Prediction of geothermal originated boron contamination by deep learning approach: at Western Anatolia Geothermal Systems in Turkey

Geothermal fluids consist of hot water, steam and gases in water-dominated reservoirs. They contain various dissolved major elements such as sodium, potassium, calcium, silica, bicarbonate, carbonate, chlorine, sulphate and minor elements such as boron, fluorine, lithium, iron, arsenic, mercury and bromine at different concentrations in the liquid phase. The concentration of dissolved solids depends on the temperature, gas content, reservoir geology, permeability, water mixing and fluid source of a geothermal system. Some of these species exhibit a toxic effect at high concentrations and require precaution after the discharging of geothermal water. Boron is one of the important constituents and can be observed as boric acid (H₃BO₃) or HBO₂ in the water phase. The concentration of B changes between 10 and 50 ppm in chloride-type fluids and can occur in greater quantities than these values in organic-rich sedimentary rocks in geothermal fluids. Although boron is considered toxic, it is also one of the crude minerals and can be used in different industries, such as oil and gas chemistry, vehicle technologies, agriculture, ceramics, and adhesive and coating, among others. Machine learning is a method of data analytics for identifying patterns in data and using them to automatically make predictions about new data points. Deep learning is a machine learning subset that uses artificial neural networks with multiple layers. Deep learning can automatically learn representations from data without hand-coded rules or domain knowledge; this is the primary difference between deep learning and traditional machine learning techniques. In this study, a deep neural network model has been developed to predict boron concentrations based on hydrogeochemistry data for different geothermal systems. To compare the prediction performance of our proposed deep neural network model, two well-known regression approaches, linear regression and linear support vector machine (SVM), were performed, and the results have been presented. The performance comparison revealed that our deep neural network (DNN) model achieved better prediction performance than traditional machine learning techniques—linear regression and linear SVM.