TLDR: A new AI workflow called DIVE extracts and organizes experimental data from scientific literature, especially from figures, with high accuracy. This data is used to build the DigHyd database and an AI agent that can rapidly design novel hydrogen storage materials, significantly speeding up clean energy material discovery.
The quest for clean energy solutions is more urgent than ever, and hydrogen stands out as a promising candidate for large-scale energy storage. However, a major hurdle in its widespread adoption is the lack of efficient, safe, and cost-effective hydrogen storage materials. Traditional methods for discovering these materials are slow, often relying on manual data extraction from scientific literature, where crucial information is frequently embedded in complex figures and tables.
A new breakthrough in artificial intelligence (AI) is set to change this. Researchers have developed a multi-agent AI workflow called Descriptive Interpretation of Visual Expression (DIVE) that systematically reads and organizes experimental data from graphical elements in scientific papers. This innovative approach addresses a critical challenge: unlocking the vast amount of materials data currently trapped in unstructured formats, making it accessible for AI-driven design.
The DIVE workflow is designed to be highly accurate and efficient. Unlike conventional methods that feed entire articles directly into AI models, DIVE first uses a lightweight model to scan figure captions for key information. If relevant figures are found, a second, more powerful AI model extracts specific data points from curves and graphs, transforming visual information into descriptive text. This text then replaces the original figures in the article, allowing a third AI model to perform the final data extraction with much higher precision.
This multi-stage process significantly improves data extraction performance. Benchmarking tests showed that DIVE achieved a 10-15% gain in accuracy and data completeness over leading commercial AI models, and an impressive over 30% improvement compared to open-source models. This enhanced accuracy is crucial for building reliable materials databases.
Leveraging the DIVE workflow, the researchers compiled an extensive database of solid-state hydrogen storage materials. This database, known as the Digital Hydrogen Platform (DigHyd), currently contains over 30,000 entries from more than 4,000 publications spanning from 1972 to 2025. This comprehensive collection provides an unprecedented resource for understanding and designing new materials.
Building on this rich dataset, the team developed an AI agent, also named DigHyd, which can interact with the database using natural language. More importantly, this agent incorporates a machine-learning-based verifier trained on the extracted data. This allows for a rapid inverse design workflow: a researcher can specify desired material properties (like hydrogen density, pressure, and temperature), and the DigHyd agent will propose novel candidate compositions. These candidates are then evaluated by the machine learning model, and the agent can even suggest further optimizations.
Remarkably, this AI-driven process can identify previously unreported hydrogen storage compositions in as little as two minutes. For example, the agent successfully designed optimized Mg2Ni-based systems with small additions of Lanthanum (La) or Yttrium (Y), which are known to improve hydrogen storage properties. It also proposed entirely new compositions like Mg2Fe0.6Co0.2Mn0.2, demonstrating its ability to explore novel material spaces.
Also Read:
- Unraveling Molecular Mysteries: AI Identifies Why Hydrogen Bonds Break
- Retro-Expert: A Framework for Explainable Chemical Synthesis
The DIVE workflow and the DigHyd AI agent represent a significant leap forward in materials discovery. By automating and enhancing the data extraction process, and then integrating this data with powerful AI design capabilities, this framework offers a broadly transferable methodology for accelerating the development of new materials across diverse fields, particularly for critical clean energy technologies. For more details, you can read the full research paper here.
