Designing brackish-water desalination plants beyond the range of real data: AI methods and a physics-structured surrogate compared

Brackish-water desalination plants of 100,000–600,000 m³ d⁻¹ are being planned, yet the largest operating brackish plant with published costs we found is about 68,000 m³ d⁻¹, so AI-assisted designs at this scale extrapolate. On a physics-based model with Egyptian feed-water presets, NSGA-II and multi-objective Bayesian optimisation gave equal cost–energy hypervolume, and Bayesian optimisation covered brine-pond area better than 40-member NSGA-II in all 26 runs. Reinforcement learning allocated capacity across four sites better than random but not better than a greedy rule. Data-driven surrogates failed silently beyond their training range; a hybrid that fits a capacity scaling law cut the error from 3.2% to 0.2% but not when the true scaling changed. Against few large-scale references, the model’s capital cost was 51% low to 129% high; an element-by-element membrane simulation (unvalidated) gave 1–118% lower energy use, most at high recovery. Re-optimised with the simulated energy, the cheapest design needed 50–59% less evaporation-pond area. In a value-of-information analysis, once ponds were costed, none of eight uncertain cost inputs changed the best recovery, which was set by the feasible limit. In this study, model uncertainty and missing large-plant data, not the choice of algorithm, limited how far designs can be trusted.