On the evaluation of permeability of heterogeneous carbonate reservoirs using rigorous data-driven techniques

Abstract

This study probes the application of Cascade Forward Neural Network (CFNN), Least Square Support Vector Machine (LSSVM), Multilayer Perceptron (MLP), and Generalized Regression Neural Network (GRNN) techniques for modeling the absolute permeability of carbonate rocks in terms of pore specific surface area, porosity, and irreducible water saturation. The control parameters of the MLP and CFNN models were tuned through Levenberg Marquardt Algorithm (LMA) and Bayesian Regularization (BR) optimizers, and the LSSVM paradigm was optimized using Gravitational Search Algorithm (GSA). Accordingly, six intelligent schemes, viz. MLP-BR, MLP-LMA, LSSVM-GSA, CFNN-BR, CFNN-LMA, and GRNN were trained by utilizing 80% of a valuable set of core data compiled from reliable literature and were tested through the rest of the data points (20%). The accuracy of the proposed paradigms was evaluated using several statistical and graphical assessments. The overall results were fulfilling and fair enough for the scope of this study. The proposed MLP-BR, MLP-LMA, LSSVM-GSA, CFNN-BR, CFNN-LMA, and GRNN models were associated with the Root Mean Square Errors of 6.8019, 5.6225, 165.8852, 6.6841, 5.2136, and 11.1799, respectively. The results were endorsed through 3-fold cross-validation. Furthermore, outlier detection was carried out by means of the plot of standardized residuals versus Leverage values. For all models, the majority of the points were valid values distributing in the applicability domain of the models. In the end, the developed models were compared against two literature smart models and a traditional correlation. The results demonstrated that the recently generated models offer remarkably higher accuracy than other alternatives followed by the Gene Expression Programming (GEP) modeling approach.