New connectionist tools for prediction of CO2 diffusion coefficient in brine at high pressure and temperature ― implications for CO2 sequestration in deep saline aquifers

Saline aquifers are considered as a prime target for permanent geological storage of anthropogenic CO2 owing their relative abundance and suitable storage capacity. CO2 diffusion coefficient in brine is a vital parameter in modeling CO2 sequestration operations. Although, CO2 diffusion coefficient can be determined precisely in laboratory but such tests are constrained by their relatively high cost, time intensive nature, and safety concerns because of high-pressure high-temperature tests. Development of precise, fast, and robust models and correlations for a broad range of pressure and temperature is vital. The main objective of this research work was to develop connectionist predictive models and a correlation using four neural network algorithms to predict CO2 diffusion coefficient in CO2-brine system for pressures of up to 100 MPa and temperatures of up to 673 degrees K (novelty). The Artificial Intelligence (AI) algorithms used include multilayer perceptron (MLP), cascade forward neural network (CFNN), recurrent neural network (RNN), and gene expression programming (GEP). These algorithms have never been used before for this purpose (novelty). A database was assembled with 191 experimental and molecular dynamics (MD) simulation data reported in the literature. This is so far the largest database ever reported on this subject. The input parameters include pressure, temperature, and density of brine determined using a rigorous feature selection strategy and with considering of physics of CO2-brine storage conditions in deep geological formations. Density of brine has never been considered before in the existing data modeling works which is a major drawback as it has the greatest impact on diffusion coefficient of CO2 in brine. The Bayesian regularization algorithm was used to optimize the developed smart models. The MLP model yielded the most accurate predictions with a root mean squared error (RMSE) and coefficient of determination (R2) of 2.945 and 0.998, respectively. The proposed MLP model is superior to the existing smart models and correlations considering both precision and range of pressure and temperature. Having an accurate estimation of CO2 diffusion in brine can be quite useful during preliminary assessment and modeling of CO2 sequestration operations in deep saline aquifers.