Machine learning-based simulation of groundwater DIC distribution and source apportionment

Groundwater is an important dissolved inorganic carbon (DIC)-rich reservoir and a key link connecting atmospheric, soil, rock, and surface-water carbon pools, thereby playing an important role in regional and global carbon cycling. However, systematic understanding of DIC migration and transformation mechanisms, spatial distribution patterns, and source identification in groundwater remains limited. In this study, a random forest model was used to capture the nonlinear relationships between groundwater DIC and the available spatial and hydrochemical predictors and to generate independently validated DIC estimates at the sampling locations. These estimates were subsequently interpolated in ArcGIS to characterize the regional spatial pattern. Geodetector analysis was used to quantify the spatial explanatory power of the hydrochemical variables, whereas δ¹³C-DIC, pCO₂, hydrochemical characteristics, and the regional hydrogeological setting were jointly used to interpret DIC sources and formation processes. The results showed that, except for the loess area in the southwestern basin, groundwater DIC concentrations generally increased from the surrounding mountainous recharge areas toward the discharge area around Tao'ahazi Lake. Mg²⁺ showed the highest explanatory power for the spatial variability of groundwater DIC among the tested hydrochemical variables, suggesting that Mg-bearing carbonate dissolution and related water-rock interaction were closely associated with DIC enrichment, and interactions among K⁺-F^-, Mg²⁺-F^-, and Mg²⁺-SiO₂ markedly enhanced the explanatory power for DIC variation. This indicates that the DIC distribution was jointly controlled by magnesian carbonate dissolution, clastic-rock weathering, evaporative concentration, and local sulfate-related processes. δ¹³C-DIC values mainly ranged from - 8 to - 1‰ and were concentrated near - 6‰, while most pCO₂ values exceeded the modern atmospheric level, suggesting that groundwater DIC was primarily produced by carbonate mineral dissolution driven by soil CO₂ input. Overall, the spatial variation of groundwater DIC in the Bojianghaizi Basin was mainly constrained by water-rock interaction and mineralization evolution under the regional recharge-runoff-discharge framework. Local sulfate-related processes in mining areas may further enhance rock-carbon input and modify the isotopic composition of DIC. This study provides a scientific basis for identifying groundwater carbon-cycling processes and supporting regional groundwater environmental management in arid and semi-arid regions.