Factors controlling groundwater radioactivity in arid environments: An automated machine learning approach

Groundwater resources in the Kingdom of Saudi Arabia (KSA) have high levels of natural radioactivity. Within the northwestern KSA, gross alpha (alpha) and gross beta (j3) levels exceed national and international drinking-water limits. In this study, we developed and used an automated machine learning (AML) approach to quantify relationships be-tween gross alpha and gross j3 activities and different geological, hydrogeological, and geochemical conditions. Two AML model groups (group I for gross alpha; group II for gross j3) were constructed, using water samples collected from 360 irrigation and water supply wells, to define a robust model that explains the spatial variability in gross alpha and gross j3 activities, as well as variables that control the gross activities. Each group contained four model families: deep neural network (DNN), gradient boosting machine (GBM), generalized linear model (GLM), and distributed ran-dom forest (DRF). Model inputs include chemical compositions as well as geological and hydrogeological conditions. Three performance metrics were used to evaluate the models during training and testing: normalized root mean square error (NRMSE), Pearson's correlation coefficient (r), and Nash-Sutcliff efficiency (NSE) coefficient. Results indicate that (1) the GBM model outperformed (training: NRMSE: 0.37 +/- 0.10; r: 0.92 +/- 0.05; NSE: 0.85 +/- 0.09; testing: NRMSE: 0.71 +/- 0.08; r: 0.72 +/- 0.08; NSE: 0.49 +/- 0.12) the DNN, DRF, and GLM models when modelling gross alpha activities; (2) gross alpha activities are controlled by pH, stream density, nitrate, manganese, and vegetation index; (3) the DRF model outperformed (training: NRMSE: 0.41 +/- 0.05; r: 0.92 +/- 0.02; NSE: 0.83 +/- 0.04; testing: NRMSE: 0.67 +/- 0.09; r: 0.77 +/- 0.07; NSE: 0.54 +/- 0.12) the GBM, DNN, and GLM models when modelling gross j3 activities; (4) input variables that affect the gross j3 actives are pH, temperature, stream density, lithology, and nitrate; and (5) no single model could be used to model both gross alpha and gross j3 activities-instead, a combination of AML models should be used. Our computationally efficient approach provides a framework and insights for using AML tech-niques in water quality investigations and promotes more and improved use of different geological, hydrogeological, and geochemical datasets by the scientific community and decision makers to develop guidelines for mitigation.