Comparative Evaluation of Machine Learning Algorithms for Permeability Prediction from Wireline Well Log Data

Abstract

Accurate permeability prediction from wireline well logs remain a persistent challenge in reservoir characterization, as conventional empirical correlations often fail to capture the heterogeneous nature of reservoirs, while core-derived permeability data are costly and rarely available across all wells. Machine learning offers a data-driven alternative capable of modelling these complex relationships, but most existing studies rely on random train-test splitting, which ignores the spatial autocorrelation inherent in well log data and can overstate model generalizability. In this research, five different models of machine learning: random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), one-dimensional convolutional neural network (1D-CNN), and CNN–LSTM network were compared in terms of their ability to predict formation permeability based on well log data collected from three wells (Jay_01, Jay_02, Jay_03) in the Jay field, Niger Delta Basin, Nigeria. Four log curves, namely gamma ray (GR), neutron porosity (NPHI), deep resistivity (RT), and bulk density (RHOB), were used as inputs. The methodology used in this paper compares model performance through two validation procedures: random train-test splits (80% and 20%), and Leave-One-Well-Out (LOWO) cross-validation where every well in turn was left out as test data to take into account spatial autocorrelation in sequential in-depth logging data. R² values estimated through random splits (0.83-0.97) overestimate generalization capacity of the models in comparison with LOWO R² values (0.63-0.73). Among the five models, XGBoost performs the best under LOWO procedure (R² = 0.734±0.163, MAE = 0.654 log-mD), followed by Random Forest (R² = 0.725±0.169).