Failure-Aware Evaluation for Bernoulli-Driven Water Separation: From Laboratory Evidence to Operational Control

Abstract

High average desalination performance does not establish that a Bernoulli-driven system will remain safe or useful under blockage, feed variation, sensor error, or control delay. This review organizes evaluation around failures that matter to operation rather than a single laboratory efficiency measure. It considers AI techniques for rare-event detection, regime-aware classification, calibrated risk estimation, and counterfactual stress testing, while retaining the hydraulic mechanism as a constraint on plausible predictions. Evaluation is decomposed into analytical validity, robustness under distribution shift, calibration, decision utility, and post-release monitoring. The synthesis emphasizes class imbalance, asymmetric error costs, temporal leakage, and the danger of choosing comparators that make normal conditions look representative of the whole operating space. It proposes reporting failure-specific recall, uncertainty intervals, alarm burden, time-to-detection, and recovery behavior alongside conventional flux and rejection metrics. Because the article is a critical synthesis rather than an experimental study, it does not claim a new model result. Its central outcome is an AI evaluation protocol that links laboratory evidence to explicit operating thresholds, human escalation, and revalidation.