Knowledge Extraction and Traceability for Bernoulli-Driven Water Separation: Measurement and Comparator Design

Abstract

Evidence on Bernoulli-driven separation is distributed across mechanism studies, measurement reports, manufacturing analogies, and operational guidance. This structured review asks how AI-based knowledge extraction can organize that literature without converting textual association into unsupported causal fact. It outlines a source-grounded pipeline in which domain-adapted language models identify entities such as pressure gradient, channel geometry, flux, rejection, contamination, and failure mode; relation candidates are then linked to exact passages, bibliographic records, and confidence estimates in a provenance-aware knowledge graph. Comparator design is treated as part of the graph rather than an afterthought: each claimed improvement must retain its baseline, boundary conditions, measurement method, and uncertainty. The synthesis discusses human verification, contradiction detection, retrieval-based checking, and separation of reported findings from model-generated inference. It further shows how graph queries can reveal missing comparators and unsupported translation steps before evidence informs design or control. The article reports no new performance benchmark; its contribution is a traceable AI workflow for turning heterogeneous literature into reviewable, decision-relevant knowledge while preserving the limits of every source.