Predicting Groundwater Hydrochemical Facies in Three Dimensions with Random Forest Classification, USA

Abstract

A random forest classification (RFC) model was developed to predict hydrochemical facies (HCFs) of groundwater in three dimensions across the conterminous United States (CONUS). Major-ion data from 152,673 sites were used to categorize groundwater into one of six HCFs (CaMg-HCO₃, NaK-HCO₃, CaMg-SO₄, NaK-SO₄, Cl, or Mixed). These six HCFs were used as targets for RFC modeling. Model features that represent relevant geochemical processes and/or physical conditions were derived from previously published data. Additional model features were specifically engineered to support this analysis: elevation of the bottom of a well relative to the base of drinking water (ERDW) and six flags that relate geologic units to HCFs. The most important model feature was ERDW. The model was used to map HCFs at a 1-km² resolution across CONUS and to depths of 400 m below the base of drinking water (which varies from 22 m to 2 km). Model predictions are consistent with expectations. CaMg-HCO₃ is predicted to occur near the water table in more humid settings, and areas underlain by carbonate or crystalline rocks. At depths below the base of drinking-water supplies, the model predicts a rapid transition from HCO₃ HCFs to Cl. Model predictions are accurate based on point data, and data averaged across hydrogeologic regions and with depth. Model predictions of HCFs could be used for multiple purposes, including the mapping of salinity and other groundwater characteristics.