Machine learning prediction of permeability distribution in the X field Malay Basin using elastic properties

Accurate estimation of porosity and permeability distribution is vital for reservoir characterization, particularly challenging in areas with limited well data. This study introduces a new approach to permeability estimation by implementing feature selection analysis and using elastic properties extracted from simultaneous seismic inversion method. Shapley Additive Explanations (SHAP) analysis was implemented to study the association between elastic properties, porosity, and permeability, while Recursive Feature Elimination with CrossValidation (RFECV) determined the optimal feature configuration. Together, they enhance model interpretability and optimize predictive performance of machine learning models. Several ensemble-based models, including Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LightBoost), Categorical Gradient Boosting (CatBoost), Bagging Regressor, Random Forest and Stacking, were evaluated for predictive performance, along with Multi-Layered Perceptron Neural Network algorithms. Additionally, Tree-structured Parzen Estimator-Bayesian optimization method was utilized to optimize the machine learning model's hyperparameters and improve accuracy. The results show that although some elastic properties lack direct correlation with permeability, they still contribute to its prediction. Bulk density and Quality factors of S-wave (SQs) display a close relationship with permeability. Moreover, combining porosity with elastic properties significantly improves model accuracy compared to using either feature independently. The LightBoost model achieved the highest accuracy (R2 = 0.97, RMSLE = 0.012) when porosity is integrated with the elastic properties, outperforming all other models. In contrast, XGBoost model performed better (R2 = 0.87, RMSLE = 0.195) using only elastic properties as features. This research highlights a robust method for predicting permeability distribution using elastic properties, which can significantly enhance the efficiency of reservoir assessment. By enabling accurate permeability predictions with minimal well data, this approach facilitates the rapid identification of highpotential zones, potentially replacing traditional geophysical methods.