Machine unlearning for water distribution systems: a scoping review and verification-guided transferability framework

Machine unlearning (MU) enables trained models to remove or reduce the influence of selected data and has become important for privacy, data correction, and trustworthy artificial intelligence. This scoping review examines whether and how MU can be transferred to water distribution systems (WDSs), where machine-learning models increasingly support leak detection, demand prediction, hydraulic monitoring, and water-quality assessment. Searches of Scopus, Web of Science, and Google Scholar were performed on 10 December 2024. The primary MU–WDS intersection search identified no eligible direct application. Supplementary screening retained 51 core MU records and 831 WDS-focused ML/DL records, revealing two active but structurally disconnected research domains. Through cross-domain synthesis, this review distinguishes legitimate unlearning targets - faulty, mislabelled, manipulated, obsolete, poisoned, or privacy-sensitive records - from verified leaks, pressure transients, contamination events, and operational changes that should generally be retained as safety-critical evidence. The review maps exact, approximate, partition-based, federated, and certified unlearning approaches against WDS data structures, temporal and spatial dependencies, operational risks, retraining baselines, and validation needs. All proposed MU–WDS linkages remain evidence-informed hypotheses rather than validated applications. Future work should develop verification-oriented benchmarks, compare MU with full retraining, and evaluate forgetting, predictive utility, physical consistency, false-negative risk, auditability, and operational safety.