Pore-Structure-Aware Prediction of Pressure-Dependent Pore-Volume Compressibility in Ultra-Deep Fractured-Vuggy Carbonate Reservoirs

Pressure-dependent pore-volume deformation is a critical but poorly constrained variable in dynamic reserve assessment for ultra-deep fractured-vuggy carbonate reservoirs, where fractures, dissolution pores, and vugs respond differently to effective-stress loading. In this work, a pore-structure-aware evaluation strategy was developed by integrating high-temperature and high-pressure volumetric measurements with data-driven regression. Twelve carbonate core plugs from the Ordovician Yijianfang and Yingshan formations of the Fuman Oilfield were selected to represent matrix-pore, dissolution-pore, fracture-vug, and fracture-dominated pore systems. Stepwise net-pressure experiments were performed under simulated reservoir conditions, and pore-volume compressibility (Cp) was calculated from corrected pore-volume changes. Measured Cp values reveal a distinct stress-sensitive response: Cp declines sharply during the low-net-pressure stage and then tends toward a quasi-stable level as net pressure increases, indicating progressive closure of mechanically compliant fractures, narrow throats, and weakly supported dissolution pores. Although porosity is positively associated with Cp, samples with comparable porosity display markedly different compressibility values, confirming that pore-space geometry and fracture-related compliance must be considered. Eight representative regression algorithms were then compared, using net pressure, porosity, permeability, initial pore volume, surface porosity, temperature, and a pore-structure index as model inputs. To further assess model generalization to completely unseen core plugs, additional core-ID-based leave-one-core-out (LOCO) validation was performed for k-nearest neighbors and AdaBoost. Under this grouped validation, k-nearest neighbors yielded an RMSE of 13.5978 × 10−4 MPa−1 and an R2 of 0.8408, whereas AdaBoost achieved an RMSE of 10.0160 × 10−4 MPa−1 and an R2 of 0.9136, indicating greater cross-core robustness of AdaBoost. Permutation-importance analysis of the split-specific KNN model indicated that net pressure, porosity, surface porosity, and pore-structure index made the largest predictive contributions within that model. Moreover, the predicted normalized Cp values reproduced the experimentally observed decreasing trend with increasing net pressure, supporting the physical consistency of the k-nearest neighbors predictions. The proposed experimental–machine learning framework offers a pressure-dependent method for estimating pore-volume compressibility within the geological and petrophysical domain represented by the investigated Fuman Oilfield cores, and provides more representative inputs for material-balance analysis, dynamic reserve evaluation, and production adjustment.