Permeability Prediction of Strong Heterogeneous Carbonates by Interpretable Machine Learning Approach Using Pore Structure Parameters Extracted from Thin Sections

Permeability prediction is crucial for carbonate reservoir characterization. Traditional measurement techniques are often limited by sample availability and cost. Permeability prediction using thin sections through digital image analysis (DIA) can overcome these limitations. Previous studies mainly focused on using single pore structure parameter (PSP) extraction method to predict the permeability of small sample size and type, with few applications in strong heterogeneous carbonates. In this study, a large sample size of carbonates with complex pore types from the Cretaceous Mishrif and Rumaila formations was collected to establish a reliable permeability prediction model based on PSPs and interpretable machine learning methods. We proposed a comprehensive framework allowing for the simultaneous extraction of 16 PSPs, including pore geometric parameters (PGPs) and pore-throat network parameters (PTNPs), from thin sections. The extracted PSPs can capture the subtle differences in pore structure of various pore types and effectively improve the permeability estimation of carbonates. Several machine learning models were developed and evaluated to obtain the optimal model for predicting permeability from PSPs. The random forest model optimized with genetic algorithm (RF-GA) achieved the best performance, with a coefficient of determination (R 2 ) of 0.756 and mean squared error (MSE) of 0.199, demonstrating strong predictive capability for highly heterogeneous carbonates. Interpretive methods were applied to the optimal model through evaluating the importance and contribution of each PSP, which reveals that both PGPs and PTNPs are crucial for permeability prediction, and therefore, both should be extracted to achieve the optimal prediction. The optimal model established in this study has good predictive performance for samples of various pore types, with 97% of the samples having permeability predictions within 1 order of magnitude of the actual values, making it suitable for strong heterogeneous carbonates. This study establishes a complete framework for low-cost and rapid estimation of carbonate permeability from thin sections and provides insights of the complex mechanisms from PSPs to permeability, which has a broad prospect in reservoir characterization in petroleum, geothermal heat and CO 2 storage fields.