Quantifying and Predicting Heterogeneity in Marine Clastic Reservoirs Through Machine Learning: Methodology and Applications
摘要
In tide-dominated marine environments, clastic reservoir sand bodies exhibit rapid vertical and lateral variability, making quantitative heterogeneity characterization essential for precise reservoir evaluation. By integrating core thin sections analysis, petrophysical data, production dynamics, and conventional well logs, a method was developed for quantitatively characterizing and predicting clastic reservoir heterogeneity using Pearson correlation, adaptive multi-dimensional kernel principal components analysis (AM-KPCA) for data integration, and stochastic gradient boosting decision tree (SGBDT) for prediction: (1) Nine heterogeneity-sensitive factors were identified at both macro- and microscales: Face, Kaot, Glau, Cal, Por, Perm, FZIm, R35, and σ, corresponding to sedimentary environments, rock types, and microstructural. (2) AM-KPCA eliminated redundant information and mitigated dimensionality issues, allowing for the integrated construction of the homogeneous composite factor (HCF), which effectively addressed nonlinear characterization challenges. (3) Pearson correlation analysis guided the selection of conventional heterogeneity-sensitive logging curves. When combined with SGBDT, these inputs formed a comprehensive model for heterogeneity prediction. This approach effectively integrates geological and geophysical knowledge and holds promising prospects for application, as evidenced by recent drilling activities in the T sandstone reservoirs of the Napo Formation in the Cretaceous Napo Formation of the Oriente Basin, South America. The evaluation results closely matched mercury injection test data, thereby enhancing the accuracy of complex reservoir structure characterization. The predicted favorable zones and remaining hydrocarbon distributions were confirmed through drilling, demonstrating significant potential for broader application.