Urban drinking-water systems increasingly face a dual challenge: trace metal contamination and disturbed mineral balance, yet many utilities still rely on descriptive monitoring rather than optimisation-driven management. Although machine learning and multi-objective evolutionary optimisation are widely applied in environmental modelling, their end-to-end integration under small-sample monitoring constraints remains under-demonstrated for actionable utility decision support. This study develops a surrogate-assisted multi-objective optimisation framework that transforms routine laboratory measurements into implementable management strategies for urban drinking-water quality. Fourteen household taps across seven distribution zones in Kénitra (Morocco) were analysed for health-relevant trace elements and macro-minerals. Gradient-boosted tree models (XGBoost) were trained under leave-one-out cross-validation to quantify predictive skill under small-sample conditions. Predictive performance was element-dependent, with R² ≈ 0.70 for Ni, 0.53 for P, 0.29 for Cr, and 0.08 for Ag, consistent with stronger signal for Ni/P and attenuated learnability for near-detection and highly variable trace elements. The trained surrogates were then coupled to a four-objective NSGA-III optimisation to simultaneously reduce regulatory exceedance (sanitary risk), compress inter-zone disparities (homogeneity), improve Ca/Mg/Na/K mineral balance, and constrain intervention effort under contrasting sanitary-priority and mineral-priority profiles. The resulting Pareto fronts reveal a narrow compromise region in which sanitary risk and mineral imbalance are jointly suppressed with marginal increases in operational effort. From this region complemented by extreme non-dominated points seven operator-facing scenarios were derived, linking explicit reduction fractions and mineral adjustments to predicted system-wide outcomes (e.g., exceedance objective as low as 0.0023, inter-zone variance down to ~0.0000–0.0004, mineral deviation as low as 10.5075, and effort proxy as low as 2.4162, in the reported objective units). By demonstrating robust optimisation under small-sample conditions typical of municipal monitoring programmes, this study provides a transferable modelling architecture for data-limited urban utilities and strengthens the integration of machine learning with environmental decision-making.