A Pattern Recognition Approach to Reservoir Modeling: Comparative Performance of CNN and Gradient Boosting in Heterogeneous Carbonate Reservoirs

Missing intervals or unreliable well-log measurements pose a persistent challenge for subsurface characterization, particularly for curves critical to porosity, lithology, and velocity analysis. This study investigates two machine-learning strategies for reconstructing missing logs: a hybrid Multivariate Imputation by Chained Equations with Gradient Boosted Trees (MICE+GBT), and also Convolutional Neural Network (CNN). The prediction targets are three essential logs-neutron porosity (NPHI), bulk density (RHOB), and compressional travel time (DT). Each reconstructed artificial log from complementary measurements including resistivity, Photoelectric Factor (PEF), and spectral gamma-ray. A rigorous preprocessing workflow was applied, followed by evaluation under a well-level cross-validation scheme to simulate deployment on unseen wells. Performance was measured using Mean Squared Error (MSE), Mean Absolute Error (MAE), correlation coefficient (R), and coefficient of determination (R²). Results indicate that the MICE+GBT approach consistently outperforms Convolutional LSTM across all three target logs, particularly for intervals affected by washouts or tool measurement failures. Beyond statistical performance, the reconstructed curves preserved geological consistency in density–neutron and sonic–density trends, ensuring reliability for downstream reservoir interpretation. The findings demonstrate the practical benefits of adapting ensemble-based imputation methods to petrophysical data, providing a robust and interpretable framework for improving log data alliance in reservoir studies.

کلیدواژه ها:

gradient boosting tree ، petrophysical well logs ، convolutional long short-term memory (convlstm) ، convolutional neural network ، machine learning