Prediction of Formation Porosity from Well Log Data using Selected Machine Learning Algorithms

Abstract

Formation porosity is a critical petrophysical parameter for reservoir characterization, hydrocarbon resource assessment, and field development planning. Traditional porosity estimation methods rely on expensive and time-consuming core acquisition and laboratory analysis, which provide only sparse depth coverage and introduce operational delays. This study developed and evaluated machine learning-based predictive models for accurate, cost-effective porosity estimation directly from conventional well-log measurements, eliminating the need for core-analysis while providing continuous depth coverage. Four distinct machine learning algorithms were systematically implemented and compared: Random Forest Regressor, Gradient Boosting Regressor, Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptron (MLP) Neural Network. The target variable, RMS neutron-density porosity (PHI_ND), was computed from density and neutron porosity logs using standard petrophysical relationships. The dataset comprised 258,708 well log measurements from 21 wells spanning diverse geological settings, lithologies, and depth intervals from 500 to 4,400 meters. Cross-validation strategy allocated 80:20 % of data to training and testing, respectively, ensuring robust evaluation of model generalization to unseen geological conditions. Results showed Random Forest achieved near-perfect accuracy with R² = 0.9997 and RMSE = 0.00194 (0.19 porosity units), surpassing typical laboratory core-analysis reproducibility (±0.5 porosity units). Gradient Boosting delivered similarly outstanding results (R² = 0.9987, RMSE = 0.00278). XGBoost maintained strong performance (R² = 0.9981, RMSE = 0.00371), while MLP significantly underperformed (R² = 0.9776, RMSE = 0.01227), demonstrating approximately 6-fold higher error than Random Forest. Per-well analysis revealed that ensemble methods generalized excellently to all test wells, with individual well (R² = 0.9918 to 0.9999). Feature importance analysis validated that model learned physically consistent porosity-log relationships. Bulk-density gave the dominant predictor (permutation importance = 0.42), followed by neutron porosity and sonic measurements exactly the features emphasized in petrophysical analysis. This study showed ensemble machine learning provides accurate and reliable porosity prediction suitable for reservoir characterization.