Deep Learning Framework for Low-Quality Reservoir Classification and Production Zone Optimization in the Late-Mature Taybah Waterflood Field, Indonesia

Abstract The Taybah, a massive and mature waterflood oil field that operates over 2,000 wells, 80 drillings, and 800 workovers annually, faces challenges with high water-cut environments (99.4%), while bypassed oil remains under-optimized due to the inaccuracy of current formation evaluation in lower-quality reservoirs (LQR). This study utilizes an AI-driven approach to model the reservoirs, automate LQR play-type classification, correlate perforation intervals to water-cut, and recommend optimal perforation intervals for production optimization. This study uses five AI-driven deep learning models to analyze well log data across depth intervals, including gamma-ray, permeability, and resistivity. Using a state-of-the-art architecture with windowing and merging, the UTime model classifies log patterns into five play-types: Fining Upward (FU), Shaly Sand (SS), Thin Sand, Clean, and Shale. Long Short-Term Memory (LSTM) models further classify FU and SS into subtypes, while a combination of One-dimensional Convolutional Neural Network (1D-CNN) and LSTM classifies and predicts water-cut values, identifying optimal perforation intervals. The objectives are FU, SS, and Thin Sand labels, which serve as reservoir targets for minimizing water-cut. The performance of each deep learning model was evaluated using various metrics. The first model, UTime, classifies well log patterns into five play types with 74% accuracy, improving to 80% when Thin Sand is classified separately. The second and third models, LSTM, refine FU and SS into subtypes, achieving accuracies of 88% and 98%, respectively, outperforming rule-based methods that require extensive tuning. The fourth model, a combination of 1D-CNN and LSTM, achieves a Mean Absolute Error (MAE) of 0.46 in ordinal classification of water-cut categories using binary encoding. The fifth model, with sub-models for each water-cut class, achieves MAEs below 25% in water-cut prediction, meeting expert standards. The strong performance of these models is attributed to comprehensive hyperparameter and architecture optimization, including adjustments to network depth, filter size, and learning rate. Each model's loss function was customized to address imbalanced data and unique distributions. Despite the persistent challenge of subsurface uncertainty in predicting water-cut values, these models demonstrated superior performance compared to traditional methods. This approach enables more accurate recommendations for perforation intervals and water-cut predictions, highlighting its substantial potential for practical applications in large-scale reservoir management. The research underscores its relevance and effectiveness in real-world scenarios. This paper presents a novel deep-learning-driven approach for automated LQR classification, perforation recommendation, and water-cut prediction, enhancing reservoir management by reducing analysis time and minimizing interpretation biases. By applying deep learning models to capture depth-specific features, it enables large-scale field analysis where manual methods would be impractical. This empowers engineers to develop higher-quality drilling and workover programs, optimize production strategies, and improve success rates, offering a practical, reliable tool for complex reservoir challenges across scales.