A cost-effective framework for drinking groundwater quality prediction based on explainable artificial intelligence

Urbanization and agricultural intensification are accelerating groundwater quality deterioration, yet most predictive studies focus primarily on accuracy while overlooking computational efficiency and economic feasibility, thereby constraining their applicability in sustainable groundwater management. To address these limitations, this study proposes a comprehensive framework for drinking water quality prediction that integrates multi-model prediction with Explainable Artificial Intelligence (XAI) techniques to achieve both efficiency and interpretability assessment. For validation, we collected 77 groundwater samples in a plain region in 2022. Hydrochemical analysis indicated that the HCO₃-Ca type was the predominant hydrochemical facies, accounting for 54.55% of the samples. EWQI results showed that Na⁺ (0.228), SO₄²⁻ (0.218), and NO₃-N (0.188) exhibited higher weights than other parameters, highlighting the significant influence of the mirabilite-rich geological background on groundwater quality. 88.31% of groundwater samples were suitable for direct drinking, with 67.53% rated as excellent and 20.78% as good. In contrast, feature selection and importance analyses based on ElasticNet, SVM, XGBoost, and DNN models consistently identified NO₃-N as the most influential parameter (SHAP contribution >50%), emphasizing the dominant role of anthropogenic inputs in groundwater quality deterioration. Among the evaluated models, the SVM demonstrated the best overall performance, achieving the highest predictive accuracy (R² = 0.997), the lowest uncertainty (mean standard uncertainty = 0.72), and an acceptable time cost. These findings provide a cost-effective and robust framework to support sustainable groundwater management and informed policymaking.