Machine learning models for the prediction of diffusivities in supercritical CO2 systems

Abstract

The molecular diffusion coefficient is fundamental to estimate dispersion coefficients, convective mass transfer coefficients, etc. Since experimental diffusion data is scarce, there is significant demand for accurate models capable of providing reliable diffusion coefficient estimations.

In this work we applied machine learning algorithms to develop predictive models to estimate diffusivities of solutes in supercritical carbon dioxide. A database of experimental data containing 13 properties for 174 binary systems totaling 4917 data points was used in the training of the models. Five machine learning algorithms were evaluated and the results were compared with three commonly used classic models.

The best results were found using the Gradient Boosted algorithm which showed an average absolute relative deviation (AARD) of 2.58 % (pure prediction). This model has five parameters: temperature, density, solute molar mass, solute critical pressure and solute acentric factor. For the same dataset, the classic Wilke-Chang equation showed AARD of 12.41 %. The developed model is provided as command line program.