Modeling hydrogen solubility in alcohols using machine learning models and equations of state

Abstract

Knowledge of hydrogen (H2) solubility in alcohols is important for designing and performing various processes in chemical plants. Accurate predictions of H2 solubility in alcohols can affect the quality and applications of pharmaceuticals, perfume, cosmetics, flavor, and many others. In this work, deep echo state network (DeepESN), extreme gradient boosting (XGBoost), extreme learning machine (ELM), and multivariate adaptive regression splines (MARS) as four advanced machine learning models were utilized for predicting the H2 solubility in alcohols. To this end, a complete set of H2 solubility data (673 experimental data points) for 26 different alcohols or alcohol-containing solvents is gathered over a wide range of operating pressure (0.101–110.3 MPa) and temperature (213.15–524.9 K). The XGBoost model was obtained as the best model for estimation H2 solubility in alcohols based on graphical and statistical analyses having a root mean square error of 0.0022 and coefficient of determination of 0.9946. Four well-known equations of state (EOSs) were also utilized to estimate H2 solubility in alcohols, among which Redlich-Kwong EOS had the best performance. However, the accuracy of machine learning models was much higher than the EOSs. Based on sensitivity analysis, pressure, temperature, and molecular weight of alcohols have the highest impact on the solubility of H2 in alcohols, respectively. Eventually, the Leverage approach was utilized to recognize the applicability domain of the XGBoost model and probable outlier data, the results of which show that this model has high credit for estimating the solubility of H2 in alcohols. The outcome of this study can help to design the hydrogenation process in chemical plants, and the XGBoost model can act as an efficient predictor for predicting H2 solubility in alcohols.