out in a single pass.

Mining 30,000 Data Points from 4,000 Papers

Using DIVE, the team processed over 4,000 scientific papers related to hydrogen storage materials. The result: more than 30,000 experimental data points, organized into a machine-readable database.

This database, called DigHyd (Digital Hydrogen Platform), has been made publicly available online, meaning researchers worldwide can access and use it for their own hydrogen storage studies.

For those unfamiliar with the field, hydrogen storage materials are substances — often metals or metal alloys — that can safely absorb hydrogen at high density and release it when needed. They are considered essential technology for building a hydrogen-based clean energy infrastructure.

New Material Candidates in Just 2 Minutes

DIVE's value goes well beyond data extraction. The team also built an "inverse design" workflow on top of the DigHyd database. Inverse design flips the traditional research process: instead of testing random materials and hoping for good results, researchers specify the performance they want, and the AI works backward to find materials that match those requirements.

Here's how it works: a researcher inputs the desired material properties; the AI generates candidate compositions from the literature database; a machine learning model predicts hydrogen weight density for each candidate; and the system iteratively refines its suggestions to meet the target specifications.

The entire process — from inputting requirements to receiving new material candidates — takes approximately 2 minutes. For a human researcher, the equivalent work of surveying thousands of papers and identifying matching materials could take weeks or even months.

Beyond Hydrogen: A Platform for All Materials Research

The DIVE technology isn't limited to hydrogen storage. According to the research team, the same approach can be applied to batteries, catalysts, thermoelectric materials, and other fields where valuable experimental data sits locked in paper figures.

Future plans include expanding DIVE to handle even more diverse figure formats and developing more autonomous AI-driven material design workflows. If scientific data trapped in images can be systematically unlocked across all fields, it could mark a new era in materials research.

This research was supported by the Japan Science and Technology Agency (JST) through the "GteX" program and was published in the journal Chemical Science on February 3, 2026.

Toward an Era Where AI "Reads" Research

Data that humans could read but computers struggled to use — that's what was locked inside paper figures. DIVE showed that a team of AI agents can dig it back out. If the approach spreads from hydrogen storage to batteries and catalysts, it could change the pace of materials research itself.

In Japan, using AI as a genuine accelerator for discovery — not just a tool for automation — is becoming a clear trend. How is AI being applied to scientific research in your country? We'd love to hear your perspective.

References