Intelligent Porosity Prediction for Sandstone Reservoirs Using Machine Learning Techniques

Authors:

Abstract
Accurate porosity prediction is essential for reliable reservoir characterization in data-limited and heterogeneous formation. Traditional approaches generally have a difficulty handling the inherent complexities and uncertainties of well log data. This study applies and compares three machine learning (ML) approaches, including Artificial Neural Network optimized with Levenberg-Marquardt (ANN-LM), Random Forest (RF), Fuzzy Logic (FL) along with a baseline Multiple Linear Regression (MLR) model, to estimate total porosity from standard geophysical well-logs in three wells from the Mazalai Gas Field (MGF), Kohat Basin, Pakistan. The models utilize sonic, neutron porosity, bulk density, and gamma ray as input parameters. The ANN-LM model was trained using backpropagation and K-fold cross-validation. RF was implemented as an ensemble of decision trees with feature ranking, FL employed Gaussian membership functions in ten bins, and MLR served as a baseline linear method. Model performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE). ANN-LM showed the strongest generalizability and robustness, achieving R² = 0.99 and RMSE = 3.5 pu by effectively minimizing errors in complex, nonlinear and heterogenous data. RF and FL performed reasonably well achieving R 2 equal to 0.89 and 0.85 respectively, but showed reduced generalization to unseen data. MLR demonstrated the lowest performance acquiring R 2 =0.82. Additionally, A Taylor diagram analysis revealed that ANN-LM provided the most accurate and statistically consistent predictions, closely matching the reference data. These results show machine learning, especially well-optimised neural networks, greatly improves porosity prediction from logs, strengthening reservoir evaluation and development planning in MGF-like settings.