Among geological settings suitable for CO₂ geological storage, saline aquifers offer larger storage capacity than other storage types and, with favorable permeability and porosity, can stably sequester CO₂ over the long term. Saline aquifers are d...
Among geological settings suitable for CO₂ geological storage, saline aquifers offer larger storage capacity than other storage types and, with favorable permeability and porosity, can stably sequester CO₂ over the long term. Saline aquifers are distributed worldwide, and their geological characteristics vary by region, making standardized storage capacity evaluation and the derivation of injection conditions challenging. Accordingly, this study aims to develop a machine-learning-based predictive model capable of evaluating storage potential in saline aquifers by considering diverse geological characteristics and injection conditions.
For training data generation, a base model was constructed using the GEM simulator developed by CMG. A total of 20,000 cases were generated using the Latin Hypercube Sampling (LHS) method embedded in CMOST for geological characteristics such as formation thickness, permeability, and porosity, and the results were converted into output variables.
For each case, the bottomhole pressure (Pwf) computed from the geological characteristics and injection rate was used as training data for the machine learning models. We developed (i) a bottomhole pressure predictive model and (ii) a classification model to determine storage feasibility and the optimal injection rate. The machine learning techniques used were Random Forest, XGBoost, and Support Vector Machine (SVM). Model accuracy for the predictive models was evaluated using R², RMSE, and MAE, while classification model performance was assessed using the confusion matrix and F1-score. As a result, XGBoost achieved the highest accuracy of 99.53% for bottomhole pressure prediction, and for storage feasibility classification, all three techniques achieved accuracy above 93%.
To verify field application feasibility of the developed models, we compared numerical simulation results and machine learning predictions for the Gunsan Basin case in the West Sea of Korea. In the Gunsan Basin-based numerical model, the bottomhole pressure at the end of injection was 23,113 kPa, whereas the machine learning model predicted 24,672 kPa, showing a deviation of approximately 1,500 kPa. This deviation may be attributable to simplifying actual reservoir heterogeneity and geological properties into single average property values for the inputs and not directly incorporating factors that are sensitive to bottomhole pressure, such as the perforation interval/range, as input variables.
In future work, we aim to reduce prediction error by expanding the input variables to include additional factors influencing bottomhole pressure and by constructing training data that more precisely reflect field conditions. The predictive model developed in this study is expected to support pre-assessment and decision-making for evaluating storage potential by considering the geological characteristics and injection conditions of saline aquifers