Exploring the performance of machine learning models to predict carbon monoxide solubility in underground pure/saline water

Abstract

The released carbon monoxide (CO) into the atmosphere is a threat to human life and environmental safety. CO storage in surface and underground seawater/water may be viewed as a potential scenario to decrease the concentration of this dangerous gas in the atmosphere. A reliable tool to calculate CO solubility in aqueous media is a prerequisite for accomplishing such a process. Since the least-squares support vector regression (LSVR), CatBoost, extreme gradient boosting, light gradient boosting, random forest, and extra tree regression can extract even the most complex relationships among a series of independent-dependent variables, they are also potential candidates for modeling CO solubility in pure and saline water as a function of temperature and salt concentration. The present work performs relevancy tests, model construction, the best model selection, accuracy assessment, and trend monitoring using 232 literature records of CO solubility in aquatic solutions containing different salt concentrations. Relevancy analysis by the multiple linear regression as well as Pearson's method approve that CO solubility in water decreases by increasing the temperature and salinity. Moreover, trial and error justified that the LSVR with the Gaussian kernel function has the highest accuracy among the six checked models to estimate CO solubility in aqueous solutions. The acceptable agreement between literature and calculated CO solubility in aquatic solutions is also approved by comprehensive numerical and graphical investigations. According to the results, the LSVR predictions for the CO-water and CO-brine equilibrium behavior correspond well with the literature records (mean square error = 6.18 × 10−8, summation of absolute error = 0.02581 cm3 CO/mL H2O, correlation coefficient = 0.99844, and mean absolute percentage error = 0.48 %).