Collaborative Research: Disciplinary Improvements: Agentic FAIR: Context, Skills, and Benchmarks for Trustworthy, Reproducible Materials Research

This project will establish new technical foundations for trustworthy agentic science in materials research by integrating formalized experimental context, reusable skill abstractions, and benchmark-driven evaluation. It include methods for representing scientific context in ways that support machine-actionable provenance and policy-aware execution; modular agent skills that can be composed across experimental, data, and computational tasks; and benchmark suites that measure reproducibility, robustness, and scientific fidelity. These advances will help move agentic materials research from promising demonstrations toward a more rigorous, testable, and cumulative research paradigm. The project (Agentic FAIR) addresses a central bottleneck in data-intensive materials research: current AI-enabled workflows remain difficult to trust, reproduce, and reuse across instruments, laboratories, and organizations because the scientific context needed for reliable automated action is fragmented, inconsistently represented, and rarely benchmarked. This project will develop a coherent foundation for agentic, FAIR, and reproducible materials research by advancing three tightly coupled elements: (1) machine-actionable context specifications that capture the entities, provenance, life-cycle stages, governance constraints, and repository expectations needed for reliable automated action on materials data and models; (2) a curated library of reusable agentic skills as bounded, testable modules following a formal (C, , T, R) contract, for recurring FAIR and research-data-management operations including metadata extraction, schema crosswalk, provenance capture, FAIR packaging, and quality assessment; and (3) benchmark suites and scoring protocols that evaluate whether agentic systems can perform realistic materials-research tasks with consistent, traceable, and scientifically defensible behavior across datasets, instruments, and institutional settings. The work builds on community infrastructure established through the Materials Research Data Alliance (MaRDA) and its NSF FAIROS-funded Materials Research Coordination Network (MaRCN). It targets integration with existing national cyberinfrastructure including PARADIM, the Materials Data Facility, Garden, MaterialsMine, and NanoHUB. Case studies in experimental tensile testing, computational DFT workflows, and ML model packaging will validate the framework across the materials development continuum. All artifacts (schemas, software, benchmarks, and training materials) will be released openly to support adoption by the broader materials and open-science communities. This project is a new foundation at the intersection of FAIR infrastructure and agentic AI and formalizes the contextual, procedural, and evaluative requirements for trustworthy autonomous action in materials research. It includes: a five-layer context model (research-object, lifecycle, provenance, standards/schema, and user/task); the first community-curated skill library for materials FAIR/RDM operations, specified with applicability conditions, termination criteria, provenance side-effects, and failure modes; and benchmark protocols that assess both task completion and scientific fidelity and robustness. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.