Geologically Deep Learning for High-Resolution Sparse 3D Oil Reservoir Modeling

• Geo-RealESRGAN reconstructs 3D reservoirs from sparse and low-resolution data. • Geostatistical loss maintains spatial autocorrelation in heterogeneous zones. • Tri-directional fusion preserves 3D geological continuity and structure. • A hybrid loss function balances numerical precision and structural fidelity. Accurate three-dimensional oil reservoir modeling is essential for optimizing unconventional oil and gas development, yet traditional methods often fail to characterize high-resolution properties under sparse sampling and extreme geological heterogeneity. This study develops Geo-RealESRGAN, a deep learning model that integrates perceptual guidance, adversarial learning, and geostatistical constraints to achieve high-fidelity property reconstruction. The methodology utilizes a Slice-Reconstruction-Fusion workflow with tri-directional slicing and multi-channel image encoding to capture complex spatial continuity and physical coupling relationships among reservoir parameters. Comprehensive ablation studies and comparisons with the Sequential Gaussian Simulation model and the Real-ESRGAN model demonstrate that the Geo-RealESRGAN model achieves lower root mean square error and and relative error rate in predicting all continuous properties. Specifically, the relative error for permeability reconstruction is reduced by over 50%, while classification accuracies for lithofacies and oil phase reach 98% and 97%, respectively. 3D property visualization further confirms the Geo-RealESRGAN model's superior capability in preserving geological structures, particularly near heterogeneous boundaries and transition zones. These results validate the effectiveness of deep learning in high-resolution reservoir modeling, offering a viable pathway for intelligent reservoir modeling, drilling and completion design.