Abstract Large‐scale groundwater quality prediction is often constrained by sparse sampling, limiting the reliability of spatial assessments. This study introduces Tabular Prior‐data Fitted Network (TabPFN), a machine learning model based on prior‐fitting networks, to address this challenge and enable high‐precision mapping of groundwater bicarbonate (HCO3−) concentrations across China. Results show that under identical sampling conditions, TabPFN markedly outperforms conventional machine learning models. Its predictive accuracy achieves an R2 of 0.830, compared to 0.384–0.771 for Random Forest, XGBoost, and Support Vector Machines. The root mean square error is reduced to 30.138, relative to 34.603–57.395 for the other models. On this basis, the first nationwide, high‐resolution distribution map of groundwater HCO3− concentrations in China was generated. By integrating explainable artificial intelligence methods, the study further identified key environmental drivers of spatial heterogeneity. Scaling risk assessment for underground pipelines indicates that areas with high HCO3− concentrations, prone to scaling in water conveyance systems, are predominantly distributed in semi‐humid (337.83 mg/L) and semi‐arid (304.35 mg/L) regions of northern and northwestern China. This work provides a novel methodological pathway for groundwater quality assessment under data‐scarce conditions. Beyond methodological advancement, the findings offer a scientific basis for improving groundwater resource management, mitigating scaling‐related infrastructure risks, and supporting policy development for sustainable water use.
Sun et al. (Fri,) studied this question.