Improved Assessment and Prediction of Groundwater Drinking Quality Integrating Game Theory and Machine Learning in the Nyangchu River Basin, Southwestern Qinghai–Tibet Plateau

To address the limitations of traditional groundwater quality assessment and prediction methods, this study integrates game theory and machine learning to investigate the drinking quality of groundwater in the southwestern Qinghai-Tibet Plateau. The results showed that the groundwater in the study area is generally weakly alkaline (mean pH: 8.08) and dominated by freshwater (mean TDS: 302.58 mg/L), with hardness levels mostly ranging from soft to medium. Major cations follow the concentration order: Ca²⁺ > Na⁺ > Mg²⁺ > K⁺; anions are in the sequence of HCO₃^- > SO₄^2- > Cl^-. The hydrochemical type is mainly Ca-HCO₃. A few samples exceed the limit values specified in the Groundwater Quality Standard. Through multivariate statistical analysis, ion ratio analysis, and saturation index calculations, water-rock interaction is identified as the primary factor influencing groundwater chemistry. It consists of carbonate dissolution and silicate weathering, accompanied by cation exchange. The water quality index improved based on game theory, integrated subjective weights (from analytic hierarchy process) and objective weights (from entropy-weighted method), shows that the overall groundwater quality in the study area is good: 95.97% of the samples are high-quality water (WQI ≤ 50), more than 99% of the samples have a WQI < 150, which is suitable as drinking water sources; only 0.81% of the samples are of extremely poor quality, presumably related to local pollution. Linear regression achieved the best performance (R² = 0.99, RMSE≈0.00) with strong stability, followed by support vector machines (test R² = 0.98), while the extreme gradient boosting model showed overfitting. This study provides a scientific basis for groundwater management in river basins.