Comparative Machine Learning Framework for Permeability Prediction in a Heterogeneous Carbonate Reservoir

Abstract The use of traditional well logs to predict the permeability of heterogeneous carbonate reservoirs remains a challenge because of complicated connectivity of pore networks and nonlinear petrophysical processes. Although many machine learning processes are suggested, many studies use small validation processes and one-model evaluations, which are bound to give encouraging results. The article reports a statistically transparent and benchmarked machine learning process of making predictions of permeability on a single dataset of 130 matched core-log measurements of the Mishrif carbonate reservoir in southern Iraq. To represent the variability of the permeability on a multi-order-of-magnitude scale and stabilise the regression behaviour, the permeability was scaled on log scale (log10(k)). As opposed to the preceding studies, a systematic comparison of four machine learning models, i.e., Artificial Neural Network (ANN), Support Vector Regression (SVR), Random Forest (RF), and Gradient Boosting Regressor (GBR) was conducted under identical conditions of 5-fold cross-validation to warrant that there was no bias in the analysis of generalisation. The findings indicate how powerful ensemble tree-based methods are whereby the optimum predictive performance (R 2 = 0.92) was implemented using the Random Forest, then Gradient Boosting (R 2 = 0.90), ANN (R 2 = 0.88), and SVR (R 2 = 0.85). In addition to predictive accuracy, the paper incorporates feature importance analysis, nonlinear sensitivity analysis, and two-dimensional interaction surface to obtain physical understanding of how porosity and shale volume interact with each other to affect the permeability behaviour. The results show that the increase in permeability as a result of the increase in porosity can be neutralized to some extent by shale‑associated flow limiting factors and this suggests the importance of modelling nonlinear processes in carbonate systems. The suggested framework offers a sensible and reproducible direction of log‑based permeability estimation and a goal benchmarking criterion of machine learning applications in heterogeneous reservoirs.