“DIVE” into Hydrogen Storage Materials Discovery with AI Agents

Developing new materials often involves a vast amount of trial and error across different compositions and configurations. In recent years, artificial intelligence (AI) has attracted growing attention in energy-materials research for its potential to accelerate this process. However, fully autonomous workflows that reliably connect high-precision experimental knowledge to the discovery of credible new energy-related materials remain at an early stage.

The research team developed a multi-agent workflow called Descriptive Interpretation of Visual Expression (DIVE) to streamline the materials research process. DIVE automatically extracts experimental information from images in the scientific literature, analyzing a database of more than 30,000 figures collected from over 4,000 publications. Based on this information, the system can propose promising hydrogen-storage materials within minutes.

DIVE systematically organizes experimental knowledge on solid-state hydrogen storage materials that is often embedded in figures and tables. Its data-extraction accuracy and coverage were found to be 10–15% higher than those of commercial models and more than 30% higher than open-source models. The system is designed for ease of use: researchers can interact with DIVE through a conversational interface by specifying desired material properties, allowing the system to rapidly suggest candidate materials. Notably, DIVE also demonstrated the ability to propose previously unreported material compositions.

This work establishes a reliable, end-to-end pipeline that converts key experimental results—traditionally difficult to access because they are embedded in published figures—into high-quality, machine-readable data. Using this pipeline, the team constructed DigHyd, a large curated database of solid-state hydrogen-storage materials derived from thousands of experimental and computational studies. The Digital Hydrogen Platform (DigHyd; www.dighyd.org ) represents the first dedicated digital platform for hydrogen-storage materials design and, to the authors’ knowledge, the largest such database reported to date.

The findings were published in Chemical Science on February 3, 2026.

Comparison between DIVE’s multi-agent workflow and conventional methods, and the distribution of collected publications in the hydrogen storage materials database. ©Hao Li et al.

Papers  
Journal: Chemical Science
Title: “DIVE” into Hydrogen Storage Materials Discovery with AI Agents
Authors: Di Zhang, Xue Jia, Hung Ba Tran, Seong Hoon Jang, Linda Zhang, Ryuhei Sato, Yusuke Hashimoto, Toyoto Sato, Kiyoe Konno, Shin-ichi Orimo, Hao Li
DOI: https://doi.org/10.1039/d5sc09921h