Scientific Data Skills Enhance AI Agent Discovery and Interpretation

Scientific data is often fragmented and designed primarily for human consumption, posing significant challenges for AI agents attempting autonomous discovery, interpretation, and utilization. To address this, researchers have developed the Scientific Data Skill (SciDSK), an innovative, agent-ready representation that encapsulates dataset-specific knowledge and operational guidance as a reusable skill for AI agents. A SciDSK integrates comprehensive information such as dataset descriptions, scientific context, file organization, usage procedures, quality checks, and provenance, all while allowing the underlying data to remain in its original repository. A structured specification for SciDSK has been defined, alongside a systematic construction pipeline that grounds each SciDSK in authoritative dataset records. Furthermore, a "Scientific Data Skill Bank" has been established as a unified platform to publish SciDSK resources across six scientific disciplines, ensuring package access, persistent identification, and traceability. Evaluations demonstrate that SciDSK significantly enhances agent-driven dataset discovery and provides more precise and actionable support for dataset interpretation. These findings underscore the substantial value of organizing dataset-specific knowledge in a format optimized for AI interaction, paving the way for more autonomous and efficient scientific data utilization.