Evidence Boundaries and Scale-Up for Carbon-Nanotube Desalination Membranes: Traceable Translation under Uncertainty

Abstract

Nanoscale simulations and coupon tests can reveal extraordinary transport behavior in carbon-nanotube membranes, but they do not by themselves predict module-scale reliability or water-system performance. This review defines the evidence boundaries that an AI model must respect when supporting scale-up. It separates molecular transport, membrane fabrication, module hydraulics, feed variability, fouling, cleaning, and downstream water-quality decisions, then examines physics-informed surrogate models and domain-adaptation methods that might bridge those levels. Translation is accepted only when the source and target domains, comparator, uncertainty, and failure consequences are explicit. Out-of-distribution detection and calibrated prediction intervals are used to identify cases where additional measurement is more defensible than a point estimate. A provenance layer records material batches, simulation assumptions, sensor calibration, and model changes so every scale-up recommendation remains traceable. This synthesis reports no new module trial. Its central contribution is an uncertainty-aware AI pathway that distinguishes legitimate cross-scale inference from convenient extrapolation and specifies what evidence is needed before laboratory promise becomes an accountable engineering decision.