A hybrid framework of rock typing mechanism and data driven modeling for permeability prediction in heterogeneous carbonate rocks

Accurate permeability prediction is essential for characterizing and efficiently developing heterogeneous carbonate reservoirs. Conventional core analysis is not only costly but also inherently limited in spatial representativeness. Purely data driven models, while capable of capturing complex nonlinear relationships between pore-structure descriptors and permeability, often suffer from limited generalizability and poor physical interpretability. To address these limitations, we proposed a hybrid permeability prediction framework that integrated the rock-typing guided physical constraints with data-driven learning. Based on the hydraulic rock-typing theory, we first developed a binary permeability–porosity model specifically adapted to carbonate lithologies. Four microstructural parameters extracted from cast thin sections, namely porosity, dominant pore size, perimeter-to-area ratio, and pore roundness, were then used as inputs to train different machine learning models under a nested cross-validation scheme comprising 10 outer folds and 5 inner folds for estimating the flow zone index (FZI) and discrete rock type (DRT). The predicted FZI/DRT was subsequently coupled with the rock-typing constrained permeability–porosity model to enable indirect, thin-section-scale permeability inference. Compared with the purely data driven models and DRT-based coupled models, the FZI-based coupled model achieves substantially improved performance, yielding the R 2 values of 0.689 on the testing dataset. Particularly in the medium-permeability interval, it achieves the highest R 2 (0.63) alongside the lowest RMSE (0.51), MAE (0.34), and mean residual (− 0.05) on the log-scale. These results demonstrate that mechanism-informed data fusion improves both the predictive accuracy and the model interpretability for heterogeneous carbonate samples, offering an alternative and cost-effective pathway for rapid reservoir evaluation using routine thin-section analysis.