The application of supervised machine learning techniques for multivariate modelling of gas component viscosity: A comparative study

Abstract Accurate prediction of gas component viscosity is crucial in gas processing, heat and mass transfer and flow calculations, as well as gas reserves estimation. Many models have been proposed to predict the viscosity of gas components, but they have limited temperature and pressure ranges. Furthermore, machine learning-based methods have not been widely used, and it remains unclear if there are benefits from these approaches. In this study, we explore gas component viscosity prediction versus molecular weight, critical properties, acentric factor, normal boiling point, dipole moment, and temperature using 38 supervised machine learning algorithms. The algorithms are tested by using 4673 data sets for 1602 organic and inorganic gas components collected from the literature. In addition, we compare the outputs of the best predictive model with the viscosity models provided in the literature. The results show that the best performing algorithm is a hidden layer neural network model containing eight neurons with a Bayesian regularization algorithm as a training algorithm and tan-sigmoid and linear transfer functions. Examining the model with an experimental dataset, which has not been used in the model development process, shows that the performance of the proposed model has an absolute mean relative error of 7.2% from experimental measurements.