Integrated machine learning and 3D reservoir modelling for CO₂ adsorption capacity prediction based on mineralogical composition

Over the last five decades, carbon capture and storage (CCS) has been a popular technology to mitigate carbon emissions. The most critical component for early-stage exploration in CCS applications is identifying, assessing, and selecting suitable sites to capture, transport, and store CO2. The CO2 storage mechanism, adsorption, has gained a lot of attention in CO2 storage capacity estimation. The CO2 molecules generated from the industrial activities are transported to geological storage sites, where it migrates, adsorb onto mineral surfaces, react geochemically, and are stored permanently within the solid mineral surface. Common minerals, such as kaolinite, montmorillonite, illite, and chlorite, exhibit a unique physicochemical structure that offers distinct adsorption capacities. Despite these mineral-specific interactions, conventional CO2 storage capacity estimation methods primarily rely on bulk reservoir properties, such as porosity and permeability, thereby neglecting mineralogical controls on adsorption. In response to this limitation, this research aimed to characterise and utilise mineralogical composition information to estimate the CO2 adsorption capacity. For a petrophysicist, the main challenge lies in the limited well log data available, such as the spectral gamma ray log (SGR) and photoelectric factor (PEF). To address this constraint, the proposed methodology framework intends to apply an advanced machine learning (ML) technique, which is the Extreme Gradient Boosting technique (XGBoost), to develop a mineralogy predictive model using the conventional well logs, including bulk density (RHOB), neutron-porosity (NPHI), and compressional slowness (DTC) logs. The prediction results indicate that the XGBoost-trained model achieves the RMSE, MAE, and R2 score of 0.009,0.021, and 0.884, respectively, on the blind dataset. Notably, the study found that increasing the number of input logs does not necessarily improve accuracy, suggesting that optimized three-log inputs provide a more efficient predictive framework than traditional six-log methods. Analysis of the X Field reveals a lithology dominated by clay minerals (65.29%), particularly illite (59.80%), which was found to have a significantly higher CO2 adsorption potential and stronger reactivity compared to quartz and carbonate minerals. A critical finding of this work is that neglecting mineral-specific interactions can lead to a significant underestimation of the reactive surface area and total CO2 storage capacity. Furthermore, 3D static modeling and simulations show that while CO2 adsorption increases with pressure and temperature under supercritical conditions, the rate of increase diminishes at higher pressures. By integrating mineral-specific adsorption data into a 3D reservoir framework, this approach provides a more accurate assessment of spatial heterogeneity, identifies favorable storage zones, and reduces the risks of overpressure and leakage compared to standard estimation methods. These findings underscore the necessity of mineral-based modeling to enhance long-term storage security and trapping efficiency in geological reservoirs.