Enhancing PFAS data integrity: An LLM-based FAIR+Environmental principle for improved evaluation of environmental contaminants and related constituent databases.

Per- and polyfluoroalkyl substances (PFAS) are persistent, bioaccumulative contaminants of emerging concern, yet data sharing of their environmental occurrence and monitoring remains fragmented. We propose a "FAIR+Environmental" framework that extends the Findable, Accessible, Interoperable, Reusable (FAIR) principles to environmental-specific matrices, assessed by a semi-automated Large Language Model (LLM) pipeline that uses rule-based scoring, few-shot prompting, and Chain-of-Thought (CoT) reasoning. We applied the framework to >100 U.S. PFAS datasets across groundwater, surface water, drinking water, and soil matrices. Few-shot CoT LLMs streamlined FAIR evaluations, reducing the manual effort required for expert assessments. Among environmental matrices, surface water datasets achieved the highest FAIR-Score (53.6 %), followed by drinking water (52.6 %), soil (49.3 %), and groundwater (45.2 %). Multi-state datasets consistently outperformed single-state datasets, particularly in Interoperability and Reusability criteria. Compared to geoscience databases, PFAS environmental datasets lag in FAIR adherence, highlighting the urgent need for centralized, standardized, and FAIR-compliant data management in the field. PFOA occurrence indicated overall PFAS pollution hotspots because of its frequent detection. PFOA contamination was most severe in soil and groundwater. Surface water and drinking water showed lower concentrations but remain critical public health exposures, with ~34 % of drinking water samples exceeding the 4 ng/L maximum contaminant level. Temporal trends indicated little significant change in PFOA concentrations across most states.