Using boosted regression tree models to predict salinity in Mississippi embayment aquifers, central United States

High salinity limits groundwater use in parts of the Mississippi embayment. Machine learning was used to create spatially continuous and three‐dimensional predictions of salinity across drinking‐water aquifers in the embayment. Boosted regression tree (BRT) models, a type of machine learning, were used to predict specific conductance (SC) and chloride (Cl), and total dissolved solids (TDS) was calculated from a correlation with SC. Explanatory variables for BRT models included well location and construction, surficial variables (e.g., soils and land use), and variables extracted from a groundwater‐flow model, including simulated groundwater ages. BRT model fits (r²) were 0.74 (SC and Cl) and 0.62 (TDS). BRT models provided spatially continuous salinity predictions across surficial and deeper aquifers where discrete water‐quality samples were missing. Uncertainty was smaller where salinity was lower, and models tended to underpredict in areas of highest salinity. Despite this, BRT models were able to capture areas of documented high salinity that exceed the TDS secondary maximum contaminant level for drinking water of 500 mg/L. Variables that served as surrogates for position along groundwater flowpaths were the most important predictors, indicating that much of the control on dissolved solids is related to rock‐water interaction as residence time increases. BRT models additionally support hypotheses of both surficial and deep sources of salinity.