Data Provenance and Reproducibility for Lithium-Sulfur Separator Engineering: Traceable Translation under Uncertainty

Abstract

AI models for lithium-sulfur separator design can reproduce a computation while failing to reproduce the scientific conclusion if material batches, cell assembly, cycling protocol, or preprocessing are missing. This review builds a provenance model that connects physical specimens and experimental conditions to learned features, checkpoints, uncertainty estimates, and engineering decisions. It considers graph-based lineage for separator composition, pore structure, polysulfide behavior, electrochemical signals, and downstream reliability outcomes. The synthesis identifies provenance gaps that distort cross-laboratory comparison, including unrecorded negative results, incompatible baselines, selective cycle windows, and model tuning on the evaluation set. It recommends versioned data packages, executable preprocessing, calibration records, source-level claim links, and explicit transfer assumptions. Where data cannot be centralized, federated analysis is discussed with local custody and auditable update histories. No new dataset is created in this article. The contribution is a reproducibility and AI-governance framework that keeps uncertainty and lineage attached as separator evidence moves from laboratory studies into manufacturing qualification and operational decisions.