Modeling gypsum (calcium sulfate dihydrate) solubility in aqueous electrolyte solutions using extreme learning machine

Due to various operational conditions, experimental determination of gypsum solubility in water-based systems comprising multiple ionic compounds is often impractical or costly. In this regard, computer-based intelligent approaches can provide highly effective and economical alternatives. In this study, after providing a database with 2288 experimental data-points gathered from 31 various literatures, four advanced artificial intelligence techniques were developed: Adaptive Boosting-Support Vector Regression (AdaBoost-SVR), Extreme Learning Machine (ELM), Gradient Boosting-Support Vector Regression (GB-SVR), and Multivariate Adaptive Regression Splines (MARS). The aim of these techniques was to predict the gypsum solubility in aqueous electrolyte solutions as a function of the solutions' temperature and 22 distinct salts' Molal concentrations. After performing various statistical and graphical analyses, it was observed that the ELM model has a more accurate prediction capability for gypsum solubility in aqueous electrolyte solutions compared to the other three models. For this model, the highest Coefficient of Determination (R2 = 0.9926) and the lowest Root Mean Square Error (RMSE = 0.00373) were obtained. Furthermore, after conducting a sensitivity analysis to calculate the relevancy factors of the ELM model's input and output variables, the variation trend of gypsum solubility was evaluated. According to the multiple evaluations, the ELM model predictions reasonably agreed with the experimental data on gypsum solubility in the chosen solutions and temperature ranges. The ELM model can be implemented in numerical modeling frameworks for improving the treatment and reclamation operations of saline water sources, optimizing the oil and gas recovery and production processes, and controlling the complicated geotechnical issues.