Porosity prediction using bagging ensemble machine learning in CCUS reservoirs. A case study: Darling Basin, Australia

Estimating CO₂ storage capacity plays an important role in carbon capture, utilization and storage (CCUS) assessment since the feasibility of the identified reservoir largely depends on it. Porosity is a critical parameter in calculating CO₂ storage capacity and it is generally estimated using core analysis which is expensive and time consuming. As an alternative, machine learning (ML) has been utilized in the literature to estimate porosity. However, one of the widely studied ML algorithm categories, bagging ensembles has not been extensively studied in porosity prediction in CCUS. To address this research gap the study investigated the applicability of two grid search optimized regression-friendly bagging ensemble models: random forest regression (RFR) and extra trees regression (ETR) to estimate porosity of a sandstone layer in Darling basin in Australia and compared their performances with 4 optimized traditional ML models: multilayer perceptron, support vector regression (SVR), k-nearest neighbors regression and decision tree regression. The ML models were developed using 5 well logs and calculated porosity. The study revealed that except SVR rest of the ML models performed robustly against the dataset during porosity prediction. However, RFR and ETR models outperformed the traditional ML models achieving blind test coefficient of determination (R²) values of 0.9668 and 0.9569 respectively. Even though RFR had a higher R² than ETR, its computational time was 3 times that of ETR. Overall, bagging ensembles can be suggested as an alternative to predict porosity of the sandstone layers of Darling basin during CCUS assessment.