Porosity Prediction using Bagging Ensemble Machine Learning in CCUS Reservoirs. A Case Study: Darling Basin, Australia

Machine learning (ML), a subset of artificial intelligence, has been utilised in many engineering fields, such as computer engineering, electrical engineering, civil engineering petroleum engineering. Bagging ensemble algorithms have been employed for parameter prediction, as they theoretically outperform traditional ML algorithms. Carbon dioxide capture and storage (CCS) is a strategy implemented to mitigate carbon dioxide emissions. A vital aspect of CCS assessment is determining carbon storage capacity, which estimates the amount of CO ₂ that can be stored in the subsurface. Porosity is a critical parameter in calculating this capacity. In this study, the applicability of regression friendly bagging ensemble ML models; random forest regression (RFR) and extra tree regression (ETR) to estimate porosity of a sandstone layer as part of a CCS program was investigated. RFR models were developed considering caliper log (CAL), gamma ray log (GR), neutron log (NPHI), photoelectric factor log (PE) and deep laterolog (LLD) input features and calculated porosity as targets. Moreover, four traditional (classical) ML models, multilayer perceptron (MLP), support vector regression (SVR), k-nearest neighbor (KNN) and decision tree regression (DTR), were developed to compare them with the bagging ensemble models. The results showed that the RFR model achieved a testing model R ² value of 0.9668, while ETR model achieved a resting model R ² value of 0.9569. The higher R ² value of the RFR model makes it a better choice for predicting porosity in CCS assessment projects. However, if computational time is a critical factor, ETR could be preferable, as it required only1/3 of the computational time that of the RFR model. Furthermore, when the performance of these models was compared with the four traditional ML models the two bagging ensembles distinctly outperformed the traditional models.