Comparison of kriging, machine learning algorithms and classical thermodynamics for correlating the formation conditions for CO2 gas hydrates and semi-clathrates

The correlating capabilities of four machine learning methods, the ordinary kriging method, an adaptive neurofuzzy interference system (ANFIS), a multi-layer artificial neural network (ANN) and a Hybrid of Fuzzy logic and Genetic Algorithm (HFGA), as well as the thermodynamics-based approach of van der Waals-Platteeuw (vdWP) are compared for CO2 gas hydrates formed in the presence of thermodynamic promoters as well as for semiclathrates formed from CO2. These systems were chosen for testing the three methods due to their potential relevance in CO2 capture and due to the expectation of them being computationally challenging. This is the first time that kriging has been tested for correlating gas hydrate equilibrium conditions. Different statistical indices, including the mean square error (MSE), an average absolute relative deviation (AARD), a correlation coefficient, and minimum and maximum errors, are employed to evaluate the performance of these methods. According to these performance indices, the ANFIS method performed the best among these methods; it predicted the equilibrium pressure with the highest accuracy. Finally, an outlier diagnosis is applied to the generated results to specify the reliability and uncertainty of the machine learning-based models. The simple-to-use machine learning tools are shown to be acceptable alternative to the vdWP methods and can be easily coupled with commercial simulation software to reduce calculation times while maintaining the accuracy.