Estimation of the CO2 diffusion coefficient within oil by considering porous media condition: Machine learning approaches

Molecular diffusion plays a significant role as a production mechanism within CO2-EOR methodologies, particularly in naturally fractured, unconventional, and tight reservoirs. Accurate prediction of the CO2 diffusion coefficient in oil under reservoir conditions is crucial for enhancing the CO2-EOR performance in porous media. The CO2 diffusion coefficient can be computed using various methods, including experimental techniques, the Equation of State (EOS) approach, and conventional correlations. However, EOS models and conventional correlations have significant limitations, as they primarily rely on parameters such as pressure, temperature, and fluid properties (e.g., viscosity and density) but overlook the critical influence of porous media characteristics like permeability, porosity, and water saturation. This oversight results in inaccurate predictions, especially under reservoir conditions where these specifications are vital. Experimental techniques, such as the pressure decay method, consider porous media specifications and elevated pressure and temperature conditions, providing realistic measurements. However, these methods are resource-intensive, requiring specialized equipment and over 2 d to measure the CO2 diffusion coefficient for a single set of conditions. Consequently, their practicality is limited for large-scale applications or diverse condition studies. To address these shortcomings, this research employs two advanced artificial intelligence methods-artificial neural networks (ANNs) and least squares support vector machines (LSSVM)-to develop predictive models for the CO2 diffusion coefficient under reservoir conditions. These methods consider oil-saturated porous media containing water, offering a comprehensive and accurate representation of the diffusion process. The developed ANN and LSSVM models demonstrated excellent predictive performance, with coefficients of determination (R-2) of 0.98 and 0.9937, respectively, highlighting the importance of incorporating porous media specifications into predictive models. [GRAPHICS] .