Data-driven AI approaches are revolutionizing materials discovery, but much valuable data remains locked in unstructured figures and tables within scientific literature. The DIVE (Descriptive Interpretation of Visual Expression) multi-agent workflow addresses this challenge by systematically reading and organizing experimental data from graphical elements, specifically focusing on solid-state hydrogen storage materials. These materials are vital for future clean-energy solutions, making accurate data extraction critical. DIVE outperforms existing multimodal models, achieving 10-15% better accuracy than commercial models and over 30% improvement compared to open-source alternatives. Leveraging a curated database of more than 30,000 entries from 4,000 publications, the workflow enables rapid inverse design, identifying novel hydrogen storage compositions within minutes. This capability not only accelerates materials discovery but also opens pathways for innovation in energy storage technologies. The AI workflow and agent design are adaptable to various material classes, suggesting broad applicability. By transforming unstructured data into actionable insights, DIVE sets a new standard for AI-driven materials research. This approach mitigates risks associated with incomplete or inaccurate data extraction, enhancing reliability. The methodology encourages the integration of AI in scientific workflows, potentially reshaping how materials science research is conducted and expediting the transition to sustainable energy solutions.
👉 Pročitaj original: arXiv AI Papers
