Leveraging Machine Learning to Model Hydrocarbon-CO2 Solubility Behavior

Abstract This study investigates the prediction of alkane solubility in carbon dioxide (CO2) using advanced machine learning methods. Alkane solubility in CO2 plays a crucial role in various industrial applications, including enhanced oil recovery, carbon capture and storage, and natural gas purification. Traditional experimental approaches to measure solubility are often time-consuming and resource-intensive, necessitating the development of robust predictive models. In this work, three machine learning techniques such as: Least Squares Support Vector Machine (LS-SVM), Multilayer Perceptron (MLP), and Gene Expression Programming (GEP) were employed to model the solubility of alkanes in CO2 under varying conditions of temperature, pressure, CO2 density, alkane molecular weight, and carbon number. A comprehensive dataset comprising experimental measurements was used to train and test the models, with 80% of the data allocated for training and 20% for testing. Key statistical metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R² score were used to evaluate model performance. The results reveal that all three machine learning methods achieved high prediction accuracy, with the MLP model demonstrating the best performance (R² = 0.9824). The GEP model, while slightly less accurate, provided valuable insights into the relationships between variables, offering a high degree of interpretability. The SVR method achieved competitive accuracy with the advantage of computational efficiency. This study highlights the potential of machine learning techniques as reliable tools for predicting alkane solubility in CO2, significantly reducing the need for experimental data. The findings have broad implications for optimizing industrial processes and improving the efficiency and sustainability of CO2-based applications.