Abstract: Mining activities significantly affect ion concentration in groundwater, and changes in mining strategies exhibit notable heterogeneity in their impact on different aquifer lithologies. Traditional methods lack reliability in identifying water sources, since they are based on empirical groundwater chemical characteristics. This study adopts causal inference models to describe the evolution and heterogeneity of water chemical characteristics, and presents a groundwater traceability inference model based on Random Forest (RF) and Generalized Random Forest (GRF). Using nearly 20 years of groundwater chemical data from the Pingshuo mining area, and combining the RF model with data augmentation techniques, we have achieved intelligent traceability of aquifer lithologies with an accuracy of exceeding 97%. The results indicate adjustments in mining strategies have a significant impact on aquifer lithologies, particularly on the water quality from mining voids and sandstone, which exhibits strong heterogeneity in ion concentrations. The heterogeneity further affects the traceability model's classification ability. This study reveals the mechanisms of how certain mining strategy intervention influences the variations in water chemical characteristics in different aquifer lithologies, and helps optimize groundwater resource management in mining areas.

Key words: machine learning, intelligent traceability, geochemistry, generalized random forest (GRF), heterogeneity analysis