Integrated assessment of corrosion behaviour in drinking water distribution systems using experimental analysis, machine learning, and a unified corrosion-scaling index.

Corrosion and scaling in drinking water distribution systems negatively affects the pipe lifespan, water quality, and cost of maintenance. Empirical corrosion and scaling indices are derived based on physicochemical water parameters. Each of these indices comes from different chemical assumptions and interpretation ranges. They provide conflicting results for the same sample of water. This study employed a hybrid methodology integrating experimental investigation, statistical analysis, and machine learning modeling toward enhanced evaluation of corrosion. Water samples were collected during an extended monitoring period from a real, running drinking water distribution system and analysed for their physicochemical properties. Corrosion rate determination was performed experimentally by the weight-loss method; an entropy-weighted water quality index was also calculated. The results were validated against the experimentally taken corrosion rate. A gradient boosting machine learning (extreme gradient boosting) model was used. It predicted the corrosion behavior with high accuracy. To resolve the inconsistency problem between individual indices, a unified corrosion-scaling index (UCSI) is developed by normalising all of them to a common scale for corrosion tendency and then averaging their values into a single, easily interpretable metric. This index provides each water sample a clear corrosive, neutral, or scale-forming classification. UCSI exhibited stable and consistent classification across both training datasets (neutral or borderline: 42.12%-50.76%, scale-forming: 25.76%-36.36%, and corrosive: 21.52%-23.48%) and testing datasets (neutral or borderline: 44.52%-45.45%, scale-forming: 39.39%-39.42%, and corrosive: 15.16%-16.06%). These results show close agreement between the observed and model-predicted results.