Uncertainty-Aware Decision Support for Relational Knowledge for Scientific Manufacturing: Lifecycle Evidence and Accountable Decisions

Abstract

Scientific manufacturing decisions increasingly draw on relations extracted from papers, laboratory records, process data, and maintenance narratives, yet an AI system can express uncertain knowledge with misleading precision. This review proposes an uncertainty-aware architecture for relational evidence across the manufacturing lifecycle. Language models identify entities and relation candidates, graph neural networks integrate local structure, and probabilistic layers represent source disagreement, missing context, and distribution shift. Every edge retains a link to its passage, specimen, batch, method, model version, and validation status; generated hypotheses remain visibly distinct from reported findings. Decision support is evaluated by whether uncertainty changes action—requesting evidence, withholding a recommendation, escalating to an expert, or authorizing a bounded use—rather than by extraction accuracy alone. The synthesis also addresses causal overreach, feedback from prior decisions, and governance of updates or deletions. It introduces no new manufacturing dataset. Its contribution is a lifecycle knowledge framework in which AI can connect heterogeneous evidence while provenance, calibrated uncertainty, and named decision authority preserve accountability.