Machine Learning Approach to Estimate the Diffusion Coefficient of CO2 in Hydrocarbons

Abstract

Diffusion of the gas into the liquids is a critical part in understanding multiphase systems and engineering applications associated with these multiphase systems. The study couples multiphase pendant drop experiments and computational modelling to calculate the CO2 diffusion coefficient in n-decane. Experiments were carried out at a varied range of pressure and temperature 25–45°C and 25–65 bar. During the experiments, the change in the volume of the hydrocarbon drop due to CO2 diffusion was dynamically measured, and numerical model was developed which used the experimental data to estimate the diffusion coefficient. The current study brings in the capability of machine learning as a replacement of the computational part for prediction of the diffusion coefficient of the process. The feasibility of various machine learning models such as Gradient boosting, Gaussian Process Regression (GPR), k-NN, Decision tree etc. are checked. Firstly different algorithms were trained on the dataset and finally evaluated on the test dataset, using various statistical metrics). Finally, the most accurate algorithm is used as a surrogate model for predicting the diffusion coefficient. The chosen ML algorithm was fairly accurate in predicting the diffusion coefficient with a maximum inaccuracy of 7.5%. Therefore, ML may then be employed as an alternative to experiments and numerical methods. A case study is performed to demonstrate the proposed methodology.