Comparative analysis of machine learning techniques for estimating groundwater deuterium and oxygen-18 isotopes

Isotope techniques are most frequently used when hydrochemical analysis are insufficient to determine the origin and quality of groundwater and reveal seawater intrusion into groundwater along coastlines. In this study, the potential of the multilayer perceptron, adaptive neuro-fuzzy inference system, generalized regression neural networks, radial basis neural networks, classification and regression tree, Gaussian process regression, multiple linear regression analysis, and support vector machines were compared using known hydrochemical properties of waters for estimating deuterium (delta D) and oxygen-18 (delta O-18) isotopes in groundwater of the Bafra plain, Northern Turkey. The data were divided into training (70%) and testing (30%) sets. Cluster analysis was performed to decrease the number of input variables. The data on electrical conductivity, chloride, magnesium, and sulfate were introduced into the models after examining different combinations of these variables in the studied models. The determination coefficient (R-2), mean absolute error (MAE), and root mean square error (RMSE) were used to evaluate the performances of the models. In addition, visualization techniques (Taylor diagram and heat maps) were prepared to assess the similarities between the measured and estimated delta D and delta O-18 values. The R-2, RMSE, and MAE for delta O-18 (0.98, 0.31 and 0.20 parts per thousand, respectively), and delta D (0.95, 2.85 and 1.89 parts per thousand, respectively) values for the testing datasets revealed that the performance accuracy of multilayer perceptron is the best among the applied models tested. Therefore, the study suggests using data-driven methods, multilayer perceptron in this case, when lacking appropriate laboratory isotope analysis or facing high laboratory analysis costs.