Machine learning-based evaluation and dimensionality optimization of groundwater quality indices for drinking and...

Groundwater is the principal source of drinking and agricultural water supply in the Vietnamese Mekong Delta (VMD), yet it is increasingly threatened by salinization, redox-driven metal mobilization, and unsustainable abstraction. Reliable and cost-effective assessment approaches are therefore needed to support groundwater management in data-limited environments. This study aims to assess groundwater suitability for drinking and irrigation uses and to evaluate whether machine learning can be used not only to reconstruct groundwater quality indices for drinking (EWQI) and irrigation (IWQI), but also to identify their effective predictor dimensionality to support cost-effective monitoring strategies. A dataset of 1386 groundwater samples from nine provinces was used to compute the entropy-weighted EWQI and the composite IWQI, and nine machine-learning algorithms were applied to reconstruct these indices from hydrochemical variables. Overall, 80.6% of samples were classified as excellent to good for drinking use, whereas 19.4% fell into poor to severely hazardous categories, indicating localized but significant groundwater degradation. In contrast, irrigation suitability was much more limited, with only 20.7% of samples classified as suitable to highly suitable, while the majority ranged from marginally suitable to highly unsuitable. These patterns reflect the combined influence of salinity intrusion and redox-driven Fe and NH₄⁺ enrichment on groundwater quality. The machine-learning results showed that ensemble models, particularly Random Forest and XGBoost, achieved high reconstruction performance, with R² values of up to 0.996 for EWQI and 0.95 for IWQI. Analysis of variable subsets further showed that both indices can be effectively approximated using a limited number of dominant variables, with EWQI mainly controlled by Fe, NH₄⁺, and Cl^-, and IWQI by Na⁺, Cl^-, and HCO₃^-. Groundwater in deeper aquifers is generally more suitable for drinking use; however, because these aquifers are overlain by poorer-quality shallow groundwater, careful extraction design and well construction are required to minimize the risk of cross-contamination and downward migration of contaminated groundwater. More broadly, the results provide new insight into the effective dimensionality of groundwater quality indices and highlight the potential of reduced variable sets as efficient surrogates for screening and monitoring design, while not replacing comprehensive hydrochemical analysis.