A Hybrid Computational Fluid Dynamics – Machine Learning Approach for Characterising CO 2 based Enhanced Oil Recovery in Naturally Fractured Reservoirs

Abstract CO₂ based Enhanced Oil Recovery (EOR) in unconventional reservoirs is an emerging technology. Scientific research efforts are directed towards understanding the propagation of CO₂ front due to the complex interplay between CO₂ injection and saturation, and reservoir’s constitutive relationships. Conventional methods for characterising CO₂-EOR rely on high-fidelity numerical solutions that often result in over or under prediction of CO₂ geosequestration. In this study, we develop a novel hybrid Computational Fluid Dynamics (CFD) and Machine Learning (ML) framework that allows for rapid CO₂ geosequestration prediction and its optimal injection. Very low or high injection rates have been shown to result in low sweep efficiency or excessive entry pressure, while an intermediate injection rate offers the best balance between the two. CFD data-driven Gaussian Process Regression (GPR) and Extreme Gradient Boosting (XGBoost) models have been developed, trained and tested for predicting CO₂ saturation in the reservoir. Comparative analysis indicates that GPR outperforms XGBoost in terms of its predictive performance and robustness. Through the analysis of layer-resolved CO₂ front displacement and development of data-driven surrogate models, this study contributes a novel framework for CO₂-EOR predictive modelling and optimising injection strategies in naturally fractured reservoirs.