Integrating Digital Rock Physics with Well-Log Data for Continuous Pore-Throat Characterization via Machine Learning

Characterizing pore–throat structures at the micron scale is essential for evaluating productivity in deepwater tight sandstone reservoirs, yet conventional well-log analysis lacks the resolution required to capture these features. This study proposes a scalable machine-learning-based workflow that integrates digital rock physics (DRP) measurements with continuous well-log data to enable the full-wellbore characterization of pore–throat properties. A Stacking ensemble model combining XGBoost, Gradient Boosting, Random Forest, Extra Trees, and CatBoost is developed to establish a robust cross-scale mapping from sparse DRP samples to logging responses under small-sample conditions. Using 423 DRP samples from 65 wells in the eastern South China Sea, the model achieves an average R2 of 0.90 in 10-fold cross-validation, substantially outperforming single-model approaches. Continuous 0.1 m resolution profiles of median pore–throat radius (R50) and connected pore ratio (CPR) show strong and physically consistent correspondence with independent production test data, with high R50 and CPR intervals systematically associated with higher oil rates. The results demonstrate that the proposed workflow provides a practical and engineering-relevant solution for continuous micron-scale reservoir characterization, supporting improved sweet-spot identification and development decision-making in deepwater tight sandstone reservoirs.