Computational Fluid Dynamics Driven Mass Transfer Modelling of CO2 Corrosion in Complex Flow Geometries

Oil and gas industries, during the extraction of oil from wells, find water along with carbon dioxide (CO2), hydrogen sulphides (H2S), and organic acids. The presence of these species affects the integrity of pipelines used for transportation due to the electrochemical reactions on the metal surface. This phenomenon of degradation is known as carbon dioxide (sweet) corrosion when the presence of CO2 is much higher than H2S. Several studies have shown a comprehensive understanding of the mechanism enabling the development of mechanistic predictive tools with the help of empirical correlations of viscous sublayer thickness, turbulent diffusivity, and reaction rate kinetics. These empirical correlations are valid for fully developed flow situations and constrain the corrosion rate predictions in developing and disturbed flow conditions. This limits the scope of current prediction tools in the literature and generates a need to develop a flow and mass transfer coupled model applicable to a broader range of operating conditions. Hence, this is the subject of investigation in this research work.

A computational fluid dynamics (CFD) driven mass transfer model is developed to predict CO2 corrosion in pipelines. The model involves accurate predictions of viscous sublayer thickness and turbulent diffusivity in a horizontal pipe using CFD. These predictions then drive the 1-dimensional mass transfer model to predict CO2 corrosion. These predictions were then verified with the experimental dataset available in the literature for pH 4 to 6, velocity 1 to 10 m/s, partial pressure of CO2 (pCO2) of 1 bar and temperature of 20°C. A verification with an experimental dataset highlighted the robustness of the CFD-driven model, as the predicted values are well within the range of experimental data.

Machine learning models such as Artificial Neural Network (ANN), Gaussian Process Regression (GPR), Random Forest (RF), and Support Vector Regression (SVR) are applied to predict corrosion rate based on input variables such as pH, velocity, temperature and pCO2. Machine learning-enabled surrogate modelling is used to determine the sensitivity of electrochemical reaction rate constants and then calculate a reliable set of electrochemical reaction rate constants. Random Latin Hypercube (RLH) sampling was then applied to a range of electrochemical reaction rate constants obtained from the literature. A dimensionality reduction technique, Principal Component Analysis (PCA), is then applied to check if the initial design variables can be reduced. An optimal machine learning model is selected from ANN, GPR, RF and SVR based on evaluation metrics. A set of optimal electrochemical reaction rate constants was then used to compare them against the experimental dataset, and predictions were obtained using the current set of electrochemical reaction rate constants.

When it comes to the predictions of corrosion rates in complex flow situations, few prediction tools are available in the literature. A CFD-driven mass transfer model for the prediction of CO2 corrosion in complex flow situations is developed in the current study, which accurately couples the CFD model with the mass transfer model by setting the benchmark for complex flow corrosion modelling. The complex flow situation model considered here is a 2D expansion/constriction pipe in which expansion and constriction domains are connected gradually. This CFD model is coupled with the 1D mass transfer model that calculates the corrosion rate over a surface.

CFD-driven mass transfer model for the prediction of CO2 corrosion in horizontal pipelines predicted corrosion rates reasonably well for pH 5 and pH 6. However, for pH 4 it was found that the corrosion rate predictions were sensitive to the choice of electrochemical reaction rate constants. A systematic approach to finding an optimal set of electrochemical reaction rate constants for pH 4 with the help of supervised machine learning models showed that the GPR model consistently provided lower RMSE values compared to other models. An optimal set of electrochemical reaction rate constants provided the lowest RMSE value of 0.28 between predicted corrosion rates and experimental corrosion rates, showing the robustness of this approach. The current model has shown its capability to predict VSL conditions and corrosion rates in complex flow geometry situations.