Data-Driven and Machine Learning Prediction of Early Water Breakthrough Time in Naturally Fractured Reservoirs in the Middle East

Effective management of water-driven, Naturally Fractured Reservoirs (NFR) requires minimizing water production. This study compares two predictive approaches-statistical methods, specifically Response Surface Methodology (RSM), and artificial intelligence (AI)-to determine their effectiveness in predicting early water breakthrough time (t_bt) in NFRs. A dataset comprising 261 simulation cases was generated based on design of experiments approach, using dual-porosity dual-permeability simulation models, varying key reservoir and fracture parameters, including fracture permeability, matrix permeability, fracture spacing, storativity ratio, mobility ratio, and production rate. A log transformation was applied to normalize the dataset, improving predictive accuracy. The study employed RSM for statistical modeling and various AI-based approaches, including Neural Networks, Optimizable Ensemble of Trees, and Optimizable Gaussian Process Regression, trained on 80% of the data and tested on the remaining 20% to evaluate their performance. The log-transformed dataset significantly improved prediction accuracy, with the best performing model being Optimizable Ensemble of Trees model with a validation Root Mean Square Error (RMSE) of 0.51631 and a Coefficient of Determination (R²) of 0.86, while its test RMSE was 0.50674 with an R² of 0.90. In comparison, the RSM approach produced a validation RMSE of 0.9392, an R² of 0.9261, an adjusted R² of 0.8960, and a predicted R² of 0.8175, with an Adequate Precision of 26.9, indicating a strong predictive signal. While RSM provided valuable interpretability and parameter sensitivity insights, it struggled with complex parameter interactions and exhibited higher variability in predictions. AI models demonstrated superior capability in capturing the non-linear relationships between fracture and matrix properties and production rates, making them more reliable for practical reservoir applications. This study concludes that AI models, given sufficient data, offer a more robust and scalable solution for predicting early water breakthrough times. The integration of machine learning (ML) offers enhanced accuracy, scalability, and decision-making potential for optimizing water management strategies in fractured reservoirs.