Enhanced Permeability Prediction in Heterogeneous Tight Sandstones: A Hybrid Model Combining Hydraulic Flow Units (HFU) with Deep Learning Method

Abstract

Permeability prediction in tight sandstone reservoirs remains challenging when using conventional methods, largely due to the weak correlation among rock physical properties caused by strong heterogeneity. Models relying on pore structure and connectivity, e.g., the Swanson-based model, the R10-based model (pore-throat radius corresponding to 10% mercury injection saturation), the R50-based model (median pore-throat radius), are also inadequate, as nuclear magnetic resonance (NMR)-derived pore structure tends to be overestimated in tight hydrocarbon-bearing sandstones due to the influence of saturated hydrocarbon in the pore spaces on T2 spectra. Although the hydraulic flow unit (HFU) method is considered effective for building accurate permeability prediction models in heterogeneous reservoirs, its crucial parameter—the flow zone indicator (FZI)—currently lacks a reliable estimation method apart from empirical statistics. In this study, 1254 core samples, recovered from the Permian He 8 Formation in the western Sulige Region of the Ordos Basin, are collected and applied for routine core analysis and NMR experiments. By integrating HFU theory with NMR principles, a correlation is established between FZI and NMR-derived porosity and T2 geometric mean (T2lm). The core samples are classified into six distinct hydraulic flow units, and porosity-permeability relationships are developed for each formation type. To account for hydrocarbon effects on NMR responses, an XGBoost-based deep learning approach is employed to evaluate the sensitivity of conventional and NMR logging curves to FZI. The analysis identifies the most sensitive parameters as the deep-to-shallow resistivity ratio, total porosity, neutron, density, interval transit time and T2lm. A novel FZI prediction model is subsequently developed by integrating these key parameters. When applied to field data, the model generates a continuous FZI profile used for formation classification. Permeability is then predicted using the six cluster-specific models. Validation against core-derived data shows a relative error of only 5.95% for FZI prediction and 9.35% for permeability, confirming the reliability of the proposed method. This approach not only enables accurate permeability evaluation in heterogeneous tight sandstones but also holds promise for other reservoir types, as it addresses the critical challenge of FZI prediction within the HFU framework.