Knowledge Extraction and Traceability for Forward-Osmosis Hydrogel Systems: Failure Modes across the Evidence Chain

Abstract

Forward-osmosis hydrogel research connects polymer composition and swelling to osmotic pressure, water flux, solute leakage, regeneration, and fouling, but those relations are reported with uneven context. This review examines how AI-based knowledge extraction can organize the evidence while exposing its failure modes. Domain-adapted language models identify candidate entities and relations; a provenance-aware graph links each statement to the original passage, hydrogel formulation, test condition, comparator, and uncertainty. The synthesis focuses on extraction errors that matter scientifically: treating correlation as mechanism, merging incompatible protocols, losing negative results, and allowing generated explanations to outrun source support. Contradiction detection and human review are used to distinguish competing evidence from simple terminology variation. Graph queries then reveal missing transitions between laboratory draw-agent performance and operational desalination decisions. No new hydrogel experiment or model benchmark is claimed. The article's contribution is a traceable AI workflow that keeps extracted knowledge reviewable, separates reported findings from inference, and turns evidence gaps into explicit priorities for measurement, validation, and safe translation.