Integrating Advanced Machine Learning Models for Accurate Prediction of Porosity and Permeability in Fractured and Vuggy Carbonate Reservoirs: Insights from the Tarim Basin, Northwestern, China

Summary Accurate prediction of porosity and permeability in fractured and vuggy carbonate reservoirs is crucial for optimizing hydrocarbon recovery but remains challenging due to their extreme heterogeneity and anisotropy. Traditional methods often struggle to capture the complex geological variability, leading to suboptimal reservoir characterization. To address this, we propose a novel hybrid machine learning (ML) framework that integrates particle swarm optimization (PSO), mixed-effects random forest (MERF), and ensemble learning models, such as light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), and random forest (RF). These models were trained and validated using leave-one well-out cross-validation (LOO-CV) and the train-test split method, leveraging geophysical well-log data from the Tarim Basin’s carbonate reservoirs. Among the three models, PSO-MERF-LightGBM outperformed the others, achieving an R² of 0.9752 and root mean square error (RMSE) of 0.0606 for porosity prediction and an R2 of 0.9983 and RMSE of 0.00473 for permeability prediction during testing. Moreover, the PSO-MERF-LightGBM model demonstrates exceptional computational efficiency, completing the processing of porosity and permeability in just 11 seconds and 9 seconds, respectively. This marks a significant reduction in computation time compared with other models, making it a highly efficient alternative. These results confirm its superior ability to capture complex nonlinear relationships and spatial variability. The study demonstrates how advanced ML techniques can enhance reservoir characterization, improving decision-making in subsurface resource management. Future research should extend this framework to other geological settings to validate its broader applicability.